back to index

Jeffrey Shainline: Neuromorphic Computing and Optoelectronic Intelligence | Lex Fridman Podcast #225


small model | large model

link |
00:00:00.000
The following is a conversation with Jeff Schoenlein,
link |
00:00:03.060
a scientist at NIST
link |
00:00:04.600
interested in optoelectronic intelligence.
link |
00:00:08.440
We have a deep technical dive into computing hardware
link |
00:00:12.160
that will make Jim Keller proud.
link |
00:00:14.080
I urge you to hop onto this rollercoaster ride
link |
00:00:17.440
through neuromorphic computing
link |
00:00:19.320
and superconducting electronics
link |
00:00:21.720
and hold on for dear life.
link |
00:00:24.360
Jeff is a great communicator of technical information
link |
00:00:27.200
and so it was truly a pleasure to talk to him
link |
00:00:30.160
about some physics and engineering.
link |
00:00:33.160
To support this podcast,
link |
00:00:34.440
please check out our sponsors in the description.
link |
00:00:37.440
This is the Lex Friedman Podcast
link |
00:00:39.880
and here is my conversation with Jeff Schoenlein.
link |
00:00:44.380
I got a chance to read a fascinating paper you authored
link |
00:00:48.920
called Optoelectronic Intelligence.
link |
00:00:52.040
So maybe we can start by talking about this paper
link |
00:00:55.180
and start with the basic questions.
link |
00:00:57.060
What is optoelectronic intelligence?
link |
00:01:00.360
Yeah, so in that paper,
link |
00:01:02.040
the concept I was trying to describe
link |
00:01:04.280
is sort of an architecture
link |
00:01:06.680
for building brain inspired computing
link |
00:01:10.820
that leverages light for communication
link |
00:01:13.380
in conjunction with electronic circuits for computation.
link |
00:01:17.640
In that particular paper,
link |
00:01:18.920
a lot of the work we're doing right now
link |
00:01:20.880
in our project at NIST
link |
00:01:22.000
is focused on superconducting electronics for computation.
link |
00:01:25.840
I won't go into why that is,
link |
00:01:27.760
but that might make a little more sense in context
link |
00:01:31.440
if we first describe what that is in contrast to,
link |
00:01:35.440
which is semiconducting electronics.
link |
00:01:37.880
So is it worth taking a couple minutes
link |
00:01:39.720
to describe semiconducting electronics?
link |
00:01:42.820
It might even be worthwhile to step back
link |
00:01:45.960
and talk about electricity and circuits
link |
00:01:49.960
and how circuits work
link |
00:01:52.240
before we talk about superconductivity.
link |
00:01:54.760
Right, okay.
link |
00:01:56.560
How does a computer work, Jeff?
link |
00:01:58.080
Well, I won't go into everything
link |
00:01:59.800
that makes a computer work,
link |
00:02:01.120
but let's talk about the basic building blocks,
link |
00:02:05.200
a transistor, and even more basic than that,
link |
00:02:08.440
a semiconductor material, silicon, say.
link |
00:02:11.520
So in silicon, silicon is a semiconductor,
link |
00:02:15.440
and what that means is at low temperature,
link |
00:02:18.440
there are no free charges,
link |
00:02:20.760
no free electrons that can move around.
link |
00:02:22.880
So when you talk about electricity,
link |
00:02:24.920
you're talking about predominantly electrons
link |
00:02:28.040
moving to establish electrical currents,
link |
00:02:30.780
and they move under the influence of voltages.
link |
00:02:33.080
So you apply voltages, electrons move around,
link |
00:02:36.640
those can be measured as currents,
link |
00:02:38.420
and you can represent information in that way.
link |
00:02:40.920
So semiconductors are special
link |
00:02:43.840
in the sense that they are really malleable.
link |
00:02:46.860
So if you have a semiconductor material,
link |
00:02:50.080
you can change the number of free electrons
link |
00:02:52.600
that can move around by putting different elements,
link |
00:02:56.020
different atoms in lattice sites.
link |
00:02:58.280
So what is a lattice site?
link |
00:03:00.200
Well, a semiconductor is a crystal,
link |
00:03:02.200
which means all the atoms that comprise the material
link |
00:03:06.480
are at exact locations
link |
00:03:09.000
that are perfectly periodic in space.
link |
00:03:10.880
So if you start at any one atom
link |
00:03:12.220
and you go along what are called the lattice vectors,
link |
00:03:14.660
you get to another atom and another atom and another atom,
link |
00:03:17.320
and for high quality devices,
link |
00:03:19.440
it's important that it's a perfect crystal
link |
00:03:21.920
with very few defects,
link |
00:03:23.820
but you can intentionally replace a silicon atom
link |
00:03:27.440
with say a phosphorus atom,
link |
00:03:29.240
and then you can change the number of free electrons
link |
00:03:32.160
that are in a region of space
link |
00:03:33.980
that has that excess of what are called dopants.
link |
00:03:37.140
So picture a device that has a left terminal
link |
00:03:40.260
and a right terminal,
link |
00:03:42.040
and if you apply a voltage between those two,
link |
00:03:44.160
you can cause electrical current to flow between them.
link |
00:03:47.740
Now we add a third terminal up on top there,
link |
00:03:52.120
and depending on the voltage
link |
00:03:53.800
between the left and right terminal and that third voltage,
link |
00:03:57.160
you can change that current.
link |
00:03:58.720
So what's commonly done in digital electronic circuits
link |
00:04:02.280
is to leave a fixed voltage from left to right,
link |
00:04:06.560
and then change that voltage
link |
00:04:08.600
that's applied at what's called the gate,
link |
00:04:10.000
the gate of the transistor.
link |
00:04:11.220
So what you do is you make it to where
link |
00:04:13.960
there's an excess of electrons on the left,
link |
00:04:15.680
excess of electrons on the right,
link |
00:04:17.360
and very few electrons in the middle,
link |
00:04:18.880
and you do this by changing the concentration
link |
00:04:21.520
of different dopants in the lattice spatially.
link |
00:04:24.800
And then when you apply a voltage to that gate,
link |
00:04:27.900
you can either cause current to flow or turn it off,
link |
00:04:30.680
and so that's sort of your zero and one.
link |
00:04:33.440
If you apply voltage, current can flow,
link |
00:04:35.360
that current is representing a digital one,
link |
00:04:38.380
and from that, from that basic element,
link |
00:04:41.960
you can build up all the complexity
link |
00:04:44.200
of digital electronic circuits
link |
00:04:45.520
that have really had a profound influence on our society.
link |
00:04:50.080
Now you're talking about electrons.
link |
00:04:51.600
Can you give a sense of what scale we're talking about
link |
00:04:54.680
when we're talking about in silicon
link |
00:04:57.600
being able to mass manufacture these kinds of gates?
link |
00:05:01.560
Yeah, so scale in a number of different senses.
link |
00:05:04.480
Well, at the scale of the silicon lattice,
link |
00:05:07.520
the distance between two atoms there is half a nanometer.
link |
00:05:10.880
So people often like to compare these things
link |
00:05:14.480
to the width of a human hair.
link |
00:05:16.540
I think it's some six orders of magnitude smaller
link |
00:05:20.040
than the width of a human hair, something on that order.
link |
00:05:24.440
So remarkably small,
link |
00:05:25.840
we're talking about individual atoms here,
link |
00:05:27.280
and electrons are of that length scale
link |
00:05:29.480
when they're in that environment.
link |
00:05:31.520
But there's another sense
link |
00:05:32.440
that scale matters in digital electronics.
link |
00:05:34.380
This is perhaps the more important sense,
link |
00:05:36.460
although they're related.
link |
00:05:37.880
Scale refers to a number of things.
link |
00:05:41.340
It refers to the size of that transistor.
link |
00:05:44.560
So for example, I said you have a left contact,
link |
00:05:47.280
a right contact, and some space between them
link |
00:05:49.800
where the gate electrode sits.
link |
00:05:52.280
That's called the channel width or the channel length.
link |
00:05:56.000
And what has enabled what we think of as Moore's law
link |
00:06:00.120
or the continued increased performance
link |
00:06:03.040
in silicon microelectronic circuits
link |
00:06:05.560
is the ability to make that size, that feature size,
link |
00:06:08.960
ever smaller, ever smaller at a really remarkable pace.
link |
00:06:14.800
I mean, that feature size has decreased consistently
link |
00:06:20.160
every couple of years since the 1960s.
link |
00:06:23.760
And that was what Moore predicted in the 1960s.
link |
00:06:27.160
He thought it would continue for at least two more decades,
link |
00:06:29.480
and it's been much longer than that.
link |
00:06:30.800
And so that is why we've been able to fit ever more devices,
link |
00:06:35.800
ever more transistors, ever more computational power
link |
00:06:39.000
on essentially the same size of chip.
link |
00:06:41.440
So a user sits back and does essentially nothing.
link |
00:06:44.400
You're running the same computer program,
link |
00:06:45.960
but those devices are getting smaller, so they get faster,
link |
00:06:48.360
they get more energy efficient,
link |
00:06:50.080
and all of our computing performance
link |
00:06:51.520
just continues to improve.
link |
00:06:53.240
And we don't have to think too hard
link |
00:06:56.200
about what we're doing as, say,
link |
00:06:59.200
a software designer or something like that.
link |
00:07:00.520
I absolutely don't mean to say
link |
00:07:02.200
that there's no innovation in software or the user side
link |
00:07:05.960
of things, of course there is.
link |
00:07:07.040
But from the hardware perspective,
link |
00:07:09.400
we just have been given this gift
link |
00:07:12.240
of continued performance improvement
link |
00:07:14.520
through this scaling that is ever smaller feature sizes
link |
00:07:19.080
with very similar, say, power consumption.
link |
00:07:22.680
That power consumption has not continued to scale
link |
00:07:25.880
in the most recent decades, but nevertheless,
link |
00:07:29.360
we had a really good run there for a while.
link |
00:07:31.200
And now we're down to gates that are seven nanometers,
link |
00:07:34.160
which is state of the art right now.
link |
00:07:36.080
Maybe GlobalFoundries is trying to push it
link |
00:07:38.760
even lower than that.
link |
00:07:39.600
I can't keep up with where the predictions are
link |
00:07:42.440
that it's gonna end.
link |
00:07:43.280
But seven nanometer transistor has just a few tens of atoms
link |
00:07:49.880
along the length of the conduction pathway.
link |
00:07:51.720
So a naive semiconductor device physicist
link |
00:07:56.640
would think you can't go much further than that
link |
00:07:58.560
without some kind of revolution in the way we think
link |
00:08:02.240
about the physics of our devices.
link |
00:08:03.800
Is there something to be said
link |
00:08:04.960
about the mass manufacture of these devices?
link |
00:08:08.160
Right, right, so that's another thing.
link |
00:08:09.320
So how have we been able
link |
00:08:10.800
to make those transistors smaller and smaller?
link |
00:08:13.840
Well, companies like Intel, GlobalFoundries,
link |
00:08:17.720
they invest a lot of money in the lithography.
link |
00:08:20.280
So how are these chips actually made?
link |
00:08:22.800
Well, one of the most important steps
link |
00:08:24.320
is this what's called ion implantation.
link |
00:08:27.640
So you start with sort of a pristine silicon crystal
link |
00:08:31.920
and then using photolithography,
link |
00:08:34.040
which is a technique where you can pattern
link |
00:08:36.440
different shapes using light,
link |
00:08:38.760
you can define which regions of space
link |
00:08:41.360
you're going to implant with different species of ions
link |
00:08:45.600
that are going to change
link |
00:08:46.560
the local electrical properties right there.
link |
00:08:49.520
So by using ever shorter wavelengths of light
link |
00:08:52.640
and different kinds of optical techniques
link |
00:08:54.360
and different kinds of lithographic techniques,
link |
00:08:56.560
things that go far beyond my knowledge base,
link |
00:09:00.920
you can just simply shrink that feature size down.
link |
00:09:03.280
And you say you're at seven nanometers.
link |
00:09:04.640
Well, the wavelength of light that's being used
link |
00:09:07.200
is over a hundred nanometers.
link |
00:09:08.560
That's already deep in the UV.
link |
00:09:10.040
So how are those minute features patterned?
link |
00:09:14.520
Well, there's an extraordinary amount of innovation
link |
00:09:16.720
that has gone into that,
link |
00:09:18.000
but nevertheless, it stayed very consistent
link |
00:09:20.080
in this ever shrinking feature size.
link |
00:09:21.720
And now the question is, can you make it smaller?
link |
00:09:24.560
And even if you do, do you still continue
link |
00:09:26.760
to get performance improvements?
link |
00:09:28.120
But that's another kind of scaling
link |
00:09:30.040
where these companies have been able to...
link |
00:09:34.080
So, okay, you picture a chip that has a processor on it.
link |
00:09:36.840
Well, that chip is not made as a chip.
link |
00:09:38.600
It's made on a wafer.
link |
00:09:40.680
And using photolithography,
link |
00:09:43.160
you basically print the same pattern on different dyes
link |
00:09:47.000
all across the wafer, multiple layers,
link |
00:09:49.160
tens, probably a hundred some layers
link |
00:09:53.120
in a mature foundry process.
link |
00:09:55.000
And you do this on ever bigger wafers too.
link |
00:09:57.360
That's another aspect of scaling
link |
00:09:58.800
that's occurred in the last several decades.
link |
00:10:00.680
So now you have this 300 millimeter wafer.
link |
00:10:02.720
It's like as big as a pizza
link |
00:10:04.040
and it has maybe a thousand processors on it.
link |
00:10:06.360
And then you dice that up using a saw.
link |
00:10:08.640
And now you can sell these things so cheap
link |
00:10:11.520
because the manufacturing process was so streamlined.
link |
00:10:14.920
I think a technology as revolutionary
link |
00:10:17.160
as silicon microelectronics has to have
link |
00:10:19.880
that kind of manufacturing scalability,
link |
00:10:23.480
which I will just emphasize,
link |
00:10:25.400
I believe is enabled by physics.
link |
00:10:28.920
It's not, I mean, of course there's human ingenuity
link |
00:10:31.480
that goes into it, but at least from my side where I sit,
link |
00:10:35.840
it sure looks like the physics of our universe
link |
00:10:38.760
allows us to produce that.
link |
00:10:40.880
And we've discovered how more so than we've invented it,
link |
00:10:45.720
although of course we have invented it,
link |
00:10:47.160
humans have invented it,
link |
00:10:48.440
but it's almost as if it was there
link |
00:10:50.960
waiting for us to discover it.
link |
00:10:52.960
You mean the entirety of it
link |
00:10:54.160
or are you specifically talking about
link |
00:10:55.920
the techniques of photolithography,
link |
00:10:58.320
like the optics involved?
link |
00:10:59.680
I mean, the entirety of the scaling down
link |
00:11:02.560
to the seven nanometers,
link |
00:11:04.320
you're able to have electrons not interfere with each other
link |
00:11:08.200
in such a way that you could still have gates.
link |
00:11:11.280
Like that's enabled.
link |
00:11:13.320
To achieve that scale, spatial and temporal,
link |
00:11:16.560
it seems to be very special
link |
00:11:18.720
and is enabled by the physics of our world.
link |
00:11:21.480
All of the things you just said.
link |
00:11:22.760
So starting with the silicon material itself,
link |
00:11:25.960
silicon is a unique semiconductor.
link |
00:11:28.760
It has essentially ideal properties
link |
00:11:31.560
for making a specific kind of transistor
link |
00:11:33.720
that's extraordinarily useful.
link |
00:11:35.200
So I mentioned that silicon has,
link |
00:11:39.560
well, when you make a transistor,
link |
00:11:40.640
you have this gate contact
link |
00:11:42.000
that sits on top of the conduction channel.
link |
00:11:44.520
And depending on the voltage you apply there,
link |
00:11:47.240
you pull more carriers into the conduction channel
link |
00:11:50.000
or push them away so it becomes more or less conductive.
link |
00:11:53.040
In order to have that work
link |
00:11:54.680
without just sucking those carriers right into that contact,
link |
00:11:57.120
you need a very thin insulator.
link |
00:11:59.080
And part of scaling has been to gradually decrease
link |
00:12:03.200
the thickness of that gate insulator
link |
00:12:06.040
so that you can use a roughly similar voltage
link |
00:12:09.040
and still have the same current voltage characteristics.
link |
00:12:12.080
So the material that's used to do that,
link |
00:12:14.600
or I should say was initially used to do that
link |
00:12:16.880
was a silicon dioxide,
link |
00:12:18.720
which just naturally grows on the silicon surface.
link |
00:12:21.800
So you expose silicon to the atmosphere that we breathe
link |
00:12:25.080
and well, if you're manufacturing,
link |
00:12:27.560
you're gonna purify these gases,
link |
00:12:29.720
but nevertheless,
link |
00:12:30.560
that what's called a native oxide will grow there.
link |
00:12:33.520
There are essentially no other materials
link |
00:12:36.040
on the entire periodic table
link |
00:12:37.480
that have as good of a gate insulator
link |
00:12:42.360
as that silicon dioxide.
link |
00:12:43.680
And that has to do with nothing but the physics
link |
00:12:46.280
of the interaction between silicon and oxygen.
link |
00:12:49.240
And if it wasn't that way,
link |
00:12:51.240
transistors could not perform
link |
00:12:54.720
in nearly the degree of capability that they have.
link |
00:12:58.800
And that has to do with the way that the oxide grows,
link |
00:13:02.880
the reduced density of defects there,
link |
00:13:05.760
it's insulation, meaning essentially it's energy gaps.
link |
00:13:08.600
You can apply a very large voltage there
link |
00:13:10.320
without having current leak through it.
link |
00:13:12.280
So that's physics right there.
link |
00:13:15.760
There are other things too.
link |
00:13:16.880
Silicon is a semiconductor in an elemental sense.
link |
00:13:19.840
You only need silicon atoms.
link |
00:13:21.560
A lot of other semiconductors,
link |
00:13:22.800
you need two different kinds of atoms,
link |
00:13:24.880
like a compound from group three
link |
00:13:26.800
and a compound from group five.
link |
00:13:28.560
That opens you up to lots of defects that can occur
link |
00:13:31.680
where one atom's not sitting quite at the lattice site,
link |
00:13:34.360
it is and it's switched with another one
link |
00:13:35.960
that degrades performance.
link |
00:13:38.280
But then also on the side that you mentioned
link |
00:13:40.280
with the manufacturing,
link |
00:13:43.480
we have access to light sources
link |
00:13:45.840
that can produce these very short wavelengths of light.
link |
00:13:49.480
How does photolithography occur?
link |
00:13:50.880
Well, you actually put this polymer on top of your wafer
link |
00:13:54.440
and you expose it to light,
link |
00:13:56.200
and then you use a aqueous chemical processing
link |
00:14:00.080
to dissolve away the regions that were exposed to light
link |
00:14:03.440
and leave the regions that were not.
link |
00:14:05.920
And we are blessed with these polymers
link |
00:14:08.520
that have the right property
link |
00:14:09.640
where they can cause scission events
link |
00:14:13.560
where the polymer splits where a photon hits.
link |
00:14:16.160
I mean, maybe that's not too surprising,
link |
00:14:19.720
but I don't know, it all comes together
link |
00:14:21.520
to have this really complex,
link |
00:14:24.240
manufacturable ecosystem
link |
00:14:26.120
where very sophisticated technologies can be devised
link |
00:14:30.600
and it works quite well.
link |
00:14:33.000
And amazingly, like you said,
link |
00:14:34.360
with a wavelength at like 100 nanometers
link |
00:14:36.240
or something like that,
link |
00:14:37.080
you're still able to achieve on this polymer
link |
00:14:39.720
precision of whatever we said, seven nanometers.
link |
00:14:43.440
I think I've heard like four nanometers
link |
00:14:45.720
being talked about, something like that.
link |
00:14:48.400
If we could just pause on this
link |
00:14:49.800
and we'll return to superconductivity,
link |
00:14:52.080
but in this whole journey from a history perspective,
link |
00:14:56.360
what do you think is the most beautiful
link |
00:14:59.440
at the intersection of engineering and physics
link |
00:15:01.680
to you in this whole process
link |
00:15:03.240
that we talked about with silicon and photolithography,
link |
00:15:06.520
things that people were able to achieve
link |
00:15:08.320
in order to push Moore's law forward?
link |
00:15:12.280
Is it the early days,
link |
00:15:13.640
the invention of the transistor itself?
link |
00:15:16.240
Is it some particular cool little thing
link |
00:15:19.280
that maybe not many people know about?
link |
00:15:21.960
Like, what do you think is most beautiful
link |
00:15:24.480
in this whole process, journey?
link |
00:15:26.760
The most beautiful is a little difficult to answer.
link |
00:15:29.560
Let me try and sidestep it a little bit
link |
00:15:32.040
and just say what strikes me about looking
link |
00:15:35.840
at the history of silicon microelectronics is that,
link |
00:15:42.000
so when quantum mechanics was developed,
link |
00:15:44.600
people quickly began applying it to semiconductors
link |
00:15:47.440
and it was broadly understood
link |
00:15:49.360
that these are fascinating systems
link |
00:15:50.760
and people cared about them for their basic physics,
link |
00:15:52.720
but also their utility as devices.
link |
00:15:55.040
And then the transistor was invented in the late forties
link |
00:15:59.280
in a relatively crude experimental setup
link |
00:16:02.080
where you just crammed a metal electrode
link |
00:16:04.280
into the semiconductor and that was ingenious.
link |
00:16:08.040
These people were able to make it work.
link |
00:16:13.000
But so what I wanna get to that really strikes me
link |
00:16:16.840
is that in those early days,
link |
00:16:19.320
there were a number of different semiconductors
link |
00:16:21.200
that were being considered.
link |
00:16:22.120
They had different properties, different strengths,
link |
00:16:23.840
different weaknesses.
link |
00:16:24.920
Most people thought germanium was the way to go.
link |
00:16:28.480
It had some nice properties related to things
link |
00:16:33.680
about how the electrons move inside the lattice.
link |
00:16:37.320
But other people thought that compound semiconductors
link |
00:16:39.800
with group three and group five also had
link |
00:16:42.000
really, really extraordinary properties
link |
00:16:46.240
that might be conducive to making the best devices.
link |
00:16:50.080
So there were different groups exploring each of these
link |
00:16:52.560
and that's great, that's how science works.
link |
00:16:54.240
You have to cast a broad net.
link |
00:16:56.160
But then what I find striking is why is it that silicon won?
link |
00:17:02.120
Because it's not that germanium is a useless material
link |
00:17:05.280
and it's not present in technology
link |
00:17:06.760
or compound semiconductors.
link |
00:17:08.080
They're both doing exciting and important things,
link |
00:17:12.640
slightly more niche applications
link |
00:17:14.400
whereas silicon is the semiconductor material
link |
00:17:18.080
for microelectronics which is the platform
link |
00:17:20.120
for digital computing which has transformed our world.
link |
00:17:22.760
Why did silicon win?
link |
00:17:24.200
It's because of a remarkable assemblage of qualities
link |
00:17:28.720
that no one of them was the clear winner
link |
00:17:32.120
but it made these sort of compromises
link |
00:17:34.560
between a number of different influences.
link |
00:17:36.680
It had that really excellent gate oxide
link |
00:17:40.520
that allowed us to make MOSFETs,
link |
00:17:43.240
these high performance transistors,
link |
00:17:45.400
so quickly and cheaply and easily
link |
00:17:47.200
without having to do a lot of materials development.
link |
00:17:49.360
The band gap of silicon is actually,
link |
00:17:53.400
so in a semiconductor there's an important parameter
link |
00:17:56.280
which is called the band gap
link |
00:17:57.480
which tells you there are sort of electrons
link |
00:18:00.600
that fill up to one level in the energy diagram
link |
00:18:04.600
and then there's a gap where electrons aren't allowed
link |
00:18:06.800
to have an energy in a certain range
link |
00:18:08.280
and then there's another energy level above that.
link |
00:18:11.320
And that difference between the lower sort of filled level
link |
00:18:14.960
and the unoccupied level,
link |
00:18:16.880
that tells you how much voltage you have to apply
link |
00:18:19.640
in order to induce a current to flow.
link |
00:18:22.160
So with germanium, that's about 0.75 electron volts.
link |
00:18:27.320
That means you have to apply 0.75 volts
link |
00:18:29.640
to get a current moving.
link |
00:18:32.000
And it turns out that if you compare that
link |
00:18:34.280
to the thermal excitations that are induced
link |
00:18:38.520
just by the temperature of our environment,
link |
00:18:40.680
that gap's not quite big enough.
link |
00:18:42.120
You start to use it to perform computations,
link |
00:18:45.120
it gets a little hot and you get all these accidental
link |
00:18:47.640
carriers that are excited into the conduction band
link |
00:18:50.720
and it causes errors in your computation.
link |
00:18:53.360
Silicon's band gap is just a little higher,
link |
00:18:56.200
1.1 electron volts,
link |
00:18:58.960
but you have an exponential dependence
link |
00:19:01.280
on the number of carriers that are present
link |
00:19:04.200
that can induce those errors.
link |
00:19:06.600
It decays exponentially with that voltage.
link |
00:19:08.480
So just that slight extra energy in that band gap
link |
00:19:12.760
really puts it in an ideal position to be operated
link |
00:19:17.040
in the conditions of our ambient environment.
link |
00:19:20.200
It's kind of fascinating that, like you mentioned,
link |
00:19:22.440
errors decrease exponentially with the voltage.
link |
00:19:27.500
So it's funny because this error thing comes up
link |
00:19:32.040
when you start talking about quantum computing.
link |
00:19:34.600
And it's kind of amazing that everything
link |
00:19:36.020
we've been talking about, the errors,
link |
00:19:37.920
as we scale down, seems to be extremely low.
link |
00:19:41.480
Yes.
link |
00:19:42.320
And like all of our computation is based
link |
00:19:45.960
on the assumption that it's extremely low.
link |
00:19:47.760
Yes, well it's digital computation.
link |
00:19:49.560
Digital, sorry, digital computation.
link |
00:19:51.480
So as opposed to our biological computation in our brain,
link |
00:19:55.000
is like the assumption is stuff is gonna fail
link |
00:19:58.240
all over the place and we somehow
link |
00:19:59.800
have to still be robust to that.
link |
00:20:01.720
That's exactly right.
link |
00:20:03.060
So this also, this is gonna be the most controversial part
link |
00:20:05.820
of our conversation where you're gonna make some enemies.
link |
00:20:07.800
So let me ask,
link |
00:20:09.040
because we've been talking about physics and engineering.
link |
00:20:14.040
Which group of people is smarter
link |
00:20:15.760
and more important for this one?
link |
00:20:17.800
Let me ask the question in a better way.
link |
00:20:20.560
Some of the big innovations,
link |
00:20:22.560
some of the beautiful things that we've been talking about,
link |
00:20:25.720
how much of it is physics?
link |
00:20:26.960
How much of it is engineering?
link |
00:20:28.440
My dad is a physicist and he talks down
link |
00:20:31.640
to all the amazing engineering that we're doing
link |
00:20:34.400
in the artificial intelligence and the computer science
link |
00:20:37.920
and the robotics and all that space.
link |
00:20:39.560
So we argue about this all the time.
link |
00:20:41.640
So what do you think?
link |
00:20:42.480
Who gets more credit?
link |
00:20:43.920
I'm genuinely not trying to just be politically correct here.
link |
00:20:46.760
I don't see how you would have any of the,
link |
00:20:50.480
what we consider sort of the great accomplishments
link |
00:20:52.600
of society without both.
link |
00:20:54.120
You absolutely need both of those things.
link |
00:20:55.880
Physics tends to play a key role earlier in the development
link |
00:20:59.640
and then engineering optimization, these things take over.
link |
00:21:04.640
And I mean, the invention of the transistor
link |
00:21:09.220
or actually even before that,
link |
00:21:10.840
the understanding of semiconductor physics
link |
00:21:12.840
that allowed the invention of the transistor,
link |
00:21:14.800
that's all physics.
link |
00:21:15.640
So if you didn't have that physics,
link |
00:21:17.000
you don't even get to get on the field.
link |
00:21:20.300
But once you have understood and demonstrated
link |
00:21:24.200
that this is in principle possible,
link |
00:21:26.520
more so as engineering.
link |
00:21:28.400
Why we have computers more powerful
link |
00:21:32.200
than old supercomputers in each of our phones,
link |
00:21:36.400
that's all engineering.
link |
00:21:37.520
And I think I would be quite foolish to say
link |
00:21:41.720
that that's not valuable, that's not a great contribution.
link |
00:21:46.920
It's a beautiful dance.
link |
00:21:47.800
Would you put like Silicon,
link |
00:21:49.740
the understanding of the material properties
link |
00:21:52.760
in the space of engineering?
link |
00:21:54.320
Like how does that whole process work?
link |
00:21:55.680
To understand that it has all these nice properties
link |
00:21:58.000
or even the development of photolithography,
link |
00:22:02.240
is that basically,
link |
00:22:03.840
would you put that in a category of engineering?
link |
00:22:06.140
No, I would say that it is basic physics,
link |
00:22:09.960
it is applied physics, it's material science,
link |
00:22:12.760
it's X ray crystallography, it's polymer chemistry,
link |
00:22:17.640
it's everything.
link |
00:22:18.800
Chemistry even is thrown in there?
link |
00:22:20.280
Absolutely, yes, absolutely.
link |
00:22:22.580
Just no biology.
link |
00:22:25.320
We can get to biology.
link |
00:22:26.560
Or the biologies and the humans
link |
00:22:28.240
that are engineering the system,
link |
00:22:29.480
so it's all integrated deeply.
link |
00:22:31.200
Okay, so let's return,
link |
00:22:32.560
you mentioned this word superconductivity.
link |
00:22:35.640
So what does that have to do with what we're talking about?
link |
00:22:38.600
Right, okay, so in a semiconductor,
link |
00:22:40.560
as I tried to describe a second ago,
link |
00:22:44.280
you can sort of induce currents by applying voltages
link |
00:22:50.040
and those have sort of typical properties
link |
00:22:52.280
that you would expect from some kind of a conductor.
link |
00:22:55.400
Those electrons, they don't just flow
link |
00:22:59.000
perfectly without dissipation.
link |
00:23:00.360
If an electron collides with an imperfection in the lattice
link |
00:23:03.240
or another electron, it's gonna slow down,
link |
00:23:05.600
it's gonna lose its momentum.
link |
00:23:06.880
So you have to keep applying that voltage
link |
00:23:09.280
in order to keep the current flowing.
link |
00:23:11.440
In a superconductor, something different happens.
link |
00:23:13.440
If you get a current to start flowing,
link |
00:23:16.520
it will continue to flow indefinitely.
link |
00:23:18.280
There's no dissipation.
link |
00:23:19.880
So that's crazy.
link |
00:23:21.660
How does that happen?
link |
00:23:22.500
Well, it happens at low temperature and this is crucial.
link |
00:23:26.800
It has to be a quite low temperature
link |
00:23:30.200
and what I'm talking about there,
link |
00:23:32.840
for essentially all of our conversation,
link |
00:23:35.800
I'm gonna be talking about conventional superconductors,
link |
00:23:39.800
sometimes called low TC superconductors,
link |
00:23:42.240
low critical temperature superconductors.
link |
00:23:45.120
And so those materials have to be at a temperature around,
link |
00:23:50.640
say around four Kelvin.
link |
00:23:51.960
I mean, their critical temperature might be 10 Kelvin,
link |
00:23:54.560
something like that,
link |
00:23:55.400
but you wanna operate them at around four Kelvin,
link |
00:23:57.120
four degrees above absolute zero.
link |
00:23:59.280
And what happens at that temperature,
link |
00:24:01.640
at very low temperatures in certain materials
link |
00:24:03.880
is that the noise of atoms moving around,
link |
00:24:10.080
the lattice vibrating, electrons colliding with each other,
link |
00:24:13.580
that becomes sufficiently low
link |
00:24:15.120
that the electrons can settle into this very special state.
link |
00:24:18.600
It's sometimes referred to as a macroscopic quantum state
link |
00:24:22.280
because if I had a piece of superconducting material here,
link |
00:24:26.560
let's say niobium is a very typical superconductor.
link |
00:24:30.920
If I had a block of niobium here
link |
00:24:33.640
and we cooled it below its critical temperature,
link |
00:24:36.700
all of the electrons in that superconducting state
link |
00:24:40.840
would be in one coherent quantum state.
link |
00:24:42.840
The wave function of that state is described
link |
00:24:47.520
in terms of all of the particles simultaneously,
link |
00:24:49.760
but it extends across macroscopic dimensions,
link |
00:24:52.220
the size of whatever block of that material
link |
00:24:56.120
I have sitting here.
link |
00:24:57.040
And the way this occurs is that,
link |
00:25:01.220
let's try to be a little bit light on the technical details,
link |
00:25:03.480
but essentially the electrons coordinate with each other.
link |
00:25:06.520
They are able to, in this macroscopic quantum state,
link |
00:25:10.300
they're able to sort of,
link |
00:25:12.260
one can quickly take the place of the other.
link |
00:25:14.400
You can't tell electrons apart.
link |
00:25:15.800
They're what's known as identical particles.
link |
00:25:17.740
So if this electron runs into a defect
link |
00:25:22.040
that would otherwise cause it to scatter,
link |
00:25:25.000
it can just sort of almost miraculously avoid that defect
link |
00:25:30.440
because it's not really in that location.
link |
00:25:32.240
It's part of a macroscopic quantum state
link |
00:25:34.000
and the entire quantum state
link |
00:25:35.660
was not scattered by that defect.
link |
00:25:37.160
So you can get a current that flows without dissipation
link |
00:25:40.780
and that's called a supercurrent.
link |
00:25:42.960
That's sort of just very much scratching the surface
link |
00:25:47.840
of superconductivity.
link |
00:25:49.840
There's very deep and rich physics there,
link |
00:25:52.240
just probably not the main subject
link |
00:25:54.520
we need to go into right now.
link |
00:25:55.640
But it turns out that when you have this material,
link |
00:26:00.520
you can do usual things like make wires out of it
link |
00:26:03.560
so you can get current to flow in a straight line on a chip,
link |
00:26:06.440
but you can also make other devices
link |
00:26:08.920
that perform different kinds of operations.
link |
00:26:11.880
Some of them are kind of logic operations
link |
00:26:14.760
like you'd get in a transistor.
link |
00:26:16.680
The most common or the most,
link |
00:26:21.200
I would say, diverse in its utility component
link |
00:26:25.480
is a Josephson junction.
link |
00:26:26.920
It's not analogous to a transistor
link |
00:26:28.980
in the sense that if you apply a voltage here,
link |
00:26:31.480
it changes how much current flows from left to right,
link |
00:26:33.880
but it is analogous in sort of a sense
link |
00:26:36.360
of it's the go to component
link |
00:26:39.200
that a circuit engineer is going to use
link |
00:26:42.040
to start to build up more complexity.
link |
00:26:44.520
So these junctions serve as gates.
link |
00:26:48.840
They can serve as gates.
link |
00:26:50.680
So I'm not sure how concerned to be with semantics,
link |
00:26:55.680
but let me just briefly say what a Josephson junction is
link |
00:26:58.880
and we can talk about different ways that they can be used.
link |
00:27:02.240
Basically, if you have a superconducting wire
link |
00:27:05.280
and then a small gap of a different material
link |
00:27:09.680
that's not superconducting, an insulator or normal metal,
link |
00:27:13.520
and then another superconducting wire on the other side,
link |
00:27:15.800
that's a Josephson junction.
link |
00:27:17.040
So it's sometimes referred to
link |
00:27:18.600
as a superconducting weak link.
link |
00:27:20.320
So you have this superconducting state on one side
link |
00:27:24.200
and on the other side, and the superconducting wave function
link |
00:27:27.520
actually tunnels across that gap.
link |
00:27:30.720
And when you create such a physical entity,
link |
00:27:35.460
it has very unusual current voltage characteristics.
link |
00:27:41.360
In that gap, like weird stuff happens.
link |
00:27:44.320
Through the entire circuit.
link |
00:27:45.160
So you can imagine, suppose you had a loop set up
link |
00:27:47.600
that had one of those weak links in the loop.
link |
00:27:51.240
Current would flow in that loop independent,
link |
00:27:53.800
even if you hadn't applied a voltage to it,
link |
00:27:55.640
and that's called the Josephson effect.
link |
00:27:57.060
So the fact that there's this phase difference
link |
00:28:00.520
in the quantum wave function from one side
link |
00:28:02.940
of the tunneling barrier to the other
link |
00:28:04.400
induces current to flow.
link |
00:28:05.720
So how does you change state?
link |
00:28:07.720
Right, exactly.
link |
00:28:08.560
So how do you change state?
link |
00:28:09.560
Now picture if I have a current bias coming down
link |
00:28:13.880
this line of my circuit and there's a Josephson junction
link |
00:28:16.120
right in the middle of it.
link |
00:28:18.120
And now I make another wire
link |
00:28:19.960
that goes around the Josephson junction.
link |
00:28:21.800
So I have a loop here, a superconducting loop.
link |
00:28:24.600
I can add current to that loop by exceeding
link |
00:28:28.960
the critical current of that Josephson junction.
link |
00:28:30.860
So like any superconducting material,
link |
00:28:34.960
it can carry this supercurrent that I've described,
link |
00:28:37.700
this current that can propagate without dissipation
link |
00:28:40.520
up to a certain level.
link |
00:28:41.840
And if you try and pass more current than that
link |
00:28:44.240
through the material, it's going to become
link |
00:28:47.520
a resistive material, normal material.
link |
00:28:51.140
So in the Josephson junction, the same thing happens.
link |
00:28:54.140
I can bias it above its critical current.
link |
00:28:57.120
And then what it's going to do,
link |
00:28:58.220
it's going to add a quantized amount of current
link |
00:29:03.560
into that loop.
link |
00:29:04.400
And what I mean by quantized is it's going to come
link |
00:29:07.320
in discrete packets with a well defined value of current.
link |
00:29:11.200
So in the vernacular of some people working
link |
00:29:15.320
in this community, you would say you pop a flux on
link |
00:29:19.520
into the loop.
link |
00:29:20.360
So a flux on.
link |
00:29:21.760
You pop a flux on into the loop.
link |
00:29:23.680
Yeah, so a flux on.
link |
00:29:24.520
Sounds like skateboarder talk, I love it.
link |
00:29:26.560
Okay, sorry, go ahead.
link |
00:29:28.960
A flux on is one of these quantized sort of amounts
link |
00:29:33.640
of current that you can add to a loop.
link |
00:29:35.220
And this is a cartoon picture,
link |
00:29:36.600
but I think it's sufficient for our purposes.
link |
00:29:38.400
So which, maybe it's useful to say,
link |
00:29:41.700
what is the speed at which these discrete packets
link |
00:29:45.480
of current travel?
link |
00:29:47.000
Because we'll be talking about light a little bit.
link |
00:29:49.160
It seems like the speed is important.
link |
00:29:51.080
The speed is important, that's an excellent question.
link |
00:29:53.560
Sometimes I wonder where you, how you became so astute.
link |
00:29:57.800
But so this.
link |
00:30:00.800
Matrix 4 is coming out, so maybe that's related.
link |
00:30:04.400
I'm not sure.
link |
00:30:05.240
I'm dressed for the job.
link |
00:30:06.400
I was trying to get to become an extra on Matrix 4,
link |
00:30:09.360
didn't work out.
link |
00:30:10.680
Anyway, so what's the speed of these packets?
link |
00:30:13.280
You'll have to find another gig.
link |
00:30:15.000
I know, I'm sorry.
link |
00:30:16.600
So the speed of the pack is actually these flux ons,
link |
00:30:19.440
these sort of pulses of current
link |
00:30:24.300
that are generated by Joseph's injunctions,
link |
00:30:26.200
they can actually propagate very close
link |
00:30:28.480
to the speed of light,
link |
00:30:29.740
maybe something like a third of the speed of light.
link |
00:30:31.920
That's quite fast.
link |
00:30:32.880
So one of the reasons why Joseph's injunctions are appealing
link |
00:30:37.280
is because their signals can propagate quite fast
link |
00:30:40.600
and they can also switch very fast.
link |
00:30:43.440
What I mean by switch is perform that operation
link |
00:30:46.080
that I described where you add current to the loop.
link |
00:30:49.440
That can happen within a few tens of picoseconds.
link |
00:30:53.960
So you can get devices that operate
link |
00:30:56.880
in the hundreds of gigahertz range.
link |
00:30:58.840
And by comparison, most processors
link |
00:31:02.080
in our conventional computers operate closer
link |
00:31:04.960
to the one gigahertz range, maybe three gigahertz
link |
00:31:08.360
seems to be kind of where those speeds have leveled out.
link |
00:31:12.960
The gamers listening to this are getting really excited
link |
00:31:15.600
to overclock their system to like, what is it?
link |
00:31:18.080
Like four gigahertz or something,
link |
00:31:19.560
a hundred sounds incredible.
link |
00:31:21.980
Can I just as a tiny tangent,
link |
00:31:24.060
is the physics of this understood well
link |
00:31:26.880
how to do this stably?
link |
00:31:28.520
Oh yes, the physics is understood well.
link |
00:31:30.160
The physics of Joseph's injunctions is understood well.
link |
00:31:32.540
The technology is understood quite well too.
link |
00:31:34.520
The reasons why it hasn't displaced
link |
00:31:37.620
silicon microelectronics in conventional digital computing
link |
00:31:41.720
I think are more related to what I was alluding to before
link |
00:31:45.040
about the myriad practical, almost mundane aspects
link |
00:31:49.240
of silicon that make it so useful.
link |
00:31:52.080
You can make a transistor ever smaller and smaller
link |
00:31:55.880
and it will still perform its digital function quite well.
link |
00:31:58.840
The same is not true of a Joseph's injunction.
link |
00:32:00.780
You really, they don't, they just,
link |
00:32:02.400
it's not the same thing that there's this feature
link |
00:32:04.440
that you can keep making smaller and smaller
link |
00:32:06.280
and it'll keep performing the same operations.
link |
00:32:08.220
This loop I described, any Joseph's in circuit,
link |
00:32:11.480
well, I wanna be careful, I shouldn't say
link |
00:32:13.600
any Joseph's in circuit, but many Joseph's in circuits,
link |
00:32:17.240
the way they process information
link |
00:32:19.440
or the way they perform whatever function it is
link |
00:32:21.280
they're trying to do,
link |
00:32:22.120
maybe it's sensing a weak magnetic field,
link |
00:32:24.560
it depends on an interplay between the junction
link |
00:32:27.480
and that loop.
link |
00:32:28.800
And you can't make that loop much smaller.
link |
00:32:30.560
And it's not for practical reasons
link |
00:32:32.120
that have to do with lithography.
link |
00:32:33.480
It's for fundamental physical reasons
link |
00:32:35.680
about the way the magnetic field interacts
link |
00:32:38.960
with that superconducting material.
link |
00:32:41.160
There are physical limits that no matter how good
link |
00:32:44.360
our technology got, those circuits would,
link |
00:32:47.260
I think would never be able to be scaled down
link |
00:32:50.360
to the densities that silicon microelectronics can.
link |
00:32:54.360
I don't know if we mentioned,
link |
00:32:55.560
is there something interesting
link |
00:32:56.960
about the various superconducting materials involved
link |
00:33:00.200
or is it all?
link |
00:33:01.040
There's a lot of stuff that's interesting.
link |
00:33:02.640
And it's not silicon.
link |
00:33:04.440
It's not silicon, no.
link |
00:33:05.840
So like it's some materials that also required
link |
00:33:09.520
to be super cold, four Kelvin and so on.
link |
00:33:12.560
So let's dissect a couple of those different things.
link |
00:33:15.280
The super cold part,
link |
00:33:16.280
let me just mention for your gamers out there
link |
00:33:19.640
that are trying to clock it at four gigahertz
link |
00:33:21.320
and would love to go to 400.
link |
00:33:22.160
What kind of cooling system can achieve four Kelvin?
link |
00:33:24.120
Four Kelvin, you need liquid helium.
link |
00:33:26.280
And so liquid helium is expensive.
link |
00:33:29.040
It's inconvenient.
link |
00:33:29.880
You need a cryostat that sits there
link |
00:33:32.080
and the energy consumption of that cryostat
link |
00:33:36.520
is impracticable for, it's not going in your cell phone.
link |
00:33:40.080
So you can picture holding your cell phone like this
link |
00:33:42.080
and then something the size of a keg of beer or something
link |
00:33:46.600
on your back to cool it.
link |
00:33:47.800
Like that makes no sense.
link |
00:33:49.520
So if you're trying to make this in consumer devices,
link |
00:33:54.120
electronics that are ubiquitous across society,
link |
00:33:57.000
superconductors are not in the race for that.
link |
00:33:59.280
For now, but you're saying,
link |
00:34:01.000
so just to frame the conversation,
link |
00:34:03.240
maybe the thing we're focused on
link |
00:34:05.520
is computing systems that serve as servers, like large.
link |
00:34:10.360
Yes, large systems.
link |
00:34:11.800
So then you can contrast what's going on in your cell phone
link |
00:34:14.800
with what's going on at one of the supercomputers.
link |
00:34:19.400
Colleague Katie Schuman invited us out to Oak Ridge
link |
00:34:22.080
a few years ago, so we got to see Titan
link |
00:34:24.080
and that was when they were building Summit.
link |
00:34:26.120
So these are some high performance supercomputers
link |
00:34:29.240
out in Tennessee and those are filling entire rooms
link |
00:34:32.480
the size of warehouses.
link |
00:34:33.920
So once you're at that level, okay,
link |
00:34:36.280
there you're already putting a lot of power into cooling.
link |
00:34:39.080
Cooling is part of your engineering task
link |
00:34:42.240
that you have to deal with.
link |
00:34:43.600
So there it's not entirely obvious
link |
00:34:45.600
that cooling to four Kelvin is out of the question.
link |
00:34:49.520
It has not happened yet and I can speak to why that is
link |
00:34:53.240
in the digital domain if you're interested.
link |
00:34:55.520
I think it's not going to happen.
link |
00:34:57.520
I don't think superconductors are gonna replace
link |
00:35:01.240
semiconductors for digital computation.
link |
00:35:05.880
There are a lot of reasons for that,
link |
00:35:07.640
but I think ultimately what it comes down to
link |
00:35:09.800
is all things considered cooling errors,
link |
00:35:13.440
scaling down to feature sizes, all that stuff,
link |
00:35:16.080
semiconductors work better at the system level.
link |
00:35:19.400
Is there some aspect of just curious
link |
00:35:22.720
about the historical momentum of this?
link |
00:35:25.560
Is there some power to the momentum of an industry
link |
00:35:28.200
that's mass manufacturing using a certain material?
link |
00:35:31.200
Is this like a Titanic shifting?
link |
00:35:33.680
Like what's your sense when a good idea comes along,
link |
00:35:37.120
how good does that idea need to be
link |
00:35:39.920
for the Titanic to start shifting?
link |
00:35:42.640
That's an excellent question.
link |
00:35:44.200
That's an excellent way to frame it.
link |
00:35:46.520
And you know, I don't know the answer to that,
link |
00:35:51.320
but what I think is, okay,
link |
00:35:53.600
so the history of the superconducting logic
link |
00:35:56.400
goes back to the 70s.
link |
00:35:58.000
IBM made a big push to do
link |
00:35:59.760
superconducting digital computing in the 70s.
link |
00:36:02.400
And they made some choices about their devices
link |
00:36:06.080
and their architectures and things that in hindsight,
link |
00:36:09.440
were kind of doomed to fail.
link |
00:36:11.000
And I don't mean any disrespect for the people that did it,
link |
00:36:13.120
it was hard to see at the time.
link |
00:36:14.280
But then another generation of superconducting logic
link |
00:36:17.880
was introduced, I wanna say the 90s,
link |
00:36:22.320
someone named Lykarev and Seminov,
link |
00:36:25.000
they proposed an entire family of circuits
link |
00:36:28.280
based on Joseph's injunctions
link |
00:36:29.920
that are doing digital computing based on logic gates
link |
00:36:33.440
and or not these kinds of things.
link |
00:36:37.920
And they showed how it could go hundreds of times faster
link |
00:36:41.560
than silicon microelectronics.
link |
00:36:43.200
And it's extremely exciting.
link |
00:36:45.360
I wasn't working in the field at that time,
link |
00:36:47.040
but later when I went back and read the literature,
link |
00:36:49.560
I was just like, wow, this is so awesome.
link |
00:36:53.040
And so you might think, well,
link |
00:36:56.000
the reason why it didn't display silicon
link |
00:36:58.280
is because silicon already had so much momentum
link |
00:37:00.400
at that time.
link |
00:37:01.720
But that was the 90s.
link |
00:37:02.960
Silicon kept that momentum
link |
00:37:04.320
because it had the simple way to keep getting better.
link |
00:37:06.960
You just make features smaller and smaller.
link |
00:37:08.720
So it would have to be,
link |
00:37:11.800
I don't think it would have to be that much better
link |
00:37:13.560
than silicon to displace it.
link |
00:37:15.440
But the problem is it's just not better than silicon.
link |
00:37:17.800
It might be better than silicon in one metric,
link |
00:37:19.960
speed of a switching operation
link |
00:37:21.440
or power consumption of a switching operation.
link |
00:37:24.400
But building a digital computer is a lot more
link |
00:37:26.680
than just that elemental operation.
link |
00:37:28.720
It's everything that goes into it,
link |
00:37:31.040
including the manufacturing, including the packaging,
link |
00:37:33.320
including the various materials aspects of things.
link |
00:37:38.840
So the reason why,
link |
00:37:40.600
and even in some of those early papers,
link |
00:37:42.800
I can't remember which one it was,
link |
00:37:44.120
Lykarev said something along the lines of,
link |
00:37:47.480
you can see how we could build an entire family
link |
00:37:49.920
of digital electronic circuits based on these components.
link |
00:37:52.800
They could go a hundred or more times faster
link |
00:37:55.000
than semiconductor logic gates.
link |
00:37:59.320
But I don't think that's the right way
link |
00:38:00.920
to use superconducting electronic circuits.
link |
00:38:02.680
He didn't say what the right way was,
link |
00:38:04.320
but he basically said digital logic,
link |
00:38:07.480
trying to steal the show from silicon
link |
00:38:11.280
is probably not what these circuits
link |
00:38:13.440
are most suited to accomplish.
link |
00:38:16.360
So if we can just linger and use the word computation.
link |
00:38:20.840
When you talk about computation, how do you think about it?
link |
00:38:24.080
Do you think purely on just the switching,
link |
00:38:28.920
or do you think something a little bit larger scale,
link |
00:38:31.320
a circuit taken together,
link |
00:38:32.720
performing the basic arithmetic operations
link |
00:38:36.940
that are then required to do the kind of computation
link |
00:38:40.400
that makes up a computer?
link |
00:38:42.160
Because when we talk about the speed of computation,
link |
00:38:44.400
is it boiled down to the basic switching,
link |
00:38:46.960
or is there some bigger picture
link |
00:38:48.400
that you're thinking about?
link |
00:38:49.240
Well, all right, so maybe we should disambiguate.
link |
00:38:52.240
There are a variety of different kinds of computation.
link |
00:38:55.600
I don't pretend to be an expert
link |
00:38:57.200
in the theory of computation or anything like that.
link |
00:39:00.180
I guess it's important to differentiate though
link |
00:39:02.640
between digital logic,
link |
00:39:05.800
which represents information as a series of bits,
link |
00:39:09.800
binary digits, which you can think of them
link |
00:39:13.000
as zeros and ones or whatever.
link |
00:39:14.160
Usually they correspond to a physical system
link |
00:39:17.400
that has two very well separated states.
link |
00:39:21.240
And then other kinds of computation,
link |
00:39:22.860
like we'll get into more the way your brain works,
link |
00:39:25.280
which it is, I think,
link |
00:39:27.600
indisputably processing information,
link |
00:39:30.400
but where the computation begins and ends
link |
00:39:34.080
is not anywhere near as well defined.
link |
00:39:36.360
It doesn't depend on these two levels.
link |
00:39:39.680
Here's a zero, here's a one.
link |
00:39:41.320
There's a lot of gray area
link |
00:39:42.640
that's usually referred to as analog computing.
link |
00:39:45.640
Also in conventional digital computers
link |
00:39:49.860
or digital computers in general,
link |
00:39:54.240
you have a concept of what's called arithmetic depth,
link |
00:39:57.300
which is jargon that basically means
link |
00:39:59.820
how many sequential operations are performed
link |
00:40:03.860
to turn an input into an output.
link |
00:40:07.680
And those kinds of computations in digital systems
link |
00:40:10.900
are highly serial, meaning that data streams,
link |
00:40:14.500
they don't branch off too far to the side.
link |
00:40:16.500
You do, you have to pull some information over there
link |
00:40:18.900
and access memory from here and stuff like that.
link |
00:40:20.900
But by and large, the computation proceeds
link |
00:40:24.340
in a serial manner.
link |
00:40:26.220
It's not that way in the brain.
link |
00:40:27.740
In the brain, you're always drawing information
link |
00:40:30.740
from different places.
link |
00:40:31.580
It's much more network based computing.
link |
00:40:33.820
Neurons don't wait for their turn.
link |
00:40:35.680
They fire when they're ready to fire.
link |
00:40:37.180
And so it's asynchronous.
link |
00:40:39.220
So one of the other things about a digital system
link |
00:40:41.680
is you're performing these operations on a clock.
link |
00:40:44.500
And that's a crucial aspect of it.
link |
00:40:46.700
Get rid of a clock in a digital system,
link |
00:40:48.900
nothing makes sense anymore.
link |
00:40:50.460
The brain has no clock.
link |
00:40:51.580
It builds its own timescales based on its internal activity.
link |
00:40:56.580
So you can think of the brain as kind of like this,
link |
00:40:59.500
like network computation,
link |
00:41:00.940
where it's actually really trivial, simple computers,
link |
00:41:05.700
just a huge number of them and they're networked.
link |
00:41:08.980
I would say it is complex, sophisticated little processors
link |
00:41:12.940
and there's a huge number of them.
link |
00:41:14.420
Neurons are not, are not simple.
link |
00:41:16.180
I don't mean to offend neurons.
link |
00:41:17.620
They're very complicated and beautiful and yeah,
link |
00:41:19.780
but we often oversimplify them.
link |
00:41:21.980
Yes, they're actually like there's computation happening
link |
00:41:24.820
within a neuron.
link |
00:41:25.660
Right, so I would say to think of a transistor
link |
00:41:29.520
as the building block of a digital computer is accurate.
link |
00:41:32.340
You use a few transistors to make your logic gates.
link |
00:41:34.660
You build up more, you build up processors
link |
00:41:37.060
from logic gates and things like that.
link |
00:41:39.140
So you can think of a transistor
link |
00:41:40.600
as a fundamental building block,
link |
00:41:42.300
or you can think of,
link |
00:41:43.380
as we get into more highly parallelized architectures,
link |
00:41:46.360
you can think of a processor
link |
00:41:47.700
as a fundamental building block.
link |
00:41:49.300
To make the analogy to the neuro side of things,
link |
00:41:53.180
a neuron is not a transistor.
link |
00:41:55.320
A neuron is a processor.
link |
00:41:57.300
It has synapses, even synapses are not transistors,
link |
00:42:00.220
but they are more,
link |
00:42:02.180
they're lower on the information processing hierarchy
link |
00:42:04.820
in a sense.
link |
00:42:05.660
They do a bulk of the computation,
link |
00:42:08.180
but neurons are entire processors in and of themselves
link |
00:42:13.580
that can take in many different kinds of inputs
link |
00:42:16.300
on many different spatial and temporal scales
link |
00:42:18.780
and produce many different kinds of outputs
link |
00:42:20.820
so that they can perform different computations
link |
00:42:23.820
in different contexts.
link |
00:42:24.860
So this is where enters this distinction
link |
00:42:27.440
between computation and communication.
link |
00:42:30.740
So you can think of neurons performing computation
link |
00:42:34.140
and the inter, the networking,
link |
00:42:36.580
the interconnectivity of neurons
link |
00:42:39.000
is communication between neurons.
link |
00:42:40.940
And you see this with very large server systems.
link |
00:42:43.500
I've been, I mentioned offline,
link |
00:42:45.020
we've been talking to Jim Keller,
link |
00:42:46.180
whose dream is to build giant computers
link |
00:42:48.140
that, you know, the bottom like there
link |
00:42:51.220
is often the communication
link |
00:42:52.380
between the different pieces of computing.
link |
00:42:54.700
So in this paper that we mentioned,
link |
00:42:57.380
Optoelectronic Intelligence,
link |
00:42:59.660
you say electrons excel at computation
link |
00:43:03.220
while light is excellent for communication.
link |
00:43:08.380
Maybe you can linger and say in this context,
link |
00:43:11.060
what do you mean by computation and communication?
link |
00:43:13.980
What are electrons, what is light
link |
00:43:17.420
and why do they excel at those two tasks?
link |
00:43:20.660
Yeah, just to first speak to computation
link |
00:43:23.620
versus communication,
link |
00:43:25.620
I would say computation is essentially taking in
link |
00:43:30.340
some information, performing operations
link |
00:43:33.860
on that information and producing new,
link |
00:43:37.220
hopefully more useful information.
link |
00:43:39.060
So for example, imagine you have a picture in front of you
link |
00:43:45.020
and there is a key in it
link |
00:43:48.020
and that's what you're looking for,
link |
00:43:48.940
for whatever reason, you wanna find the key,
link |
00:43:50.700
we all wanna find the key.
link |
00:43:51.580
So the input is that entire picture
link |
00:43:56.540
and the output might be the coordinates where the key is.
link |
00:43:59.060
So you've reduced the total amount of information you have
link |
00:44:01.540
but you found the useful information
link |
00:44:03.020
for you in that present moment,
link |
00:44:04.380
that's the useful information.
link |
00:44:05.220
And you think about this computation
link |
00:44:07.180
as the controlled synchronous sequential?
link |
00:44:10.820
Not necessarily, it could be,
link |
00:44:12.700
that could be how your system is performing the computation
link |
00:44:16.220
or it could be asynchronous,
link |
00:44:19.300
there are lots of ways to find the key.
link |
00:44:21.420
It depends on the nature of the data,
link |
00:44:23.700
it depends on, that's a very simplified example,
link |
00:44:27.540
a picture with a key in it,
link |
00:44:28.700
what about if you're in the world
link |
00:44:30.540
and you're trying to decide the best way
link |
00:44:32.500
to live your life?
link |
00:44:35.940
It might be interactive,
link |
00:44:37.020
it might be there might be some recurrence
link |
00:44:38.580
or some weird asynchrony, I got it.
link |
00:44:41.340
But there's an input and there's an output
link |
00:44:43.260
and you do some stuff in the middle
link |
00:44:44.460
that actually goes from the input to the output.
link |
00:44:46.020
You've taken in information
link |
00:44:47.340
and output different information,
link |
00:44:49.100
hopefully reducing the total amount of information
link |
00:44:51.820
and extracting what's useful.
link |
00:44:53.820
Communication is then getting that information
link |
00:44:57.780
from the location at which it's stored
link |
00:44:59.460
because information is physical as Landauer emphasized
link |
00:45:02.660
and so it is in one place
link |
00:45:04.940
and you need to get that information to another place
link |
00:45:07.860
so that something else can use it
link |
00:45:10.100
for whatever computation it's working on.
link |
00:45:12.020
Maybe it's part of the same network
link |
00:45:13.460
and you're all trying to solve the same problem
link |
00:45:15.020
but neuron A over here just deduced something
link |
00:45:20.500
based on its inputs
link |
00:45:21.620
and it's now sending that information across the network
link |
00:45:25.180
to another location
link |
00:45:26.460
so that would be the act of communication.
link |
00:45:28.420
Can you linger on Landauer
link |
00:45:29.900
and saying information is physical?
link |
00:45:31.820
Rolf Landauer, not to be confused with Lev Landauer.
link |
00:45:35.340
Yeah, and he made huge contributions
link |
00:45:38.020
to our understanding of the reversibility of information
link |
00:45:42.980
and this concept that energy has to be dissipated
link |
00:45:46.700
in computing when the computation is irreversible
link |
00:45:50.100
but if you can manage to make it reversible
link |
00:45:52.140
then you don't need to expend energy
link |
00:45:55.060
but if you do expend energy to perform a computation
link |
00:45:59.660
there's sort of a minimal amount that you have to do
link |
00:46:02.300
and it's KT log two.
link |
00:46:04.460
And it's all somehow related
link |
00:46:05.900
to the second law of thermodynamics
link |
00:46:07.660
and that the universe is an information process
link |
00:46:09.620
and then we're living in a simulation.
link |
00:46:11.500
So okay, sorry, sorry for that tangent.
link |
00:46:13.980
So that's the defining the distinction
link |
00:46:17.140
between computation and communication.
link |
00:46:19.580
Let me say one more thing just to clarify.
link |
00:46:21.940
Communication ideally does not change the information.
link |
00:46:27.100
It moves it from one place to another
link |
00:46:28.900
but it is preserved.
link |
00:46:30.940
Got it, okay.
link |
00:46:32.500
All right, that's beautiful.
link |
00:46:33.700
So then the electron versus light distinction
link |
00:46:38.620
and why are electrons good at computation
link |
00:46:42.380
and light good at communication?
link |
00:46:44.540
Yes, there's a lot that goes into it I guess
link |
00:46:48.820
but just try to speak to the simplest part of it.
link |
00:46:54.100
Electrons interact strongly with one another.
link |
00:46:56.980
They're charged particles.
link |
00:46:58.340
So if I pile a bunch of them over here
link |
00:47:02.020
they're feeling a certain amount of force
link |
00:47:03.860
and they wanna move somewhere else.
link |
00:47:05.700
They're strongly interactive.
link |
00:47:06.900
You can also get them to sit still.
link |
00:47:08.900
You can, an electron has a mass
link |
00:47:10.660
so you can cause it to be spatially localized.
link |
00:47:15.860
So for computation that's useful
link |
00:47:18.100
because now I can make these little devices
link |
00:47:20.140
that put a bunch of electrons over here
link |
00:47:21.940
and then I change the state of a gate
link |
00:47:25.620
like I've been describing,
link |
00:47:26.500
put a different voltage on this gate
link |
00:47:28.380
and now I move the electrons over here.
link |
00:47:29.980
Now they're sitting somewhere else.
link |
00:47:31.220
I have a physical mechanism
link |
00:47:33.980
with which I can represent information.
link |
00:47:36.020
It's spatially localized and I have knobs
link |
00:47:38.140
that I can adjust to change where those electrons are
link |
00:47:41.220
or what they're doing.
link |
00:47:42.380
Light by contrast, photons of light
link |
00:47:45.220
which are the discrete packets of energy
link |
00:47:48.100
that were identified by Einstein,
link |
00:47:50.740
they do not interact with each other
link |
00:47:54.540
especially at low light levels.
link |
00:47:56.260
If you're in a medium and you have a bright high light level
link |
00:48:00.380
you can get them to interact with each other
link |
00:48:02.540
through the interaction with that medium that they're in
link |
00:48:05.340
but that's a little bit more exotic.
link |
00:48:07.780
And for the purposes of this conversation
link |
00:48:10.340
we can assume that photons don't interact with each other.
link |
00:48:13.180
So if you have a bunch of them
link |
00:48:16.100
all propagating in the same direction
link |
00:48:17.580
they don't interfere with each other.
link |
00:48:19.140
If I wanna send, if I have a communication channel
link |
00:48:22.900
and I put one more photon on it,
link |
00:48:24.460
it doesn't screw up with those other ones.
link |
00:48:26.020
It doesn't change what those other ones were doing at all.
link |
00:48:29.060
So that's really useful for communication
link |
00:48:31.260
because that means you can sort of allow
link |
00:48:33.660
a lot of these photons to flow
link |
00:48:37.060
without disruption of each other
link |
00:48:38.820
and they can branch really easily and things like that.
link |
00:48:41.060
But it's not good for computation
link |
00:48:42.700
because it's very hard for this packet of light
link |
00:48:46.340
to change what this packet of light is doing.
link |
00:48:48.700
They pass right through each other.
link |
00:48:50.180
So in computation you want to change information
link |
00:48:53.260
and if photons don't interact with each other
link |
00:48:55.700
it's difficult to get them to change the information
link |
00:48:58.020
represented by the others.
link |
00:48:59.380
So that's the fundamental difference.
link |
00:49:01.580
Is there also something about the way they travel
link |
00:49:04.780
through different materials
link |
00:49:07.460
or is that just a particular engineering?
link |
00:49:10.700
No, it's not, that's deep physics I think.
link |
00:49:12.580
So this gets back to electrons interact with each other
link |
00:49:17.060
and photons don't.
link |
00:49:18.140
So say I'm trying to get a packet of information
link |
00:49:22.380
from me to you and we have a wire going between us.
link |
00:49:25.820
In order for me to send electrons across that wire
link |
00:49:29.020
I first have to raise the voltage on my end of the wire
link |
00:49:32.180
and that means putting a bunch of charges on it
link |
00:49:34.580
and then that charge packet has to propagate along the wire
link |
00:49:39.140
and it has to get all the way over to you.
link |
00:49:41.260
That wire is gonna have something that's called capacitance
link |
00:49:44.380
which basically tells you how much charge
link |
00:49:46.940
you need to put on the wire
link |
00:49:48.060
in order to raise the voltage on it
link |
00:49:49.980
and the capacitance is gonna be proportional
link |
00:49:52.500
to the length of the wire.
link |
00:49:54.060
So the longer the length of the wire is
link |
00:49:56.900
the more charge I have to put on it
link |
00:49:59.140
and the energy required to charge up that line
link |
00:50:03.060
and move those electrons to you
link |
00:50:04.980
is also proportional to the capacitance
link |
00:50:06.860
and goes as the voltage squared.
link |
00:50:08.500
So you get this huge penalty if you wanna send electrons
link |
00:50:13.780
across a wire over appreciable distances.
link |
00:50:16.620
So distance is an important thing here
link |
00:50:19.140
when you're doing communication.
link |
00:50:20.780
Distance is an important thing.
link |
00:50:22.140
So is the number of connections I'm trying to make.
link |
00:50:25.340
Me to you, okay one, that's not so bad.
link |
00:50:27.620
If I want to now send it to 10,000 other friends
link |
00:50:31.420
then all of those wires are adding tons
link |
00:50:34.380
of extra capacitance.
link |
00:50:35.460
Now not only does it take forever
link |
00:50:37.660
to put the charge on that wire
link |
00:50:39.540
and raise the voltage on all those lines
link |
00:50:41.820
but it takes a ton of power
link |
00:50:43.540
and the number 10,000 is not randomly chosen.
link |
00:50:46.980
That's roughly how many connections
link |
00:50:49.100
each neuron in your brain makes.
link |
00:50:50.620
So a neuron in your brain needs to send 10,000 messages
link |
00:50:55.020
every time it has something to say.
link |
00:50:56.780
You can't do that if you're trying to drive electrons
link |
00:51:00.100
from here to 10,000 different places.
link |
00:51:02.060
The brain does it in a slightly different way
link |
00:51:03.660
which we can discuss.
link |
00:51:04.860
How can light achieve the 10,000 connections
link |
00:51:07.020
and why is it better?
link |
00:51:09.340
In terms of like the energy use required
link |
00:51:12.580
to use light for the communication of the 10,000 connections.
link |
00:51:15.260
Right, right.
link |
00:51:16.100
So now instead of trying to send electrons
link |
00:51:17.700
from me to you, I'm trying to send photons.
link |
00:51:19.380
So I can make what's called a wave guide
link |
00:51:21.540
which is just a simple piece of a material.
link |
00:51:25.140
It could be glass like an optical fiber
link |
00:51:27.060
or silicon on a chip.
link |
00:51:29.860
And I just have to inject photons into that wave guide
link |
00:51:34.140
and independent of how long it is,
link |
00:51:35.800
independent of how many different connections I'm making,
link |
00:51:39.620
it doesn't change the voltage or anything like that
link |
00:51:43.140
that I have to raise up on the wire.
link |
00:51:45.460
So if I have one more connection,
link |
00:51:47.940
if I add additional connections,
link |
00:51:49.820
I need to add more light to the wave guide
link |
00:51:51.760
because those photons need to split
link |
00:51:53.280
and go to different paths.
link |
00:51:55.040
That makes sense but I don't have a capacitive penalty.
link |
00:51:58.860
Sometimes these are called wiring parasitics.
link |
00:52:01.300
There are no parasitics associated with light
link |
00:52:03.460
in that same sense.
link |
00:52:04.420
So this might be a dumb question
link |
00:52:07.500
but how do I catch a photon on the other end?
link |
00:52:11.380
Is it material?
link |
00:52:12.500
Is it the polymer stuff you were talking about
link |
00:52:15.020
for a different application for photolithography?
link |
00:52:19.380
Like how do you catch a photon?
link |
00:52:20.980
There's a lot of ways to catch a photon.
link |
00:52:22.500
It's not a dumb question.
link |
00:52:23.620
It's a deep and important question
link |
00:52:25.940
that basically defines a lot of the work
link |
00:52:29.180
that goes on in our group at NIST.
link |
00:52:31.380
One of my group leaders, Seywoon Nam,
link |
00:52:34.260
has built his career around
link |
00:52:35.780
these superconducting single photon detectors.
link |
00:52:38.420
So if you're going to try to sort of reach a lower limit
link |
00:52:42.460
and detect just one particle of light,
link |
00:52:45.260
superconductors come back into our conversation
link |
00:52:47.660
and just picture a simple device
link |
00:52:50.140
where you have current flowing
link |
00:52:51.600
through a superconducting wire and...
link |
00:52:54.580
A loop again or no?
link |
00:52:56.560
Let's say yes, you have a loop.
link |
00:52:57.820
So you have a superconducting wire
link |
00:52:59.580
that goes straight down like this
link |
00:53:01.000
and on your loop branch, you have a little ammeter,
link |
00:53:04.260
something that measures current.
link |
00:53:05.800
There's a resistor up there too.
link |
00:53:07.980
Go with me here.
link |
00:53:09.020
So your current biasing this,
link |
00:53:12.020
so there's current flowing
link |
00:53:13.060
through that superconducting branch.
link |
00:53:14.400
Since there's a resistor over here,
link |
00:53:16.560
all the current goes through the superconducting branch.
link |
00:53:18.940
Now a photon comes in, strikes that superconductor.
link |
00:53:22.280
We talked about this superconducting
link |
00:53:24.300
macroscopic quantum state.
link |
00:53:25.760
That's going to be destroyed by the energy of that photon.
link |
00:53:28.480
So now that branch of the circuit is resistive too.
link |
00:53:32.080
And you've properly designed your circuit
link |
00:53:33.780
so that the resistance on that superconducting branch
link |
00:53:36.700
is much greater than the other resistance.
link |
00:53:38.420
Now all of your current's going to go that way.
link |
00:53:40.780
Your ammeter says, oh, I just got a pulse of current.
link |
00:53:43.260
That must mean I detected a photon.
link |
00:53:45.140
Then where you broke that superconductivity
link |
00:53:47.220
in a matter of a few nanoseconds,
link |
00:53:49.100
it cools back off, dissipates that energy
link |
00:53:51.100
and the current flows back
link |
00:53:52.820
through that superconducting branch.
link |
00:53:54.300
This is a very powerful superconducting device
link |
00:53:59.260
that allows us to understand quantum states of light.
link |
00:54:02.300
I didn't realize a loop like that
link |
00:54:04.880
could be sensitive to a single photon.
link |
00:54:07.160
I mean, that seems strange to me because,
link |
00:54:13.060
I mean, so what happens when you just barrage it
link |
00:54:15.700
with photons?
link |
00:54:16.620
If you put a bunch of photons in there,
link |
00:54:18.340
essentially the same thing happens.
link |
00:54:19.860
You just drive it into the normal state,
link |
00:54:21.660
it becomes resistive and it's not particularly interesting.
link |
00:54:25.420
So you have to be careful how many photons you send.
link |
00:54:27.980
Like you have to be very precise with your communication.
link |
00:54:30.100
Well, it depends.
link |
00:54:31.220
So I would say that that's actually in the application
link |
00:54:34.220
that we're trying to use these detectors for.
link |
00:54:37.060
That's a feature because what we want is for,
link |
00:54:41.140
if a neuron sends one photon to a synaptic connection
link |
00:54:46.660
and one of these superconducting detectors is sitting there,
link |
00:54:49.860
you get this pulse of current.
link |
00:54:51.200
And that synapse says event,
link |
00:54:54.000
then I'm gonna do what I do when there's a synapse event,
link |
00:54:56.020
I'm gonna perform computations, that kind of thing.
link |
00:54:58.660
But if accidentally you send two there or three or five,
link |
00:55:02.100
it does the exact same.
link |
00:55:03.380
Got it.
link |
00:55:04.200
And so this is how in the system that we're devising here,
link |
00:55:10.040
communication is entirely binary.
link |
00:55:12.740
And that's what I tried to emphasize a second ago.
link |
00:55:15.020
Communication should not change the information.
link |
00:55:17.780
You're not saying, oh, I got this kind of communication
link |
00:55:21.300
event for photons.
link |
00:55:22.280
No, we're not keeping track of that.
link |
00:55:23.700
This neuron fired, this synapse says that neuron fired,
link |
00:55:26.580
that's it.
link |
00:55:27.420
So that's a noise filtering property of those detectors.
link |
00:55:31.460
However, there are other applications
link |
00:55:33.140
where you'd rather know the exact number of photons
link |
00:55:36.140
that can be very useful in quantum computing with light.
link |
00:55:39.300
And our group does a lot of work
link |
00:55:41.820
around another kind of superconducting sensor
link |
00:55:44.580
called a transition edge sensor that Adrian Alita
link |
00:55:48.260
in our group does a lot of work on that.
link |
00:55:49.940
And that can tell you based on the amplitude
link |
00:55:53.980
of the current pulse you divert exactly how many photons
link |
00:55:58.180
were in that pulse.
link |
00:56:00.900
What's that useful for?
link |
00:56:02.500
One way that you can encode information
link |
00:56:04.700
in quantum states of light is in the number of photons.
link |
00:56:07.460
You can have what are called number states
link |
00:56:09.300
and a number state will have a well defined number
link |
00:56:11.980
of photons and maybe the output of your quantum computation
link |
00:56:16.400
encodes its information in the number of photons
link |
00:56:19.780
that are generated.
link |
00:56:20.620
So if you have a detector that is sensitive to that,
link |
00:56:23.020
it's extremely useful.
link |
00:56:24.300
Can you achieve like a clock with photons
link |
00:56:29.240
or is that not important?
link |
00:56:30.260
Is there a synchronicity here?
link |
00:56:33.300
In general, it can be important.
link |
00:56:36.880
Clock distribution is a big challenge
link |
00:56:39.380
in especially large computational systems.
link |
00:56:43.300
And so yes, optical clocks, optical clock distribution
link |
00:56:47.940
is a very powerful technology.
link |
00:56:51.140
I don't know the state of that field right now,
link |
00:56:53.180
but I imagine that if you're trying to distribute a clock
link |
00:56:55.620
across any appreciable size computational system,
link |
00:56:58.960
you wanna use light.
link |
00:57:00.340
Yeah, I wonder how these giant systems work,
link |
00:57:04.300
especially like supercomputers.
link |
00:57:07.380
Do they need to do clock distribution
link |
00:57:09.380
or are they doing more ad hoc parallel
link |
00:57:14.260
like concurrent programming?
link |
00:57:15.540
Like there's some kind of locking mechanisms or something.
link |
00:57:18.140
That's a fascinating question,
link |
00:57:19.320
but let's zoom in at this very particular question
link |
00:57:23.900
of computation on a processor
link |
00:57:28.200
and communication between processors.
link |
00:57:31.560
So what does this system look like
link |
00:57:36.440
that you're envisioning?
link |
00:57:38.220
One of the places you're envisioning it
link |
00:57:40.100
is in the paper on optoelectronic intelligence.
link |
00:57:43.140
So what are we talking about?
link |
00:57:44.740
Are we talking about something
link |
00:57:46.300
that starts to look a lot like the human brain
link |
00:57:48.740
or does it still look a lot like a computer?
link |
00:57:51.300
What are the size of this thing?
link |
00:57:52.980
Is it going inside a smartphone or as you said,
link |
00:57:55.140
does it go inside something that's more like a house?
link |
00:57:58.580
Like what should we be imagining?
link |
00:58:01.180
What are you thinking about
link |
00:58:02.260
when you're thinking about these fundamental systems?
link |
00:58:05.460
Let me introduce the word neuromorphic.
link |
00:58:07.380
There's this concept of neuromorphic computing
link |
00:58:09.960
where what that broadly refers to
link |
00:58:12.580
is computing based on the information processing principles
link |
00:58:17.740
of the brain.
link |
00:58:19.260
And as digital computing seems to be pushing
link |
00:58:23.820
towards some fundamental performance limits,
link |
00:58:26.480
people are considering architectural advances,
link |
00:58:29.180
drawing inspiration from the brain,
link |
00:58:30.860
more distributed parallel network kind of architectures
link |
00:58:33.660
and stuff.
link |
00:58:34.500
And so there's this continuum of neuromorphic
link |
00:58:37.420
from things that are pretty similar to digital computers,
link |
00:58:42.740
but maybe there are more cores
link |
00:58:45.720
and the way they send messages is a little bit more
link |
00:58:49.180
like the way brain neurons send spikes.
link |
00:58:52.780
But for the most part, it's still digital electronics.
link |
00:58:56.100
And then you have some things in between
link |
00:58:58.700
where maybe you're using transistors,
link |
00:59:02.060
but now you're starting to use them
link |
00:59:03.220
instead of in a digital way, in an analog way.
link |
00:59:06.140
And so you're trying to get those circuits
link |
00:59:08.180
to behave more like neurons.
link |
00:59:10.320
And then that's a little bit,
link |
00:59:12.140
quite a bit more on the neuromorphic side of things.
link |
00:59:14.940
You're trying to get your circuits,
link |
00:59:17.080
although they're still based on silicon,
link |
00:59:19.140
you're trying to get them to perform operations
link |
00:59:22.480
that are highly analogous to the operations in the brain.
link |
00:59:24.660
And that's where a great deal of work is
link |
00:59:26.700
in neuromorphic computing,
link |
00:59:27.700
people like Giacomo Indoveri and Gert Kauenberg,
link |
00:59:30.920
Jennifer Hasler, countless others.
link |
00:59:32.980
It's a rich and exciting field going back to Carver Mead
link |
00:59:36.900
in the late 1980s.
link |
00:59:39.460
And then all the way on the other extreme of the continuum
link |
00:59:44.240
is where you say, I'll give up anything related
link |
00:59:48.580
to transistors or semiconductors or anything like that.
link |
00:59:51.420
I'm not starting with the assumption
link |
00:59:53.620
that I'm gonna use any kind
link |
00:59:55.060
of conventional computing hardware.
link |
00:59:57.020
And instead, what I wanna do is try and understand
link |
00:59:59.300
what makes the brain powerful
link |
01:00:00.700
at the kind of information processing it does.
link |
01:00:03.180
And I wanna think from first principles
link |
01:00:05.540
about what hardware is best going to enable us
link |
01:00:10.580
to capture those information processing principles
link |
01:00:14.020
in an artificial system.
link |
01:00:16.020
And that's where I live.
link |
01:00:17.420
That's where I'm doing my exploration these days.
link |
01:00:21.740
So what are the first principles
link |
01:00:25.800
of brain like computation communication?
link |
01:00:29.960
Right, yeah, this is so important
link |
01:00:32.580
and I'm glad we booked 14 hours for this because.
link |
01:00:35.540
I only have 13, I'm sorry.
link |
01:00:38.380
Okay, so the brain is notoriously complicated.
link |
01:00:41.500
And I think that's an important part
link |
01:00:44.060
of why it can do what it does.
link |
01:00:46.300
But okay, let me try to break it down.
link |
01:00:49.620
Starting with the devices, neurons, as I said before,
link |
01:00:54.580
they're sophisticated devices in and of themselves
link |
01:00:57.100
and synapses are too.
link |
01:00:58.220
They can change their state based on the activity.
link |
01:01:03.060
So they adapt over time.
link |
01:01:04.900
That's crucial to the way the brain works.
link |
01:01:06.980
They don't just adapt on one timescale,
link |
01:01:09.380
they can adapt on myriad timescales
link |
01:01:12.460
from the spacing between pulses,
link |
01:01:16.060
the spacing between spikes that come from neurons
link |
01:01:18.700
all the way to the age of the organism.
link |
01:01:23.100
Also relevant, perhaps I think the most important thing
link |
01:01:28.100
that's guided my thinking is the network structure
link |
01:01:32.320
of the brain, so.
link |
01:01:33.880
Which can also be adjusted on different scales.
link |
01:01:36.600
Absolutely, yes, so you're making new,
link |
01:01:39.440
you're changing the strength of contacts,
link |
01:01:41.360
you're changing the spatial distribution of them,
link |
01:01:44.120
although spatial distribution doesn't change that much
link |
01:01:46.900
once you're a mature organism.
link |
01:01:49.400
But that network structure is really crucial.
link |
01:01:52.880
So let me dwell on that for a second.
link |
01:01:55.400
You can't talk about the brain without emphasizing
link |
01:01:58.960
that most of the neurons in the neocortex
link |
01:02:02.880
or the prefrontal cortex, the part of the brain
link |
01:02:04.840
that we think is most responsible for high level reasoning
link |
01:02:08.080
and things like that,
link |
01:02:09.080
those neurons make thousands of connections.
link |
01:02:11.400
So you have this network that is highly interconnected.
link |
01:02:15.560
And I think it's safe to say that one of the primary reasons
link |
01:02:19.880
that they make so many different connections
link |
01:02:23.180
is that allows information to be communicated very rapidly
link |
01:02:26.880
from any spot in the network
link |
01:02:28.420
to any other spot in the network.
link |
01:02:30.320
So that's a sort of spatial aspect of it.
link |
01:02:33.920
You can quantify this in terms of concepts
link |
01:02:38.480
that are related to fractals and scale invariants,
link |
01:02:41.020
which I think is a very beautiful concept.
link |
01:02:43.240
So what I mean by that is kind of,
link |
01:02:46.200
no matter what spatial scale you're looking at in the brain
link |
01:02:50.720
within certain bounds, you see the same
link |
01:02:53.520
general statistical pattern.
link |
01:02:54.960
So if I draw a box around some region of my cortex,
link |
01:02:59.280
most of the connections that those neurons
link |
01:03:02.280
within that box make are gonna be within the box
link |
01:03:04.480
to each other in their local neighborhood.
link |
01:03:06.200
And that's sort of called clustering, loosely speaking.
link |
01:03:09.280
But a non negligible fraction
link |
01:03:10.920
is gonna go outside of that box.
link |
01:03:12.600
And then if I draw a bigger box,
link |
01:03:14.080
the pattern is gonna be exactly the same.
link |
01:03:16.400
So you have this scale invariants,
link |
01:03:18.400
and you also have a non vanishing probability
link |
01:03:22.720
of a neuron making connection very far away.
link |
01:03:25.400
So suppose you wanna plot the probability
link |
01:03:28.720
of a neuron making a connection as a function of distance.
link |
01:03:32.420
If that were an exponential function,
link |
01:03:34.240
it would go e to the minus radius
link |
01:03:36.720
over some characteristic radius,
link |
01:03:38.700
and it would drop off up to some certain radius,
link |
01:03:41.840
the probability would be reasonably close to one,
link |
01:03:44.920
and then beyond that characteristic length R zero,
link |
01:03:49.120
it would drop off sharply.
link |
01:03:51.280
And so that would mean that the neurons in your brain
link |
01:03:53.560
are really localized, and that's not what we observe.
link |
01:03:58.640
Instead, what you see is that the probability
link |
01:04:00.680
of making a longer distance connection, it does drop off,
link |
01:04:03.760
but it drops off as a power law.
link |
01:04:05.760
So the probability that you're gonna have a connection
link |
01:04:08.420
at some radius R goes as R to the minus some power.
link |
01:04:13.180
And that's more, that's what we see with forces in nature,
link |
01:04:16.800
like the electromagnetic force
link |
01:04:18.440
between two particles or gravity
link |
01:04:20.440
goes as one over the radius squared.
link |
01:04:23.000
So you can see this in fractals.
link |
01:04:24.420
I love that there's like a fractal dynamics of the brain
link |
01:04:28.600
that if you zoom out, you draw the box
link |
01:04:31.460
and you increase that box by certain step sizes,
link |
01:04:35.040
you're gonna see the same statistics.
link |
01:04:36.720
I think that's probably very important
link |
01:04:40.000
to the way the brain processes information.
link |
01:04:41.880
It's not just in the spatial domain,
link |
01:04:43.600
it's also in the temporal domain.
link |
01:04:45.640
And what I mean by that is...
link |
01:04:48.640
That's incredible that this emerged
link |
01:04:50.640
through the evolutionary process
link |
01:04:52.320
that potentially somehow connected
link |
01:04:54.880
to the way the physics of the universe works.
link |
01:04:57.800
Yeah, I couldn't agree more that it's a deep
link |
01:05:00.360
and fascinating subject that I hope to be able
link |
01:05:02.720
to spend the rest of my life studying.
link |
01:05:04.160
You think you need to solve, understand this,
link |
01:05:07.280
this fractal nature in order to understand intelligence
link |
01:05:10.120
and communication. I do think so.
link |
01:05:11.520
I think they're deeply intertwined.
link |
01:05:13.320
Yes, I think power laws are right at the heart of it.
link |
01:05:16.880
So just to push that one through,
link |
01:05:19.400
the same thing happens in the temporal domain.
link |
01:05:21.480
So suppose your neurons in your brain
link |
01:05:26.000
were always oscillating at the same frequency,
link |
01:05:28.160
then the probability of finding a neuron oscillating
link |
01:05:31.320
as a function of frequency
link |
01:05:32.520
would be this narrowly peaked function
link |
01:05:34.520
around that certain characteristic frequency.
link |
01:05:36.520
That's not at all what we see.
link |
01:05:37.880
The probability of finding neurons oscillating
link |
01:05:40.240
or producing spikes at a certain frequency
link |
01:05:43.880
is again a power law,
link |
01:05:45.200
which means there's no defined scale
link |
01:05:49.640
of the temporal activity in the brain.
link |
01:05:53.560
At what speed do your thoughts occur?
link |
01:05:56.040
Well, there's a fastest speed they can occur
link |
01:05:58.280
and that is limited by communication and other things,
link |
01:06:01.520
but there's not a characteristic scale.
link |
01:06:03.960
We have thoughts on all temporal scales
link |
01:06:06.880
from a few tens of milliseconds,
link |
01:06:10.800
which is physiologically limited by our devices,
link |
01:06:13.360
compare that to tens of picoseconds
link |
01:06:15.720
that I talked about in superconductors,
link |
01:06:17.120
all the way up to the lifetime of the organism.
link |
01:06:19.240
You can still think about things
link |
01:06:20.720
that happened to you when you were a kid.
link |
01:06:22.560
Or if you wanna be really trippy
link |
01:06:24.040
then across multiple organisms
link |
01:06:25.840
in the entirety of human civilization,
link |
01:06:27.440
you have thoughts that span organisms, right?
link |
01:06:29.400
Yes, taking it to that level, yes.
link |
01:06:31.200
If you're willing to see the entirety of the human species
link |
01:06:34.600
as a single organism with a collective intelligence
link |
01:06:37.160
and that too on a spatial and temporal scale,
link |
01:06:39.880
there's thoughts occurring.
link |
01:06:41.080
And then if you look at not just the human species,
link |
01:06:44.000
but the entirety of life on earth
link |
01:06:46.440
as an organism with thoughts that are occurring,
link |
01:06:49.600
that are greater and greater sophisticated thoughts,
link |
01:06:51.600
there's a different spatial and temporal scale there.
link |
01:06:54.640
This is getting very suspicious.
link |
01:06:57.200
Well, hold on though, before we're done,
link |
01:06:58.640
I just wanna just tie the bow
link |
01:07:00.960
and say that the spatial and temporal aspects
link |
01:07:04.280
are intimately interrelated with each other.
link |
01:07:06.440
So activity between neurons that are very close to each other
link |
01:07:10.320
is more likely to happen on this faster timescale
link |
01:07:13.560
and information is gonna propagate
link |
01:07:15.280
and encompass more of the brain,
link |
01:07:17.200
more of your cortices, different modules in the brain
link |
01:07:20.280
are gonna be engaged in information processing
link |
01:07:23.720
on longer timescales.
link |
01:07:25.280
So there's this concept of information integration
link |
01:07:27.960
where neurons are specialized.
link |
01:07:31.960
Any given neuron or any cluster of neuron
link |
01:07:33.960
has its specific purpose,
link |
01:07:35.720
but they're also very much integrated.
link |
01:07:39.880
So you have neurons that specialize,
link |
01:07:41.880
but share their information.
link |
01:07:43.640
And so that happens through these fractal nested oscillations
link |
01:07:47.560
that occur across spatial and temporal scales.
link |
01:07:49.400
I think capturing those dynamics in hardware,
link |
01:07:53.640
to me, that's the goal of neuromorphic computing.
link |
01:07:57.040
So does it need to look,
link |
01:07:58.680
so first of all, that's fascinating.
link |
01:08:00.800
We stated some clear principles here.
link |
01:08:03.960
Now, does it have to look like the brain
link |
01:08:08.120
outside of those principles as well?
link |
01:08:09.800
Like what other characteristics
link |
01:08:11.320
have to look like the human brain?
link |
01:08:13.080
Or can it be something very different?
link |
01:08:15.840
Well, it depends on what you're trying to use it for.
link |
01:08:18.000
And so I think a lot of the community
link |
01:08:21.720
asks that question a lot.
link |
01:08:23.080
What are you gonna do with it?
link |
01:08:24.360
And I completely get it.
link |
01:08:26.600
I think that's a very important question.
link |
01:08:28.040
And it's also sometimes not the most helpful question.
link |
01:08:31.840
What if what you wanna do with it is study it?
link |
01:08:33.800
What if you just wanna see,
link |
01:08:37.400
what do you have to build into your hardware
link |
01:08:39.280
in order to observe these dynamical principles?
link |
01:08:43.200
And also, I ask myself that question every day
link |
01:08:47.520
and I'm not sure I'm able to answer that.
link |
01:08:49.880
So like, what are you gonna do
link |
01:08:51.200
with this particular neuromorphic machine?
link |
01:08:53.480
So suppose what we're trying to do with it
link |
01:08:55.320
is build something that thinks.
link |
01:08:56.960
We're not trying to get it to make us any money
link |
01:08:59.160
or drive a car.
link |
01:09:00.240
Maybe we'll be able to do that, but that's not our goal.
link |
01:09:02.640
Our goal is to see if we can get the same types of behaviors
link |
01:09:07.600
that we observe in our own brain.
link |
01:09:08.920
And by behaviors in this sense,
link |
01:09:10.480
what I mean the behaviors of the components,
link |
01:09:14.320
the neurons, the network, that kind of stuff.
link |
01:09:16.000
I think there's another element that I didn't really hit on
link |
01:09:19.120
that you also have to build into this.
link |
01:09:21.200
And those are architectural principles.
link |
01:09:22.920
They have to do with the hierarchical modular construction
link |
01:09:26.680
of the network.
link |
01:09:27.520
And without getting too lost in jargon,
link |
01:09:30.200
the main point that I think is relevant there,
link |
01:09:33.680
let me try and illustrate it with a cartoon picture
link |
01:09:35.720
of the architecture of the brain.
link |
01:09:38.200
So in the brain, you have the cortex,
link |
01:09:41.120
which is sort of this outer sheet.
link |
01:09:44.440
It's actually, it's a layered structure.
link |
01:09:46.720
You can, if you could take it out of your brain,
link |
01:09:48.480
you could unroll it on the table
link |
01:09:50.680
and it would be about the size of a pizza sitting there.
link |
01:09:53.560
And that's a module.
link |
01:09:56.400
It does certain things.
link |
01:09:57.800
It processes as Yogi Buzaki would say,
link |
01:10:00.680
it processes the what of what's going on around you.
link |
01:10:03.560
But you have another really crucial module
link |
01:10:06.200
that's called the hippocampus.
link |
01:10:08.040
And that network is structured entirely differently.
link |
01:10:10.520
First of all, this cortex that had described
link |
01:10:12.800
10 billion neurons in there.
link |
01:10:14.640
So numbers matter here.
link |
01:10:16.840
And they're organized in that sort of power law distribution
link |
01:10:20.360
where the probability of making a connection drops off
link |
01:10:22.880
as a power law in space.
link |
01:10:24.520
The hippocampus is another module that's important
link |
01:10:26.840
for understanding how, where you are,
link |
01:10:30.880
when you are keeping track of your position
link |
01:10:36.240
in space and time.
link |
01:10:37.280
And that network is very much random.
link |
01:10:39.280
So the probability of making a connection,
link |
01:10:41.960
it almost doesn't even drop off as a function of distance.
link |
01:10:44.760
It's the same probability that you'll make it here
link |
01:10:46.720
to over there, but there are only about 100 million neurons
link |
01:10:50.520
there, so you can have that huge densely connected module
link |
01:10:54.680
because it's not so big.
link |
01:10:57.280
And the neocortex or the cortex and the hippocampus,
link |
01:11:02.040
they talk to each other constantly.
link |
01:11:04.920
And that communication is largely facilitated
link |
01:11:07.920
by what's called the thalamus.
link |
01:11:09.720
I'm not a neuroscientist here.
link |
01:11:10.880
I'm trying to do my best to recite things.
link |
01:11:12.960
Cartoon picture of the brain, I gotcha.
link |
01:11:14.680
Yeah, something like that.
link |
01:11:15.560
So this thalamus is coordinating the activity
link |
01:11:18.640
between the neocortex and the hippocampus
link |
01:11:20.760
and making sure that they talk to each other
link |
01:11:23.560
at the right time and send messages
link |
01:11:25.280
that will be useful to one another.
link |
01:11:26.840
So this all taken together is called
link |
01:11:29.120
the thalamocortical complex.
link |
01:11:31.600
And it seems like building something like that
link |
01:11:34.880
is going to be crucial to capturing the types of activity
link |
01:11:39.280
we're looking for because those responsibilities,
link |
01:11:43.400
those separate modules, they do different things,
link |
01:11:45.720
that's gotta be central to achieving these states
link |
01:11:51.760
of efficient information integration across space and time.
link |
01:11:55.720
By the way, I am able to achieve this state
link |
01:11:58.960
by watching simulations, visualizations
link |
01:12:01.800
of the thalamocortical complex.
link |
01:12:03.800
There's a few people I forget from where.
link |
01:12:06.440
They've created these incredible visual illustrations
link |
01:12:09.880
of visual stimulation from the eye or something like that.
link |
01:12:14.880
And this image flowing through the brain.
link |
01:12:18.520
Wow, I haven't seen that, I gotta check that out.
link |
01:12:20.880
So it's one of those things,
link |
01:12:22.120
you find this stuff in the world,
link |
01:12:24.280
and you see on YouTube, it has 1,000 views,
link |
01:12:26.960
these visualizations of the human brain
link |
01:12:30.800
processing information.
link |
01:12:32.120
And because there's chemistry there,
link |
01:12:36.440
because this is from actual human brains,
link |
01:12:38.880
I don't know how they're doing the coloring,
link |
01:12:40.720
but they're able to actually trace
link |
01:12:42.840
the different, the chemical and the electrical signals
link |
01:12:46.680
throughout the brain, and the visual thing,
link |
01:12:48.880
it's like, whoa, because it looks kinda like the universe,
link |
01:12:51.800
I mean, the whole thing is just incredible.
link |
01:12:53.800
I recommend it highly, I'll probably post a link to it.
link |
01:12:56.640
But you can just look for, one of the things they simulate
link |
01:13:00.960
is the thalamocortical complex and just visualization.
link |
01:13:05.840
You can find that yourself on YouTube, but it's beautiful.
link |
01:13:09.520
The other question I have for you is,
link |
01:13:11.320
how does memory play into all of this?
link |
01:13:14.440
Because all the signals sending back and forth,
link |
01:13:17.120
that's computation and communication,
link |
01:13:20.880
but that's kinda like processing of inputs and outputs,
link |
01:13:26.240
to produce outputs in the system,
link |
01:13:27.560
that's kinda like maybe reasoning,
link |
01:13:29.000
maybe there's some kind of recurrence.
link |
01:13:30.920
But is there a storage mechanism that you think about
link |
01:13:33.920
in the context of neuromorphic computing?
link |
01:13:35.840
Yeah, absolutely, so that's gotta be central.
link |
01:13:37.760
You have to have a way that you can store memories.
link |
01:13:41.520
And there are a lot of different kinds
link |
01:13:43.600
of memory in the brain.
link |
01:13:45.480
That's yet another example of how it's not a simple system.
link |
01:13:49.160
So there's one kind of memory,
link |
01:13:53.000
one way of talking about memory,
link |
01:13:56.040
usually starts in the context of Hopfield networks.
link |
01:13:59.040
You were lucky to talk to John Hopfield on this program.
link |
01:14:02.440
But the basic idea there is working memory
link |
01:14:05.840
is stored in the dynamical patterns
link |
01:14:07.840
of activity between neurons.
link |
01:14:10.400
And you can think of a certain pattern of activity
link |
01:14:14.760
as an attractor, meaning if you put in some signal
link |
01:14:19.680
that's similar enough to other
link |
01:14:22.400
previously experienced signals like that,
link |
01:14:26.480
then you're going to converge to the same network dynamics
link |
01:14:29.600
and you will see these neurons
link |
01:14:31.760
participate in the same network patterns of activity
link |
01:14:36.200
that they have in the past.
link |
01:14:37.600
So you can talk about the probability
link |
01:14:39.720
that different inputs will allow you to converge
link |
01:14:42.520
to different basins of attraction
link |
01:14:44.240
and you might think of that as,
link |
01:14:46.600
oh, I saw this face and then I excited
link |
01:14:49.040
this network pattern of activity
link |
01:14:50.920
because last time I saw that face,
link |
01:14:53.080
I was at some movie and that's a famous person
link |
01:14:56.960
that's on the screen or something like that.
link |
01:14:58.120
So that's one memory storage mechanism.
link |
01:15:00.560
But crucial to the ability to imprint those memories
link |
01:15:04.400
in your brain is the ability to change
link |
01:15:07.040
the strength of connection between one neuron and another,
link |
01:15:11.360
that synaptic connection between them.
link |
01:15:13.280
So synaptic weight update is a massive field of neuroscience
link |
01:15:18.000
and neuromorphic computing as well.
link |
01:15:19.560
So there are two poles on that spectrum.
link |
01:15:26.720
Okay, so more in the language of machine learning,
link |
01:15:28.880
we would talk about supervised and unsupervised learning.
link |
01:15:32.000
And when I'm trying to tie that down
link |
01:15:33.960
to neuromorphic computing,
link |
01:15:35.520
I will use a definition of supervised learning,
link |
01:15:38.440
which basically means the external user,
link |
01:15:42.960
the person who's controlling this hardware
link |
01:15:45.520
has some knob that they can tune
link |
01:15:48.360
to change each of the synaptic weights,
link |
01:15:50.400
depending on whether or not the network
link |
01:15:52.160
is doing what you want it to do.
link |
01:15:53.560
Whereas what I mean in this conversation
link |
01:15:56.120
when I say unsupervised learning
link |
01:15:57.600
is that those synaptic weights
link |
01:15:59.400
are dynamically changing in your network
link |
01:16:03.120
based on nothing that the user is doing,
link |
01:16:05.000
nothing that there's no wire from the outside
link |
01:16:07.080
going into any of those synapses.
link |
01:16:09.040
The network itself is reconfiguring those synaptic weights
link |
01:16:12.080
based on physical properties
link |
01:16:15.760
that you've built into the devices.
link |
01:16:17.600
So if the synapse receives a pulse from here
link |
01:16:21.400
and that causes the neuron to spike,
link |
01:16:23.360
some circuit built in there with no help from me
link |
01:16:27.040
or anybody else adjust the weight
link |
01:16:29.200
in a way that makes it more likely
link |
01:16:31.400
to store the useful information
link |
01:16:34.600
and excite the useful network patterns
link |
01:16:36.360
and makes it less likely that random noise,
link |
01:16:39.360
useless communication events
link |
01:16:41.440
will have an important effect on the network activity.
link |
01:16:45.320
So there's memory encoded in the weights,
link |
01:16:48.280
the synaptic weights.
link |
01:16:49.760
What about the formation of something
link |
01:16:51.880
that's not often done in machine learning,
link |
01:16:53.680
the formation of new synaptic connections?
link |
01:16:56.280
Right, well, that seems to,
link |
01:16:57.440
so again, not a neuroscientist here,
link |
01:17:00.120
but my reading of the literature
link |
01:17:01.960
is that that's particularly crucial
link |
01:17:04.000
in early stages of brain development
link |
01:17:06.400
where a newborn is born
link |
01:17:09.160
with tons of extra synaptic connections
link |
01:17:11.680
and it's actually pruned over time.
link |
01:17:13.880
So the number of synapses decreases
link |
01:17:16.800
as opposed to growing new long distance connections.
link |
01:17:19.680
It is possible in the brain to grow new neurons
link |
01:17:22.280
and assign new synaptic connections
link |
01:17:26.080
but it doesn't seem to be the primary mechanism
link |
01:17:29.120
by which the brain is learning.
link |
01:17:31.840
So for example, like right now,
link |
01:17:34.280
sitting here talking to you,
link |
01:17:35.720
you say lots of interesting things
link |
01:17:37.000
and I learn from you
link |
01:17:38.760
and I can remember things that you just said
link |
01:17:41.240
and I didn't grow new axonal connections
link |
01:17:44.720
down to new synapses to enable those.
link |
01:17:47.360
It's plasticity mechanisms
link |
01:17:50.160
in the synaptic connections between neurons
link |
01:17:52.920
that enable me to learn on that timescale.
link |
01:17:55.960
So at the very least,
link |
01:17:57.560
you can sufficiently approximate that
link |
01:17:59.880
with just weight updates.
link |
01:18:01.360
You don't need to form new connections.
link |
01:18:02.920
I would say weight updates are a big part of it.
link |
01:18:05.040
I also think there's more
link |
01:18:06.200
because broadly speaking,
link |
01:18:08.600
when we're doing machine learning,
link |
01:18:10.440
our networks, say we're talking about feed forward,
link |
01:18:12.480
deep neural networks,
link |
01:18:14.120
the temporal domain is not really part of it.
link |
01:18:16.960
Okay, you're gonna put in an image
link |
01:18:18.200
and you're gonna get out a classification
link |
01:18:20.400
and you're gonna do that as fast as possible.
link |
01:18:22.000
So you care about time
link |
01:18:23.160
but time is not part of the essence of this thing really.
link |
01:18:27.560
Whereas in spiking neural networks,
link |
01:18:30.040
what we see in the brain,
link |
01:18:31.760
time is as crucial as space
link |
01:18:33.360
and they're intimately intertwined
link |
01:18:34.600
as I've tried to say.
link |
01:18:36.000
And so adaptation on different timescales
link |
01:18:40.280
is important not just in memory formation,
link |
01:18:44.120
although it plays a key role there,
link |
01:18:45.360
but also in just keeping the activity
link |
01:18:48.240
in a useful dynamic range.
link |
01:18:50.320
So you have other plasticity mechanisms,
link |
01:18:52.520
not just weight update,
link |
01:18:54.200
or at least not on the timescale
link |
01:18:56.760
of many action potentials,
link |
01:18:58.760
but even on the shorter timescale.
link |
01:19:00.200
So a synapse can become much less efficacious.
link |
01:19:04.720
It can transmit a weaker signal
link |
01:19:07.200
after the second, third, fourth,
link |
01:19:08.800
that can second, third, fourth action potential
link |
01:19:11.960
to occur in a sequence.
link |
01:19:13.040
So that's what's called short term synaptic plasticity,
link |
01:19:15.960
which is a form of learning.
link |
01:19:17.600
You're learning that I'm getting too much stimulus
link |
01:19:19.640
from looking at something bright right now.
link |
01:19:21.640
So I need to tone that down.
link |
01:19:24.960
There's also another really important mechanism
link |
01:19:28.080
in learning that's called metoplasticity.
link |
01:19:30.520
What that seems to be is a way
link |
01:19:33.560
that you change not the weights themselves,
link |
01:19:37.400
but the rate at which the weights change.
link |
01:19:40.280
So when I am in say a lecture hall and my,
link |
01:19:45.440
this is a potentially terrible cartoon example,
link |
01:19:48.400
but let's say I'm in a lecture hall
link |
01:19:49.680
and it's time to learn, right?
link |
01:19:51.960
So my brain will release more,
link |
01:19:54.280
perhaps dopamine or some neuromodulator
link |
01:19:57.240
that's gonna change the rate
link |
01:20:00.320
at which synaptic plasticity occurs.
link |
01:20:02.240
So that can make me more sensitive
link |
01:20:03.840
to learning at certain times,
link |
01:20:05.320
more sensitive to overriding previous information
link |
01:20:08.360
and less sensitive at other times.
link |
01:20:10.320
And finally, as long as I'm rattling off the list,
link |
01:20:13.200
I think another concept that falls in the category
link |
01:20:16.480
of learning or memory adaptation is homeostasis
link |
01:20:20.560
or homeostatic adaptation,
link |
01:20:22.440
where neurons have the ability
link |
01:20:24.960
to control their firing rate.
link |
01:20:27.800
So if one neuron is just like blasting way too much,
link |
01:20:31.200
it will naturally tone itself down.
link |
01:20:33.000
Its threshold will adjust
link |
01:20:35.520
so that it stays in a useful dynamical range.
link |
01:20:38.520
And we see that that's captured in deep neural networks
link |
01:20:41.680
where you don't just change the synaptic weights,
link |
01:20:43.320
but you can also move the thresholds of simple neurons
link |
01:20:46.680
in those models.
link |
01:20:47.520
And so to achieve the spiking neural networks,
link |
01:20:53.800
you want to use,
link |
01:20:58.360
you want to implement the first principles
link |
01:21:01.200
that you mentioned of the temporal
link |
01:21:03.400
and the spatial fractal dynamics here.
link |
01:21:07.040
So you can communicate locally,
link |
01:21:09.240
you can communicate across much greater distances
link |
01:21:13.320
and do the same thing in space
link |
01:21:16.000
and do the same thing in time.
link |
01:21:18.040
Now, you have like a chapter called
link |
01:21:21.040
Superconducting Hardware for Neuromorphic Computing.
link |
01:21:24.360
So what are some ideas that integrate
link |
01:21:27.760
some of the things we've been talking about
link |
01:21:29.080
in terms of the first principles of neuromorphic computing
link |
01:21:32.080
and the ideas that you outline
link |
01:21:34.280
in optoelectronic intelligence?
link |
01:21:38.040
Yeah, so let me start, I guess,
link |
01:21:40.920
on the communication side of things,
link |
01:21:42.520
because that's what led us down this track
link |
01:21:46.280
in the first place.
link |
01:21:47.120
By us, I'm talking about my team of colleagues at NIST,
link |
01:21:51.800
Saeed Han, Bryce Brimavera, Sonia Buckley,
link |
01:21:54.800
Jeff Chiles, Adam McCallum to name,
link |
01:21:57.200
Alex Tate to name a few,
link |
01:21:58.720
our group leaders, Saewoo Nam and Rich Mirren.
link |
01:22:01.240
We've all contributed to this.
link |
01:22:02.480
So this is not me saying necessarily
link |
01:22:05.880
just the things that I've proposed,
link |
01:22:07.560
but sort of where our team's thinking
link |
01:22:09.600
has evolved over the years.
link |
01:22:11.560
Can I quickly ask, what is NIST
link |
01:22:14.720
and where is this amazing group of people located?
link |
01:22:18.080
NIST is the National Institute of Standards and Technology.
link |
01:22:23.120
The larger facility is out in Gaithersburg, Maryland.
link |
01:22:26.720
Our team is located in Boulder, Colorado.
link |
01:22:31.960
NIST is a federal agency under the Department of Commerce.
link |
01:22:36.240
We do a lot with, by we, I mean other people at NIST,
link |
01:22:40.160
do a lot with standards,
link |
01:22:43.640
making sure that we understand the system of units,
link |
01:22:46.080
international system of units, precision measurements.
link |
01:22:49.320
There's a lot going on in electrical engineering,
link |
01:22:53.560
material science.
link |
01:22:54.760
And it's historic.
link |
01:22:56.000
I mean, it's one of those, it's like MIT
link |
01:22:58.280
or something like that.
link |
01:22:59.120
It has a reputation over many decades
link |
01:23:00.960
of just being this really a place
link |
01:23:04.240
where there's a lot of brilliant people have done
link |
01:23:06.520
a lot of amazing things.
link |
01:23:07.600
But in terms of the people in your team,
link |
01:23:10.600
in this team of people involved
link |
01:23:12.760
in the concept we're talking about now,
link |
01:23:14.600
I'm just curious,
link |
01:23:15.440
what kind of disciplines are we talking about?
link |
01:23:17.240
What is it?
link |
01:23:18.080
Mostly physicists and electrical engineers,
link |
01:23:20.240
some material scientists,
link |
01:23:23.000
but I would say,
link |
01:23:24.840
yeah, I think physicists and electrical engineers,
link |
01:23:27.240
my background is in photonics,
link |
01:23:29.480
the use of light for technology.
link |
01:23:31.040
So coming from there, I tend to have found colleagues
link |
01:23:36.840
that are more from that background.
link |
01:23:38.240
Although Adam McConn,
link |
01:23:40.240
more of a superconducting electronics background,
link |
01:23:42.720
we need a diversity of folks.
link |
01:23:44.280
This project is sort of cross disciplinary.
link |
01:23:46.840
I would love to be working more
link |
01:23:48.280
with neuroscientists and things,
link |
01:23:50.800
but we haven't reached that scale yet.
link |
01:23:53.880
But yeah.
link |
01:23:54.720
You're focused on the hardware side,
link |
01:23:56.480
which requires all the disciplines that you mentioned.
link |
01:23:59.160
And then of course,
link |
01:24:00.000
neuroscientists may be a source of inspiration
link |
01:24:02.120
for some of the longterm vision.
link |
01:24:04.360
I would actually call it more than inspiration.
link |
01:24:06.240
I would call it sort of a roadmap.
link |
01:24:11.120
We're not trying to build exactly the brain,
link |
01:24:15.000
but I don't think it's enough to just say,
link |
01:24:17.520
oh, neurons kind of work like that.
link |
01:24:19.240
Let's kind of do that thing.
link |
01:24:20.760
I mean, we're very much following the concepts
link |
01:24:25.360
that the cognitive sciences have laid out for us,
link |
01:24:27.440
which I believe is a really robust roadmap.
link |
01:24:30.520
I mean, just on a little bit of a tangent,
link |
01:24:33.040
it's often stated that we just don't understand the brain.
link |
01:24:36.080
And so it's really hard to replicate it
link |
01:24:37.960
because we just don't know what's going on there.
link |
01:24:40.200
And maybe five or seven years ago,
link |
01:24:43.560
I would have said that,
link |
01:24:44.800
but as I got more interested in the subject,
link |
01:24:47.880
I read more of the neuroscience literature
link |
01:24:50.480
and I was just taken by the exact opposite sense.
link |
01:24:53.640
I can't believe how much they know about this.
link |
01:24:55.880
I can't believe how mathematically rigorous
link |
01:24:59.320
and sort of theoretically complete
link |
01:25:02.960
a lot of the concepts are.
link |
01:25:04.240
That's not to say we understand consciousness
link |
01:25:06.680
or we understand the self or anything like that,
link |
01:25:08.560
but what is the brain doing
link |
01:25:11.040
and why is it doing those things?
link |
01:25:13.600
Neuroscientists have a lot of answers to those questions.
link |
01:25:16.000
So if you're a hardware designer
link |
01:25:17.840
that just wants to get going,
link |
01:25:19.440
whoa, it's pretty clear which direction to go in, I think.
link |
01:25:23.000
Okay, so I love the optimism behind that,
link |
01:25:28.280
but in the implementation of these systems
link |
01:25:32.640
that uses superconductivity, how do you make it happen?
link |
01:25:39.320
So to me, it starts with thinking
link |
01:25:41.880
about the communication network.
link |
01:25:43.400
You know for sure that the ability of each neuron
link |
01:25:47.560
to communicate to many thousands of colleagues
link |
01:25:50.560
across the network is indispensable.
link |
01:25:52.360
I take that as a core principle of my architecture,
link |
01:25:56.280
my thinking on the subject.
link |
01:25:58.440
So coming from a background in photonics,
link |
01:26:02.280
it was very natural to say,
link |
01:26:03.560
okay, we're gonna use light for communication.
link |
01:26:05.360
Just in case listeners may not know,
link |
01:26:08.720
light is often used in communication.
link |
01:26:10.840
I mean, if you think about radio, that's light,
link |
01:26:12.720
it's long wavelengths, but it's electromagnetic radiation.
link |
01:26:15.320
It's the same physical phenomenon
link |
01:26:17.640
obeying exactly the same Maxwell's equations.
link |
01:26:20.200
And then all the way down to fiber, fiber optics.
link |
01:26:24.920
Now you're using visible
link |
01:26:26.240
or near infrared wavelengths of light,
link |
01:26:27.800
but the way you send messages across the ocean
link |
01:26:30.360
is now contemporary over optical fibers.
link |
01:26:33.200
So using light for communication is not a stretch.
link |
01:26:37.480
It makes perfect sense.
link |
01:26:38.960
So you might ask, well, why don't you use light
link |
01:26:41.520
for communication in a conventional microchip?
link |
01:26:45.280
And the answer to that is, I believe, physical.
link |
01:26:49.280
If we had a light source on a silicon chip
link |
01:26:53.080
that was as simple as a transistor,
link |
01:26:55.880
there would not be a processor in the world
link |
01:26:58.240
that didn't use light for communication,
link |
01:26:59.760
at least above some distance.
link |
01:27:01.800
How many light sources are needed?
link |
01:27:04.080
Oh, you need a light source at every single point.
link |
01:27:06.840
A light source per neuron.
link |
01:27:08.440
Per neuron, per little,
link |
01:27:09.960
but then if you could have a really small
link |
01:27:13.120
and nice light source,
link |
01:27:15.080
your definition of neuron could be flexible.
link |
01:27:17.960
Could be, yes, yes.
link |
01:27:19.200
Sometimes it's helpful to me to say,
link |
01:27:21.720
in this hardware, a neuron is that entity
link |
01:27:24.560
which has a light source.
link |
01:27:25.720
That, and I can explain.
link |
01:27:27.960
And then there was light.
link |
01:27:29.520
I mean, I can explain more about that, but.
link |
01:27:32.280
Somehow this like rhymes with consciousness
link |
01:27:34.680
because people will often say the light of consciousness.
link |
01:27:38.240
So that consciousness is that which is conscious.
link |
01:27:41.680
I got it.
link |
01:27:43.600
That's not my quote.
link |
01:27:44.840
That's me, that's my quote.
link |
01:27:47.000
You see, that quote comes from my background.
link |
01:27:49.520
Yours is in optics, mine in light, mine's in darkness.
link |
01:27:55.440
So go ahead.
link |
01:27:56.920
So the point I was making there is that
link |
01:27:59.640
if it was easy to manufacture light sources
link |
01:28:02.960
along with transistors on a silicon chip,
link |
01:28:05.760
they would be everywhere.
link |
01:28:07.240
And it's not easy.
link |
01:28:08.880
People have been trying for decades
link |
01:28:10.160
and it's actually extremely difficult.
link |
01:28:11.960
I think an important part of our research
link |
01:28:14.040
is dwelling right at that spot there.
link |
01:28:16.960
So.
link |
01:28:17.800
Is it physics or engineering?
link |
01:28:18.720
It's physics.
link |
01:28:19.560
So, okay, so it's physics, I think.
link |
01:28:22.280
So what I mean by that is, as we discussed,
link |
01:28:26.920
silicon is the material of choice for transistors
link |
01:28:29.400
and it's very difficult to imagine
link |
01:28:33.280
that that's gonna change anytime soon.
link |
01:28:35.200
Silicon is notoriously bad at emitting light.
link |
01:28:39.160
And that has to do with the immutable properties
link |
01:28:43.320
of silicon itself.
link |
01:28:44.160
The way that the energy bands are structured in silicon,
link |
01:28:47.880
you're never going to make silicon efficient
link |
01:28:50.520
as a light source at room temperature
link |
01:28:53.520
without doing very exotic things
link |
01:28:55.640
that degrade its ability to interface nicely
link |
01:28:58.320
with those transistors in the first place.
link |
01:28:59.840
So that's like one of these things where it's,
link |
01:29:02.600
why is nature dealing us that blow?
link |
01:29:05.200
You give us these beautiful transistors
link |
01:29:07.000
and you give us all the motivation
link |
01:29:08.640
to use light for communication,
link |
01:29:10.120
but then you don't give us a light source.
link |
01:29:11.520
So, well, okay, you do give us a light source.
link |
01:29:14.040
Compound semiconductors,
link |
01:29:15.320
like we talked about back at the beginning,
link |
01:29:16.880
an element from group three and an element from group five
link |
01:29:19.640
form an alloy where every other lattice site
link |
01:29:21.800
switches which element it is.
link |
01:29:23.680
Those have much better properties for generating light.
link |
01:29:27.240
You put electrons in, light comes out.
link |
01:29:30.040
Almost 100% of the electron hold,
link |
01:29:33.960
it can be made efficient.
link |
01:29:36.080
I'll take your word for it, okay.
link |
01:29:37.440
However, I say it's physics, not engineering,
link |
01:29:39.600
because it's very difficult
link |
01:29:41.920
to get those compound semiconductor light sources
link |
01:29:45.240
situated with your silicon.
link |
01:29:47.520
In order to do that ion implantation
link |
01:29:49.400
that I talked about at the beginning,
link |
01:29:50.760
high temperatures are required.
link |
01:29:52.720
So you gotta make all of your transistors first
link |
01:29:55.680
and then put the compound semiconductors on top of there.
link |
01:29:58.560
You can't grow them afterwards
link |
01:30:00.800
because that requires high temperature.
link |
01:30:02.360
It screws up all your transistors.
link |
01:30:04.000
You try and stick them on there.
link |
01:30:05.800
They don't have the same lattice constant.
link |
01:30:07.960
The spacing between atoms is different enough
link |
01:30:10.320
that it just doesn't work.
link |
01:30:11.640
So nature does not seem to be telling us that,
link |
01:30:15.520
hey, go ahead and combine light sources
link |
01:30:17.960
with your digital switches
link |
01:30:19.960
for conventional digital computing.
link |
01:30:22.680
And conventional digital computing
link |
01:30:24.720
will often require smaller scale, I guess,
link |
01:30:27.640
in terms of like smartphone.
link |
01:30:30.600
So in which kind of systems does nature hint
link |
01:30:35.440
that we can use light and photons for communication?
link |
01:30:40.440
Well, so let me just try and be clear.
link |
01:30:42.920
You can use light for communication in digital systems,
link |
01:30:46.240
just the light sources are not intimately integrated
link |
01:30:49.640
with the silicon.
link |
01:30:50.480
You manufacture all the silicon,
link |
01:30:52.080
you have your microchip, plunk it down.
link |
01:30:54.680
And then you manufacture your light sources,
link |
01:30:56.640
separate chip, completely different process
link |
01:30:58.720
made in a different foundry.
link |
01:31:00.440
And then you put those together at the package level.
link |
01:31:03.200
So now you have some,
link |
01:31:06.920
I would say a great deal of architectural limitations
link |
01:31:09.480
that are introduced by that sort of
link |
01:31:13.920
package level integration
link |
01:31:15.400
as opposed to monolithic on the same chip integration,
link |
01:31:18.000
but it's still a very useful thing to do.
link |
01:31:19.600
And that's where I had done some work previously
link |
01:31:23.120
before I came to NIST.
link |
01:31:24.080
There's a project led by Vladimir Stoyanovich
link |
01:31:27.920
that now spun out into a company called IR Labs
link |
01:31:30.800
led by Mark Wade and Chen Sun
link |
01:31:33.200
where they're doing exactly that.
link |
01:31:34.440
So you have your light source chip,
link |
01:31:36.440
your silicon chip, whatever it may be doing,
link |
01:31:39.760
maybe it's digital electronics,
link |
01:31:40.880
maybe it's some other control purpose, something.
link |
01:31:43.600
And the silicon chip drives the light source chip
link |
01:31:47.760
and modulates the intensity of the lights.
link |
01:31:49.800
You can get data out of the package on an optical fiber.
link |
01:31:52.640
And that still gives you tremendous advantages in bandwidth
link |
01:31:56.560
as opposed to sending those signals out
link |
01:31:58.800
over electrical lines.
link |
01:32:00.760
But it is somewhat peculiar to my eye
link |
01:32:05.160
that they have to be integrated at this package level.
link |
01:32:07.880
And those people, I mean, they're so smart.
link |
01:32:09.760
Those are my colleagues that I respect a great deal.
link |
01:32:12.880
So it's very clear that it's not just
link |
01:32:16.760
they're making a bad choice.
link |
01:32:18.760
This is what physics is telling us.
link |
01:32:20.480
It just wouldn't make any sense
link |
01:32:22.280
to try to stick them together.
link |
01:32:24.240
Yeah, so even if it's difficult,
link |
01:32:28.160
it's easier than the alternative, unfortunately.
link |
01:32:30.880
I think so, yes.
link |
01:32:31.720
And again, I need to go back
link |
01:32:33.280
and make sure that I'm not taking the wrong way.
link |
01:32:35.080
I'm not saying that the pursuit
link |
01:32:36.800
of integrating compound semiconductors with silicon
link |
01:32:39.400
is fruitless and shouldn't be pursued.
link |
01:32:41.120
It should, and people are doing great work.
link |
01:32:43.440
Kai Mei Lau and John Bowers, others,
link |
01:32:45.720
they're doing it and they're making progress.
link |
01:32:48.240
But to my eye, it doesn't look like that's ever going to be
link |
01:32:53.680
just the standard monolithic light source
link |
01:32:57.720
on silicon process.
link |
01:32:58.600
I just don't see it.
link |
01:33:00.040
Yeah, so nature kind of points the way usually.
link |
01:33:02.840
And if you resist nature,
link |
01:33:04.480
you're gonna have to do a lot more work.
link |
01:33:05.680
And it's gonna be expensive and not scalable.
link |
01:33:07.720
Got it.
link |
01:33:08.560
But okay, so let's go far into the future.
link |
01:33:11.320
Let's imagine this gigantic neuromorphic computing system
link |
01:33:14.760
that simulates all of our realities.
link |
01:33:17.360
It currently is Mantra Matrix 4.
link |
01:33:19.080
So this thing, this powerful computer,
link |
01:33:23.200
how does it operate?
link |
01:33:24.880
So what are the neurons?
link |
01:33:27.520
What is the communication?
link |
01:33:29.040
What's your sense?
link |
01:33:30.000
All right, so let me now,
link |
01:33:32.480
after spending 45 minutes trashing
link |
01:33:34.720
light source integration with silicon,
link |
01:33:36.160
let me now say why I'm basing my entire life,
link |
01:33:40.160
professional life, on integrating light sources
link |
01:33:43.680
with electronics.
link |
01:33:44.960
I think the game is completely different
link |
01:33:47.040
when you're talking about superconducting electronics.
link |
01:33:49.560
For several reasons, let me try to go through them.
link |
01:33:54.240
One is that, as I mentioned,
link |
01:33:56.480
it's difficult to integrate
link |
01:33:57.960
those compound semiconductor light sources with silicon.
link |
01:34:01.280
With silicon is a requirement that is introduced
link |
01:34:04.800
by the fact that you're using semiconducting electronics.
link |
01:34:07.280
In superconducting electronics,
link |
01:34:08.840
you're still gonna start with a silicon wafer,
link |
01:34:10.840
but it's just the bread for your sandwich in a lot of ways.
link |
01:34:13.880
You're not using that silicon
link |
01:34:15.800
in precisely the same way for the electronics.
link |
01:34:17.720
You're now depositing superconducting materials
link |
01:34:20.440
on top of that.
link |
01:34:21.840
The prospects for integrating light sources
link |
01:34:24.520
with that kind of an electronic process
link |
01:34:27.400
are certainly less explored,
link |
01:34:30.480
but I think much more promising
link |
01:34:31.960
because you don't need those light sources
link |
01:34:34.280
to be intimately integrated with the transistors.
link |
01:34:36.600
That's where the problems come up.
link |
01:34:37.920
They don't need to be lattice matched to the silicon,
link |
01:34:39.920
all that kind of stuff.
link |
01:34:41.160
Instead, it seems possible
link |
01:34:43.640
that you can take those compound semiconductor light sources,
link |
01:34:47.320
stick them on the silicon wafer,
link |
01:34:49.160
and then grow your superconducting electronics
link |
01:34:51.400
on the top of that.
link |
01:34:52.320
It's at least not obviously going to fail.
link |
01:34:55.800
So the computation would be done
link |
01:34:57.280
on the superconductive material as well?
link |
01:35:00.120
Yes, the computation is done
link |
01:35:01.920
in the superconducting electronics,
link |
01:35:03.920
and the light sources receive signals
link |
01:35:06.400
that say, hey, a neuron reached threshold,
link |
01:35:08.200
produce a pulse of light,
link |
01:35:09.800
send it out to all your downstream synaptic connections.
link |
01:35:12.480
Those are, again, superconducting electronics.
link |
01:35:16.240
Perform your computation,
link |
01:35:18.000
and you're off to the races.
link |
01:35:19.640
Your network works.
link |
01:35:20.760
So then if we can rewind real quick,
link |
01:35:22.600
so what are the limitations of the challenges
link |
01:35:25.960
of superconducting electronics
link |
01:35:28.920
when we think about constructing these kinds of systems?
link |
01:35:31.480
So actually, let me say one other thing
link |
01:35:35.640
about the light sources,
link |
01:35:37.560
and then I'll move on, I promise,
link |
01:35:39.840
because this is probably tedious for some.
link |
01:35:42.320
This is super exciting.
link |
01:35:44.000
Okay, one other thing about the light sources.
link |
01:35:45.720
I said that silicon is terrible at emitting photons.
link |
01:35:48.920
It's just not what it's meant to do.
link |
01:35:50.640
However, the game is different
link |
01:35:52.800
when you're at low temperature.
link |
01:35:54.080
If you're working with superconductors,
link |
01:35:55.720
you have to be at low temperature
link |
01:35:56.960
because they don't work otherwise.
link |
01:35:58.680
When you're at four Kelvin,
link |
01:36:00.240
silicon is not obviously a terrible light source.
link |
01:36:03.440
It's still not as efficient as compound semiconductors,
link |
01:36:05.840
but it might be good enough for this application.
link |
01:36:08.680
The final thing that I'll mention about that is, again,
link |
01:36:11.320
leveraging superconductors, as I said,
link |
01:36:13.960
in a different context,
link |
01:36:15.360
superconducting detectors can receive one single photon.
link |
01:36:19.520
In that conversation, I failed to mention
link |
01:36:21.240
that semiconductors can also receive photons.
link |
01:36:23.480
That's the primary mechanism by which it's done.
link |
01:36:26.200
A camera in your phone that's receptive to visible light
link |
01:36:29.400
is receiving photons.
link |
01:36:31.080
It's based on silicon,
link |
01:36:32.200
or you can make it in different semiconductors
link |
01:36:34.360
for different wavelengths,
link |
01:36:36.600
but it requires on the order of a thousand,
link |
01:36:39.600
a few thousand photons to receive a pulse.
link |
01:36:43.560
Now, when you're using a superconducting detector,
link |
01:36:46.160
you need one photon, exactly one.
link |
01:36:48.360
I mean, one or more.
link |
01:36:50.840
So the fact that your synapses can now be based
link |
01:36:54.840
on superconducting detectors
link |
01:36:56.680
instead of semiconducting detectors
link |
01:36:58.800
brings the light levels that are required
link |
01:37:00.720
down by some three orders of magnitude.
link |
01:37:03.160
So now you don't need good light sources.
link |
01:37:06.520
You can have the world's worst light sources.
link |
01:37:08.800
As long as they spit out maybe a few thousand photons
link |
01:37:11.840
every time a neuron fires,
link |
01:37:13.800
you have the hardware principles in place
link |
01:37:17.640
that you might be able to perform
link |
01:37:19.640
this optoelectronic integration.
link |
01:37:21.720
To me optoelectronic integration is, it's just so enticing.
link |
01:37:25.040
We want to be able to leverage electronics for computation,
link |
01:37:28.680
light for communication,
link |
01:37:30.400
working with silicon microelectronics at room temperature
link |
01:37:32.800
that has been exceedingly difficult.
link |
01:37:35.000
And I hope that when we move to the superconducting domain,
link |
01:37:40.000
target a different application space
link |
01:37:41.840
that is neuromorphic instead of digital
link |
01:37:44.920
and use superconducting detectors,
link |
01:37:47.680
maybe optoelectronic integration comes to us.
link |
01:37:50.120
Okay, so there's a bunch of questions.
link |
01:37:51.720
So one is temperature.
link |
01:37:53.680
So in these kinds of hybrid heterogeneous systems,
link |
01:37:58.320
what's the temperature?
link |
01:37:59.560
What are some of the constraints to the operation here?
link |
01:38:01.600
Does it all have to be a four Kelvin as well?
link |
01:38:03.560
Four Kelvin.
link |
01:38:04.400
Everything has to be at four Kelvin.
link |
01:38:06.840
Okay, so what are the other engineering challenges
link |
01:38:09.720
of making this kind of optoelectronic systems?
link |
01:38:14.320
Let me just dwell on that four Kelvin for a second
link |
01:38:16.720
because some people hear four Kelvin
link |
01:38:18.280
and they just get up and leave.
link |
01:38:19.280
They just say, I'm not doing it, you know?
link |
01:38:21.480
And to me, that's very earth centric, species centric.
link |
01:38:25.360
We live in 300 Kelvin.
link |
01:38:27.240
So we want our technologies to operate there too.
link |
01:38:29.120
I totally get it.
link |
01:38:30.080
Yeah, what's zero Celsius?
link |
01:38:31.880
Zero Celsius is 273 Kelvin.
link |
01:38:34.520
So we're talking very, very cold here.
link |
01:38:37.640
This is...
link |
01:38:38.480
Not even Boston cold.
link |
01:38:39.400
No.
link |
01:38:40.240
This is real cold.
link |
01:38:42.400
Yeah.
link |
01:38:43.240
Siberia cold, no.
link |
01:38:44.280
Okay, so just for reference,
link |
01:38:45.680
the temperature of the cosmic microwave background
link |
01:38:47.920
is about 2.7 Kelvin.
link |
01:38:49.400
So we're still warmer than deep space.
link |
01:38:51.680
Yeah, good.
link |
01:38:52.960
So that when the universe dies out,
link |
01:38:56.520
it'll be colder than four K.
link |
01:38:57.920
It's already colder than four K.
link |
01:38:59.400
In the expanses, you know,
link |
01:39:01.760
you don't have to get that far away from the earth
link |
01:39:05.000
in order to drop down to not far from four Kelvin.
link |
01:39:08.040
So what you're saying is the aliens that live at the edge
link |
01:39:11.200
of the observable universe
link |
01:39:13.280
are using superconductive material for their computation.
link |
01:39:16.440
They don't have to live at the edge of the universe.
link |
01:39:17.880
The aliens that are more advanced than us
link |
01:39:21.040
in their solar system are doing this
link |
01:39:24.480
in their asteroid belt.
link |
01:39:26.560
We can get to that.
link |
01:39:27.800
Oh, because they can get that
link |
01:39:30.320
to that temperature easier there?
link |
01:39:31.640
Sure, yeah.
link |
01:39:32.480
All you have to do is reflect the sunlight away
link |
01:39:34.120
and you have a huge headstart.
link |
01:39:36.080
Oh, so the sun is the problem here.
link |
01:39:37.680
Like it's warm here on earth.
link |
01:39:39.000
Got it. Yeah.
link |
01:39:39.840
Okay, so can you...
link |
01:39:41.560
So how do we get to four K?
link |
01:39:42.920
What's...
link |
01:39:43.760
Well, okay, so what I want to say about temperature...
link |
01:39:47.320
Yeah.
link |
01:39:48.160
What I want to say about temperature is that
link |
01:39:50.640
if you can swallow that,
link |
01:39:52.600
if you can say, all right, I give up applications
link |
01:39:56.200
that have to do with my cell phone
link |
01:39:58.200
and the convenience of a laptop on a train
link |
01:40:02.040
and you instead...
link |
01:40:03.760
For me, I'm very much in the scientific head space.
link |
01:40:06.440
I'm not looking at products.
link |
01:40:07.960
I'm not looking at what this will be useful
link |
01:40:09.880
to sell to consumers.
link |
01:40:11.000
Instead, I'm thinking about scientific questions.
link |
01:40:13.400
Well, it's just not that bad to have to work at four Kelvin.
link |
01:40:16.560
We do it all the time in our labs at NIST.
link |
01:40:19.000
And so does...
link |
01:40:19.840
I mean, for reference,
link |
01:40:21.440
the entire quantum computing sector
link |
01:40:25.560
usually has to work at something like 100 millikelvin,
link |
01:40:28.680
50 millikelvin.
link |
01:40:29.680
So now you're talking of another factor of 100
link |
01:40:32.120
even colder than that, a fraction of a degree.
link |
01:40:35.160
And everybody seems to think quantum computing
link |
01:40:37.360
is going to take over the world.
link |
01:40:39.200
It's so much more expensive
link |
01:40:40.720
to have to get that extra factor of 10 or whatever colder.
link |
01:40:46.600
And yet it's not stopping people from investing in that area.
link |
01:40:50.280
And by investing, I mean putting their research into it
link |
01:40:53.840
as well as venture capital or whatever.
link |
01:40:55.760
So...
link |
01:40:56.600
Oh, so based on the energy of what you're commenting on,
link |
01:40:59.600
I'm getting a sense that's one of the criticism
link |
01:41:01.960
of this approach is 4K, 4 Kelvin is a big negative.
link |
01:41:06.720
It is the showstopper for a lot of people.
link |
01:41:10.680
They just, I mean, and understandably,
link |
01:41:12.880
I'm not saying that that's not a consideration.
link |
01:41:16.840
Of course it is.
link |
01:41:17.840
For some...
link |
01:41:18.800
Okay, so different motivations for different people.
link |
01:41:21.440
In the academic world,
link |
01:41:23.000
suppose you spent your whole life
link |
01:41:24.400
learning about silicon microelectronic circuits.
link |
01:41:26.760
You send a design to a foundry,
link |
01:41:28.880
they send you back a chip
link |
01:41:30.520
and you go test it at your tabletop.
link |
01:41:33.000
And now I'm saying,
link |
01:41:34.360
here now learn how to use all these cryogenics
link |
01:41:36.520
so you can do that at 4 Kelvin.
link |
01:41:38.440
No, come on, man.
link |
01:41:39.880
I don't wanna do that.
link |
01:41:41.200
That sounds bad.
link |
01:41:42.040
It's the old momentum, the Titanic of the turning.
link |
01:41:44.520
Yeah, kind of.
link |
01:41:45.600
But you're saying that's not too much of a...
link |
01:41:48.360
When we're looking at large systems
link |
01:41:50.360
and the gain you can potentially get from them,
link |
01:41:52.320
that's not that much of a cost.
link |
01:41:53.440
And when you wanna answer the scientific question
link |
01:41:55.160
about what are the physical limits of cognition?
link |
01:41:58.120
Well, the physical limits,
link |
01:41:59.800
they don't care if you're at 4 Kelvin.
link |
01:42:01.440
If you can perform cognition at a scale
link |
01:42:04.560
orders of magnitude beyond any room temperature technology,
link |
01:42:07.680
but you gotta get cold to do it,
link |
01:42:09.760
you're gonna do it.
link |
01:42:10.600
And to me, that's the interesting application space.
link |
01:42:14.600
It's not even an application space,
link |
01:42:16.120
that's the interesting scientific paradigm.
link |
01:42:18.920
So I personally am not going to let low temperature
link |
01:42:22.560
stop me from realizing a technological domain or realm
link |
01:42:29.000
that is achieving in most ways everything else
link |
01:42:33.800
that I'm looking for in my hardware.
link |
01:42:36.160
So that, okay, that's a big one.
link |
01:42:37.640
Is there other kind of engineering challenges
link |
01:42:40.000
that you envision?
link |
01:42:40.840
Yeah, yeah, yeah.
link |
01:42:41.680
So let me take a moment here
link |
01:42:43.120
because I haven't really described what I mean
link |
01:42:45.760
by a neuron or a network in this particular hardware.
link |
01:42:49.000
Yeah, do you wanna talk about loop neurons
link |
01:42:51.640
and there's so many fascinating...
link |
01:42:53.680
But you just have so many amazing papers
link |
01:42:55.960
that people should definitely check out
link |
01:42:57.720
and the titles alone are just killer.
link |
01:42:59.880
So anyway, go ahead.
link |
01:43:01.080
Right, so let me say big picture,
link |
01:43:03.680
based on optics, photonics for communication,
link |
01:43:07.720
superconducting electronics for computation,
link |
01:43:10.120
how does this all work?
link |
01:43:11.520
So a neuron in this hardware platform
link |
01:43:17.480
can be thought of as circuits
link |
01:43:19.520
that are based on Josephson junctions,
link |
01:43:21.080
like we talked about before,
link |
01:43:22.680
where every time a photon comes in...
link |
01:43:25.320
So let's start by talking about a synapse.
link |
01:43:27.120
A synapse receives a photon, one or more,
link |
01:43:29.680
from a different neuron
link |
01:43:31.400
and it converts that optical signal
link |
01:43:33.600
to an electrical signal.
link |
01:43:35.320
The amount of current that that adds to a loop
link |
01:43:38.920
is controlled by the synaptic weight.
link |
01:43:40.840
So as I said before,
link |
01:43:42.360
you're popping fluxons into a loop, right?
link |
01:43:44.520
So a photon comes in,
link |
01:43:46.440
it hits a superconducting single photon detector,
link |
01:43:49.200
one photon, the absolute physical minimum
link |
01:43:52.120
that you can communicate
link |
01:43:53.040
from one place to another with light.
link |
01:43:54.520
And that detector then converts that
link |
01:43:56.120
into an electrical signal
link |
01:43:57.360
and the amount of signal
link |
01:43:58.720
is correlated with some kind of weight.
link |
01:44:01.200
Yeah, so the synaptic weight will tell you
link |
01:44:02.940
how many fluxons you pop into the loop.
link |
01:44:05.440
It's an analog number.
link |
01:44:06.560
We're doing analog computation now.
link |
01:44:08.160
Well, can you just linger on that?
link |
01:44:09.760
What the heck is a fluxon?
link |
01:44:10.980
Are we supposed to know this?
link |
01:44:11.960
Or is this a funny,
link |
01:44:14.220
is this like the big bang?
link |
01:44:15.600
Is this a funny word for something deeply technical?
link |
01:44:18.840
No, let's try to avoid using the word fluxon
link |
01:44:21.040
because it's not actually necessary.
link |
01:44:22.960
When a photon...
link |
01:44:24.360
It's fun to say though.
link |
01:44:25.440
So it's very necessary, I would say.
link |
01:44:29.400
When a photon hits
link |
01:44:30.360
that superconducting single photon detector,
link |
01:44:32.920
current is added to a superconducting loop.
link |
01:44:36.560
And the amount of current that you add
link |
01:44:39.200
is an analog value,
link |
01:44:40.340
can have eight bit equivalent resolution,
link |
01:44:42.980
something like that.
link |
01:44:44.240
10 bits, maybe.
link |
01:44:45.440
That's amazing, by the way.
link |
01:44:46.920
This is starting to make a lot more sense.
link |
01:44:48.480
When you're using superconductors for this,
link |
01:44:50.620
the energy of that circulating current
link |
01:44:54.720
is less than the energy of that photon.
link |
01:44:58.600
So your energy budget is not destroyed
link |
01:45:01.760
by doing this analog computation.
link |
01:45:04.000
So now in the language of a neuroscientist,
link |
01:45:07.080
you would say that's your postsynaptic signal.
link |
01:45:09.180
You have this current being stored in a loop.
link |
01:45:11.680
You can decide what you wanna do with it.
link |
01:45:13.480
Most likely you're gonna have it decay exponentially.
link |
01:45:16.160
So every single synapse
link |
01:45:18.280
is gonna have some given time constant.
link |
01:45:20.960
And that's determined by putting some resistor
link |
01:45:25.360
in that superconducting loop.
link |
01:45:27.040
So a synapse event occurs when a photon strikes a detector,
link |
01:45:31.120
adds current to that loop, it decays over time.
link |
01:45:33.880
That's the postsynaptic signal.
link |
01:45:35.560
Then you can process that in a dendritic tree.
link |
01:45:38.440
Bryce Primavera and I have a paper
link |
01:45:41.080
that we've submitted about that.
link |
01:45:43.420
For the more neuroscience oriented people,
link |
01:45:45.380
there's a lot of dendritic processing,
link |
01:45:47.080
a lot of plasticity mechanisms you can implement
link |
01:45:49.440
with essentially exactly the same circuits.
link |
01:45:51.480
You have this one simple building block circuit
link |
01:45:54.460
that you can use for a synapse, for a dendrite,
link |
01:45:57.240
for the neuron cell body, for all the plasticity functions.
link |
01:46:00.320
It's all based on the same building block,
link |
01:46:02.360
just tweaking a couple parameters.
link |
01:46:03.760
So this basic building block
link |
01:46:05.040
has both an optical and an electrical component,
link |
01:46:07.240
and then you just build arbitrary large systems with that?
link |
01:46:11.680
Close, you're not at fault
link |
01:46:13.520
for thinking that that's what I meant.
link |
01:46:15.000
What I should say is that if you want it to be a synapse,
link |
01:46:18.780
you tack a superconducting detector onto the front of it.
link |
01:46:22.240
And if you want it to be anything else,
link |
01:46:23.440
there's no optical component.
link |
01:46:25.280
Got it, so at the front,
link |
01:46:28.400
optics in the front, electrical stuff in the back.
link |
01:46:32.160
Electrical, yeah, in the processing
link |
01:46:34.120
and in the output signal that it sends
link |
01:46:36.480
to the next stage of processing further.
link |
01:46:39.380
So the dendritic trees is electrical.
link |
01:46:41.480
It's all electrical.
link |
01:46:42.500
It's all electrical in the superconducting domain.
link |
01:46:44.880
For anybody who's up on their superconducting circuits,
link |
01:46:48.560
it's just based on a DC squid, the most ubiquitous,
link |
01:46:52.440
which is a circuit composed of two Joseph's injunctions.
link |
01:46:55.160
So it's a very bread and butter kind of thing.
link |
01:46:58.720
And then the only place where you go beyond that
link |
01:47:00.900
is the neuron cell body itself.
link |
01:47:03.040
It's receiving all these electrical inputs
link |
01:47:05.300
from the synapses or dendrites
link |
01:47:06.840
or however you've structured that particular unique neuron.
link |
01:47:09.840
And when it reaches its threshold,
link |
01:47:12.320
which occurs by driving a Joseph's injunction
link |
01:47:14.360
above its critical current,
link |
01:47:15.760
it produces a pulse of current,
link |
01:47:17.200
which starts an amplification sequence,
link |
01:47:19.460
voltage amplification,
link |
01:47:21.380
that produces light out of a transmitter.
link |
01:47:24.480
So one of our colleagues, Adam McCann,
link |
01:47:26.880
and Sonia Buckley as well,
link |
01:47:27.960
did a lot of work on the light sources
link |
01:47:30.980
and the amplifiers that drive the current
link |
01:47:34.660
and produce sufficient voltage to drive current
link |
01:47:37.560
through that now semiconducting part.
link |
01:47:39.160
So that light source is the semiconducting part of a neuron.
link |
01:47:43.160
And that, so the neuron has reached threshold.
link |
01:47:45.620
It produces a pulse of light.
link |
01:47:47.640
That light then fans out across a network of wave guides
link |
01:47:51.560
to reach all the downstream synaptic terminals
link |
01:47:54.520
that perform this process themselves.
link |
01:47:57.320
So it's probably worth explaining
link |
01:47:59.880
what a network of wave guides is,
link |
01:48:02.320
because a lot of listeners aren't gonna know that.
link |
01:48:04.780
Look up the papers by Jeff Chiles on this one.
link |
01:48:07.120
But basically, light can be guided in a simple,
link |
01:48:11.380
basically wire of usually an insulating material.
link |
01:48:14.880
So silicon, silicon nitride,
link |
01:48:18.280
different kinds of glass,
link |
01:48:20.040
just like in a fiber optic, it's glass, silicon dioxide.
link |
01:48:23.440
That makes it a little bit big.
link |
01:48:24.840
We wanna bring these down.
link |
01:48:26.160
So we use different materials like silicon nitride,
link |
01:48:28.040
but basically just imagine a rectangle of some material
link |
01:48:32.980
that just goes and branches,
link |
01:48:37.080
forms different branch points
link |
01:48:39.980
that target different subregions of the network.
link |
01:48:43.060
You can transition between layers of these.
link |
01:48:45.000
So now we're talking about building in the third dimension,
link |
01:48:47.160
which is absolutely crucial.
link |
01:48:48.840
So that's what wave guides are.
link |
01:48:50.720
Yeah, that's great.
link |
01:48:52.040
Why the third dimension is crucial?
link |
01:48:54.680
Okay, so yes, you were talking about
link |
01:48:56.640
what are some of the technical limitations.
link |
01:48:59.660
One of the things that I believe we have to grapple with
link |
01:49:04.280
is that our brains are miraculously compact.
link |
01:49:08.800
For the number of neurons that are in our brain,
link |
01:49:11.560
it sure does fit in a small volume,
link |
01:49:13.520
as it would have to if we're gonna be biological organisms
link |
01:49:16.200
that are resource limited and things like that.
link |
01:49:19.260
Any kind of hardware neuron
link |
01:49:20.960
is almost certainly gonna be much bigger than that
link |
01:49:23.460
if it is of comparable complexity,
link |
01:49:26.480
whether it's based on silicon transistors.
link |
01:49:28.540
Okay, a transistor, seven nanometers,
link |
01:49:30.720
that doesn't mean a semiconductor based neuron
link |
01:49:33.760
is seven nanometers.
link |
01:49:34.680
They're big.
link |
01:49:35.920
They require many transistors,
link |
01:49:38.780
different other things like capacitors and things
link |
01:49:40.640
that store charge.
link |
01:49:41.540
They end up being on the order of 100 microns
link |
01:49:44.660
by 100 microns,
link |
01:49:45.880
and it's difficult to get them down any smaller than that.
link |
01:49:48.240
The same is true for superconducting neurons,
link |
01:49:50.600
and the same is true
link |
01:49:52.080
if we're trying to use light for communication.
link |
01:49:54.000
Even if you're using electrons for communication,
link |
01:49:56.860
you have these wires where, okay,
link |
01:50:00.680
the size of an electron might be angstroms,
link |
01:50:03.360
but the size of a wire is not angstroms,
link |
01:50:05.560
and if you try and make it narrower,
link |
01:50:07.160
the resistance just goes up,
link |
01:50:08.520
so you don't actually win.
link |
01:50:10.760
To communicate over long distances,
link |
01:50:12.360
you need your wires to be microns wide,
link |
01:50:15.640
and it's the same thing for wave guides.
link |
01:50:17.160
Wave guides are essentially limited
link |
01:50:18.920
by the wavelength of light,
link |
01:50:20.280
and that's gonna be about a micron,
link |
01:50:21.920
so whereas compare that to an axon,
link |
01:50:24.680
the analogous component in the brain,
link |
01:50:26.760
which is 10 nanometers in diameter, something like that,
link |
01:50:32.160
they're bigger when they need to communicate
link |
01:50:33.680
over long distances,
link |
01:50:34.640
but grappling with the size of these structures
link |
01:50:37.960
is inevitable and crucial,
link |
01:50:39.820
and so in order to make systems of comparable scale
link |
01:50:45.000
to the human brain, by scale here,
link |
01:50:46.880
I mean number of interconnected neurons,
link |
01:50:49.840
you absolutely have to be using
link |
01:50:51.760
the third spatial dimension,
link |
01:50:53.640
and that means on the wafer,
link |
01:50:55.960
you need multiple layers
link |
01:50:57.360
of both active and passive components.
link |
01:50:59.600
Active, I mean superconducting electronic circuits
link |
01:51:03.360
that are performing computations,
link |
01:51:05.360
and passive, I mean these wave guides
link |
01:51:07.400
that are routing the optical signals to different places,
link |
01:51:10.140
you have to be able to stack those.
link |
01:51:11.720
If you can get to something like 10 planes
link |
01:51:14.720
of each of those, or maybe not even 10,
link |
01:51:16.160
maybe five, six, something like that,
link |
01:51:18.840
then you're in business.
link |
01:51:19.760
Now you can get millions of neurons on a wafer,
link |
01:51:22.860
but that's not anywhere close to the brain scale.
link |
01:51:26.380
In order to get to the scale of the human brain,
link |
01:51:27.960
you're gonna have to also use the third dimension
link |
01:51:30.000
in the sense that entire wafers
link |
01:51:32.560
need to be stacked on top of each other
link |
01:51:34.160
with fiber optic communication between them,
link |
01:51:36.080
and we need to be able to fill a space
link |
01:51:38.640
the size of this table with stacked wafers,
link |
01:51:42.040
and that's when you can get to some 10 billion neurons
link |
01:51:44.280
like your human brain,
link |
01:51:45.120
and I don't think that's specific
link |
01:51:46.680
to the optoelectronic approach that we're taking.
link |
01:51:48.800
I think that applies to any hardware
link |
01:51:51.400
where you're trying to reach commensurate scale
link |
01:51:53.560
and complexity as the human brain.
link |
01:51:55.000
So you need that fractal stacking,
link |
01:51:57.520
so stacking on the wafer,
link |
01:51:59.360
and stacking of the wafers,
link |
01:52:01.120
and then whatever the system that combines,
link |
01:52:03.840
this stacking of the tables with the wafers.
link |
01:52:06.360
And it has to be fractal all the way,
link |
01:52:07.720
you're exactly right,
link |
01:52:08.600
because that's the only way
link |
01:52:10.060
that you can efficiently get information
link |
01:52:12.500
from a small point to across that whole network.
link |
01:52:15.060
It has to have the power law connected.
link |
01:52:17.520
And photons are like optics throughout.
link |
01:52:20.640
Yeah, absolutely.
link |
01:52:21.480
Once you're at this scale, to me it's just obvious.
link |
01:52:23.680
Of course you're using light for communication.
link |
01:52:25.580
You have fiber optics given to us from nature, so simple.
link |
01:52:30.580
The thought of even trying to do
link |
01:52:32.860
any kind of electrical communication
link |
01:52:34.740
just doesn't make sense to me.
link |
01:52:37.140
I'm not saying it's wrong, I don't know,
link |
01:52:39.260
but that's where I'm coming from.
link |
01:52:40.980
So let's return to loop neurons.
link |
01:52:43.860
Why are they called loop neurons?
link |
01:52:46.460
Yeah, the term loop neurons comes from the fact,
link |
01:52:48.900
like we've been talking about,
link |
01:52:49.980
that they rely heavily on these superconducting loops.
link |
01:52:53.260
So even in a lot of forms of digital computing
link |
01:52:57.420
with superconductors,
link |
01:52:58.740
storing a signal in a superconducting loop
link |
01:53:02.540
is a primary technique.
link |
01:53:05.060
In this particular case,
link |
01:53:06.660
it's just loops everywhere you look.
link |
01:53:08.620
So the strength of a synaptic weight
link |
01:53:11.620
is gonna be set by the amount of current circulating
link |
01:53:15.060
in a loop that is coupled to the synapse.
link |
01:53:17.820
So memory is implemented as current circulating
link |
01:53:22.820
in a superconducting loop.
link |
01:53:24.140
The coupling between, say, a synapse and a dendrite
link |
01:53:27.140
or a synapse in the neuron cell body
link |
01:53:29.260
occurs through loop coupling through transformers.
link |
01:53:33.140
So current circulating in a synapse
link |
01:53:34.820
is gonna induce current in a different loop,
link |
01:53:37.580
a receiving loop in the neuron cell body.
link |
01:53:40.820
So since all of the computation is happening
link |
01:53:44.500
in these flux storage loops
link |
01:53:46.660
and they play such a central role
link |
01:53:48.380
in how the information is processed,
link |
01:53:50.260
how memories are formed, all that stuff,
link |
01:53:52.820
I didn't think too much about it,
link |
01:53:53.980
I just called them loop neurons
link |
01:53:55.540
because it rolls off the tongue a little bit better
link |
01:53:58.180
than superconducting optoelectronic neurons.
link |
01:54:02.220
Okay, so how do you design circuits for these loop neurons?
link |
01:54:08.540
That's a great question.
link |
01:54:09.740
There's a lot of different scales of design.
link |
01:54:12.340
So at the level of just one synapse,
link |
01:54:16.340
you can use conventional methods.
link |
01:54:18.980
They're not that complicated
link |
01:54:21.100
as far as superconducting electronics goes.
link |
01:54:23.220
It's just four Joseph's injunctions or something like that
link |
01:54:27.100
depending on how much complexity you wanna add.
link |
01:54:29.180
So you can just directly simulate each component in SPICE.
link |
01:54:34.940
What's SPICE?
link |
01:54:35.860
It's Standard Electrical Simulation Software, basically.
link |
01:54:39.220
So you're just explicitly solving the differential equations
link |
01:54:42.580
that describe the circuit elements.
link |
01:54:44.340
And then you can stack these things together
link |
01:54:46.220
in that simulation software to then build circuits.
link |
01:54:48.860
You can, but that becomes computationally expensive.
link |
01:54:51.580
So one of the things when COVID hit,
link |
01:54:54.100
we knew we had to turn some attention
link |
01:54:55.780
to more things you can do at home in your basement
link |
01:54:59.380
or whatever, and one of them was computational modeling.
link |
01:55:02.700
So we started working on adapting,
link |
01:55:07.780
abstracting out the circuit performance
link |
01:55:10.180
so that you don't have to explicitly solve
link |
01:55:12.540
the circuit equations, which for Joseph's injunctions
link |
01:55:15.860
usually needs to be done on like a picosecond timescale
link |
01:55:18.460
and you have a lot of nodes in your circuit.
link |
01:55:21.220
So it results in a lot of differential equations
link |
01:55:24.860
that need to be solved simultaneously.
link |
01:55:26.340
We were looking for a way to simulate these circuits
link |
01:55:29.420
that is scalable up to networks of millions or so neurons
link |
01:55:33.740
is sort of where we're targeting right now.
link |
01:55:36.620
So we were able to analyze the behavior of these circuits.
link |
01:55:40.820
And as I said, it's based on these simple building blocks.
link |
01:55:43.620
So you really only need to understand
link |
01:55:45.140
this one building block.
link |
01:55:46.340
And if you get a good model of that, boom, it tiles.
link |
01:55:48.980
And you can change the parameters in there
link |
01:55:51.140
to get different behaviors and stuff,
link |
01:55:52.860
but it's all based on now it's one differential equation
link |
01:55:56.180
that you need to solve.
link |
01:55:57.140
So one differential equation for every synapse,
link |
01:56:00.780
dendrite or neuron in your system.
link |
01:56:03.740
And for the neuroscientists out there,
link |
01:56:05.340
it's just a simple leaky integrate and fire model,
link |
01:56:08.420
leaky integrator, basically.
link |
01:56:10.300
A synapse is a leaky integrator,
link |
01:56:11.860
a dendrite is a leaky integrator.
link |
01:56:13.460
So I'm really fascinated by how this one simple component
link |
01:56:18.460
can be used to achieve lots of different types
link |
01:56:22.180
of dynamical activity.
link |
01:56:24.420
And to me, that's where scalability comes from.
link |
01:56:27.500
And also complexity as well.
link |
01:56:29.180
Complexity is often characterized
link |
01:56:30.860
by relatively simple building blocks
link |
01:56:35.420
connected in potentially simple
link |
01:56:37.820
or sometimes complicated ways,
link |
01:56:39.340
and then emergent new behavior that was hard to predict
link |
01:56:41.940
from those simple elements.
link |
01:56:44.660
And that's exactly what we're working with here.
link |
01:56:46.620
So it's a very exciting platform,
link |
01:56:49.020
both from a modeling perspective
link |
01:56:50.380
and from a hardware manifestation perspective
link |
01:56:52.740
where we can hopefully start to have this test bed
link |
01:56:57.380
where we can explore things,
link |
01:56:58.860
not just related to neuroscience,
link |
01:57:00.820
but also related to other things
link |
01:57:03.300
that connected to other physics like critical phenomenon,
link |
01:57:07.140
Ising models, things like that.
link |
01:57:08.660
So you were asking how we simulate these circuits.
link |
01:57:13.060
It's at different levels
link |
01:57:14.540
and we've got the simple spice circuit stuff.
link |
01:57:18.300
That's no problem.
link |
01:57:19.540
And now we're building these network models
link |
01:57:21.740
based on this more efficient leaky integrator.
link |
01:57:23.620
So we can actually reduce every element
link |
01:57:26.220
to one differential equation.
link |
01:57:27.460
And then we can also step through it
link |
01:57:28.860
on a much coarser time grid.
link |
01:57:30.700
So it ends up being something like a factor
link |
01:57:32.380
of a thousand to 10,000 speed improvement,
link |
01:57:35.380
which allows us to simulate,
link |
01:57:37.740
but hopefully up to millions of neurons.
link |
01:57:40.540
Whereas before we would have been limited to tens,
link |
01:57:44.660
a hundred, something like that.
link |
01:57:45.780
And just like simulating quantum mechanical systems
link |
01:57:48.780
with a quantum computer.
link |
01:57:49.860
So the goal here is to understand such systems.
link |
01:57:53.380
For me, the goal is to study this
link |
01:57:55.900
as a scientific physical system.
link |
01:57:58.940
I'm not drawn towards turning this
link |
01:58:01.940
into an enterprise at this point.
link |
01:58:03.460
I feel short term applications
link |
01:58:05.820
that obviously make a lot of money
link |
01:58:07.460
is not necessarily a curiosity driver for you at the moment.
link |
01:58:11.380
Absolutely not.
link |
01:58:12.220
If you're interested in short term making money,
link |
01:58:14.020
go with deep learning, use silicon microelectronics.
link |
01:58:16.500
If you wanna understand things like the physics
link |
01:58:20.900
of a fascinating system,
link |
01:58:23.020
or if you wanna understand something more
link |
01:58:25.620
along the lines of the physical limits
link |
01:58:27.820
of what can be achieved,
link |
01:58:29.460
then I think single photon communication,
link |
01:58:32.700
superconducting electronics is extremely exciting.
link |
01:58:35.940
What if I wanna use superconducting hardware
link |
01:58:39.620
at four Kelvin to mine Bitcoin?
link |
01:58:42.020
That's my main interest.
link |
01:58:44.220
The reason I wanted to talk to you today,
link |
01:58:45.860
I wanna say, no, I don't know.
link |
01:58:47.660
What's Bitcoin?
link |
01:58:51.540
Look it up on the internet.
link |
01:58:53.140
Somebody told me about it.
link |
01:58:54.780
I'm not sure exactly what it is.
link |
01:58:57.500
But let me ask nevertheless
link |
01:58:59.060
about applications to machine learning.
link |
01:59:01.380
Okay, so if you look at the scale of five, 10, 20 years,
link |
01:59:07.300
is it possible to, before we understand the nature
link |
01:59:11.900
of human intelligence and general intelligence,
link |
01:59:14.420
do you think we'll start falling out of this exploration
link |
01:59:19.100
of neuromorphic systems ability to solve some
link |
01:59:23.100
of the problems that the machine learning systems
link |
01:59:25.180
of today can't solve?
link |
01:59:26.420
Well, I'm really hesitant to over promise.
link |
01:59:31.620
So I really don't know.
link |
01:59:34.100
Also, I don't really understand machine learning
link |
01:59:36.740
in a lot of senses.
link |
01:59:37.580
I mean, machine learning from my perspective appears
link |
01:59:42.900
to require that you know precisely what your input is
link |
01:59:49.180
and also what your goal is.
link |
01:59:51.700
You usually have some objective function
link |
01:59:53.300
or something like that.
link |
01:59:54.140
And that's very limiting.
link |
01:59:57.300
I mean, of course, a lot of times that's the case.
link |
02:00:00.540
There's a picture and there's a horse in it, so you're done.
link |
02:00:03.940
But that's not a very interesting problem.
link |
02:00:06.500
I think when I think about intelligence,
link |
02:00:09.500
it's almost defined by the ability to handle problems
link |
02:00:13.200
where you don't know what your inputs are going to be
link |
02:00:15.940
and you don't even necessarily know
link |
02:00:17.420
what you're trying to accomplish.
link |
02:00:18.580
I mean, I'm not sure what I'm trying to accomplish
link |
02:00:21.620
in this world.
link |
02:00:22.540
Yeah, at all scales.
link |
02:00:24.540
Yeah, at all scales, right.
link |
02:00:25.900
I mean, so I'm more drawn to the underlying phenomena,
link |
02:00:33.700
the critical dynamics of this system,
link |
02:00:37.940
trying to understand how elements that you build
link |
02:00:41.900
into your hardware result in emergent fascinating activity
link |
02:00:48.580
that was very difficult to predict, things like that.
link |
02:00:51.740
So, but I gotta be really careful
link |
02:00:53.660
because I think a lot of other people who,
link |
02:00:55.580
if they found themselves working on this project
link |
02:00:57.700
in my shoes, they would say, all right,
link |
02:00:59.300
what are all the different ways we can use this
link |
02:01:01.660
for machine learning?
link |
02:01:02.500
Actually, let me just definitely mention colleague
link |
02:01:05.340
at NIST, Mike Schneider.
link |
02:01:06.660
He's also very much interested,
link |
02:01:09.340
particularly in the superconducting side of things,
link |
02:01:11.720
using the incredible speed, power efficiency,
link |
02:01:14.920
also Ken Seagal at Colgate,
link |
02:01:16.620
other people working on specifically
link |
02:01:18.780
the superconducting side of this for machine learning
link |
02:01:22.020
and deep feed forward neural networks.
link |
02:01:25.060
There, the advantages are obvious.
link |
02:01:27.400
It's extremely fast.
link |
02:01:28.780
Yeah, so that's less on the nature of intelligences
link |
02:01:31.880
and more on various characteristics of this hardware
link |
02:01:35.740
that you can use for the basic computation
link |
02:01:38.340
as we know it today and communication.
link |
02:01:40.680
One of the things that Mike Schneider's working on right now
link |
02:01:42.900
is an image classifier at a relatively small scale.
link |
02:01:46.160
I think he's targeting that nine pixel problem
link |
02:01:48.380
where you can have three different characters
link |
02:01:50.100
and you put in a nine pixel image
link |
02:01:54.240
and you classify it as one of these three categories.
link |
02:01:58.420
And that's gonna be really interesting
link |
02:02:00.160
to see what happens there,
link |
02:02:01.260
because if you can show that even at that scale,
link |
02:02:05.720
you just put these images in and you get it out
link |
02:02:08.180
and he thinks he can do it,
link |
02:02:09.540
I forgot if it's a nanosecond
link |
02:02:11.180
or some extremely fast classification time,
link |
02:02:14.040
it's probably less,
link |
02:02:14.880
it's probably a hundred picoseconds or something.
link |
02:02:17.420
There you have challenges though,
link |
02:02:18.820
because the Joseph's injunctions themselves,
link |
02:02:21.600
the electronic circuit is extremely power efficient.
link |
02:02:24.580
Some orders of magnitude for something more
link |
02:02:26.880
than a transistor doing the same thing,
link |
02:02:29.280
but when you have to cool it down to four Kelvin,
link |
02:02:31.480
you pay a huge overhead just for keeping it cold,
link |
02:02:33.900
even if it's not doing anything.
link |
02:02:35.300
So it has to work at large scale
link |
02:02:40.680
in order to overcome that power penalty,
link |
02:02:43.720
but that's possible.
link |
02:02:45.140
It's just, it's gonna have to get that performance.
link |
02:02:48.000
And this is sort of what you were asking about before
link |
02:02:49.720
is like how much better than silicon would it need to be?
link |
02:02:52.760
And the answer is, I don't know.
link |
02:02:54.120
I think if it's just overall better than silicon
link |
02:02:57.220
at a problem that a lot of people care about,
link |
02:03:00.200
maybe it's image classification,
link |
02:03:02.640
maybe it's facial recognition,
link |
02:03:03.960
maybe it's monitoring credit transactions, I don't know,
link |
02:03:07.520
then I think it will have a place.
link |
02:03:09.000
It's not gonna be in your cell phone,
link |
02:03:10.440
but it could be in your data center.
link |
02:03:12.200
So what about in terms of the data center,
link |
02:03:16.160
I don't know if you're paying attention
link |
02:03:17.680
to the various systems,
link |
02:03:19.080
like Tesla recently announced DOJO,
link |
02:03:23.880
which is a large scale machine learning training system,
link |
02:03:27.160
that again, the bottleneck there
link |
02:03:28.940
is probably going to be communication
link |
02:03:30.900
between those systems.
link |
02:03:32.800
Is there something from your work
link |
02:03:34.880
on everything we've been talking about
link |
02:03:38.720
in terms of superconductive hardware
link |
02:03:41.640
that could be useful there?
link |
02:03:43.560
Oh, I mean, okay, tomorrow, no.
link |
02:03:46.720
In the long term, it could be the whole thing.
link |
02:03:49.000
It could be nothing.
link |
02:03:49.840
I don't know, but definitely, definitely.
link |
02:03:54.040
When you look at the,
link |
02:03:55.160
so I don't know that much about DOJO.
link |
02:03:56.720
My understanding is that that's new, right?
link |
02:03:58.840
That's just coming online.
link |
02:04:01.320
Well, I don't even know where it hasn't come online.
link |
02:04:06.960
And when you announce big, sexy,
link |
02:04:09.560
so let me explain to you the way things work
link |
02:04:11.360
in the world of business and marketing.
link |
02:04:15.680
It's not always clear where you are
link |
02:04:18.400
on the coming online part of that.
link |
02:04:20.560
So I don't know where they are exactly,
link |
02:04:22.680
but the vision is from a ground up
link |
02:04:25.240
to build a very, very large scale,
link |
02:04:28.520
modular machine learning, ASIC,
link |
02:04:31.080
basically hardware that's optimized
link |
02:04:32.640
for training neural networks.
link |
02:04:34.120
And of course, there's a lot of companies
link |
02:04:36.080
that are small and big working on this kind of problem.
link |
02:04:39.360
The question is how to do it in a modular way
link |
02:04:42.680
that has very fast communication.
link |
02:04:45.640
The interesting aspect of Tesla is you have a company
link |
02:04:49.520
that at least at this time is so singularly focused
link |
02:04:54.640
on solving a particular machine learning problem
link |
02:04:57.720
and is making obviously a lot of money doing so
link |
02:05:00.880
because the machine learning problem
link |
02:05:02.200
happens to be involved with autonomous driving.
link |
02:05:05.200
So you have a system that's driven by an application.
link |
02:05:09.760
And that's really interesting because you have maybe Google
link |
02:05:15.040
working on TPUs and so on.
link |
02:05:17.720
You have all these other companies with ASICs.
link |
02:05:21.440
They're usually more kind of always thinking general.
link |
02:05:25.960
So I like it when it's driven by a particular application
link |
02:05:29.160
because then you can really get to the,
link |
02:05:32.200
it's somehow if you just talk broadly about intelligence,
link |
02:05:37.240
you may not always get to the right solutions.
link |
02:05:40.240
It's nice to couple that sometimes
link |
02:05:41.560
with specific clear illustration
link |
02:05:45.600
of something that requires general intelligence,
link |
02:05:47.640
which for me driving is one such case.
link |
02:05:49.880
I think you're exactly right.
link |
02:05:51.080
Sometimes just having that focus on that application
link |
02:05:54.360
brings a lot of people focuses their energy and attention.
link |
02:05:57.640
I think that, so one of the things that's appealing
link |
02:06:00.200
about what you're saying is not just
link |
02:06:02.440
that the application is specific,
link |
02:06:03.840
but also that the scale is big
link |
02:06:06.040
and that the benefit is also huge.
link |
02:06:10.640
Financial and to humanity.
link |
02:06:12.280
Right, right, right.
link |
02:06:13.120
Yeah, so I guess let me just try to understand
link |
02:06:15.480
is the point of this dojo system
link |
02:06:17.960
to figure out the parameters
link |
02:06:21.840
that then plug into neural networks
link |
02:06:23.840
and then you don't need to retrain,
link |
02:06:26.520
you just make copies of a certain chip
link |
02:06:29.080
that has all the other parameters established or?
link |
02:06:32.320
No, it's straight up retraining a large neural network
link |
02:06:36.760
over and over and over.
link |
02:06:38.560
So you have to do it once for every new car?
link |
02:06:41.640
No, no, you have to, so they do this interesting process,
link |
02:06:44.800
which I think is a process for machine learning,
link |
02:06:47.000
supervised machine learning systems
link |
02:06:49.200
you're going to have to do, which is you have a system,
link |
02:06:53.720
you train your network once, it takes a long time.
link |
02:06:56.400
I don't know how long, but maybe a week.
link |
02:06:58.720
Okay. To train.
link |
02:07:00.840
And then you deploy it on, let's say about a million cars.
link |
02:07:05.080
I don't know what the number is.
link |
02:07:05.920
But that part, you just write software
link |
02:07:07.840
that updates some weights in a table and yeah, okay.
link |
02:07:10.680
But there's a loop back.
link |
02:07:12.640
Yeah, yeah, okay.
link |
02:07:13.480
Each of those cars run into trouble, rarely,
link |
02:07:18.840
but they catch the edge cases
link |
02:07:23.720
of the performance of that particular system
link |
02:07:26.080
and then send that data back
link |
02:07:28.320
and either automatically or by humans,
link |
02:07:31.440
that weird edge case data is annotated
link |
02:07:34.760
and then the network has to become smart enough
link |
02:07:37.560
to now be able to perform in those edge cases,
link |
02:07:40.120
so it has to get retrained.
link |
02:07:41.800
There's clever ways of retraining different parts
link |
02:07:43.960
of that network, but for the most part,
link |
02:07:46.240
I think they prefer to retrain the entire thing.
link |
02:07:49.280
So you have this giant monster
link |
02:07:51.320
that kind of has to be retrained regularly.
link |
02:07:54.880
I think the vision with Dojo is to have
link |
02:07:58.960
a very large machine learning focused,
link |
02:08:02.320
driving focused supercomputer
link |
02:08:05.200
that then is sufficiently modular
link |
02:08:07.600
that can be scaled to other machine learning applications.
link |
02:08:11.040
So they're not limiting themselves completely
link |
02:08:12.760
to this particular application,
link |
02:08:14.000
but this application is the way they kind of test
link |
02:08:17.440
this iterative process of machine learning
link |
02:08:19.760
is you make a system that's very dumb,
link |
02:08:23.440
deploy it, get the edge cases where it fails,
link |
02:08:27.160
make it a little smarter, it becomes a little less dumb
link |
02:08:30.000
and that iterative process achieves something
link |
02:08:33.000
that you can call intelligent or is smart enough
link |
02:08:36.160
to be able to solve this particular application.
link |
02:08:37.960
So it has to do with training neural networks fast
link |
02:08:43.680
and training neural networks that are large.
link |
02:08:45.920
But also based on an extraordinary amount of diverse input.
link |
02:08:49.920
Data, yeah.
link |
02:08:50.760
And that's one of the things,
link |
02:08:51.920
so this does seem like one of those spaces
link |
02:08:54.520
where the scale of superconducting optoelectronics,
link |
02:08:58.920
the way that, so when you talk about the weaknesses,
link |
02:09:02.520
like I said, okay, well, you have to cool it down.
link |
02:09:04.120
At this scale, that's fine.
link |
02:09:05.760
Because that's not too much of an added cost.
link |
02:09:09.440
Most of your power is being dissipated
link |
02:09:10.960
by the circuits themselves, not the cooling.
link |
02:09:12.920
And also you have one centralized kind of cognitive hub,
link |
02:09:19.400
if you will.
link |
02:09:20.800
And so if we're talking about putting
link |
02:09:24.840
a superconducting system in a car, that's questionable.
link |
02:09:28.640
Do you really wanna cryostat
link |
02:09:30.080
in the trunk of everyone in your car?
link |
02:09:31.400
It'll fit, it's not that big of a deal,
link |
02:09:32.920
but hopefully there's a better way, right?
link |
02:09:35.720
But since this is sort of a central supreme intelligence
link |
02:09:39.080
or something like that,
link |
02:09:40.360
and it needs to really have this massive data acquisition,
link |
02:09:45.120
massive data integration,
link |
02:09:47.080
I would think that that's where large scale
link |
02:09:49.120
spiking neural networks with vast communication
link |
02:09:51.280
and all these things would have something
link |
02:09:53.160
pretty tremendous to offer.
link |
02:09:54.280
It's not gonna happen tomorrow.
link |
02:09:55.760
There's a lot of development that needs to be done.
link |
02:09:58.280
But we have to be patient with self driving cars
link |
02:10:01.440
for a lot of reasons.
link |
02:10:02.280
We were all optimistic that they would be here by now.
link |
02:10:04.600
And okay, they are to some extent,
link |
02:10:06.480
but if we're thinking five or 10 years down the line,
link |
02:10:09.560
it's not unreasonable.
link |
02:10:12.080
One other thing, let me just mention,
link |
02:10:15.200
getting into self driving cars and technologies
link |
02:10:17.400
that are using AI out in the world,
link |
02:10:19.680
this is something NIST cares a lot about.
link |
02:10:21.520
Elham Tabassi is leading up a much larger effort in AI
link |
02:10:25.720
at NIST than my little project.
link |
02:10:29.320
And really central to that mission
link |
02:10:32.680
is this concept of trustworthiness.
link |
02:10:35.000
So when you're going to deploy this neural network
link |
02:10:39.880
in every single automobile with so much on the line,
link |
02:10:43.320
you have to be able to trust that.
link |
02:10:45.240
So now how do we know that we can trust that?
link |
02:10:48.040
How do we know that we can trust the self driving car
link |
02:10:50.080
or the supercomputer that trained it?
link |
02:10:53.720
There's a lot of work there
link |
02:10:54.840
and there's a lot of that going on at NIST.
link |
02:10:56.960
And it's still early days.
link |
02:10:58.200
I mean, you're familiar with the problem and all that.
link |
02:11:01.840
But there's a fascinating dance in engineering
link |
02:11:04.480
with safety critical systems.
link |
02:11:06.680
There's a desire in computer science,
link |
02:11:08.840
just recently talked to Don Knuth,
link |
02:11:13.120
for algorithms and for systems,
link |
02:11:14.800
for them to be provably correct or provably safe.
link |
02:11:17.440
And this is one other difference
link |
02:11:20.320
between humans and biological systems
link |
02:11:22.320
is we're not provably anything.
link |
02:11:24.840
And so there's some aspect of imperfection
link |
02:11:29.760
that we need to have built in,
link |
02:11:32.160
like robustness to imperfection be part of our systems,
link |
02:11:37.920
which is a difficult thing for engineers to contend with.
link |
02:11:40.680
They're very uncomfortable with the idea
link |
02:11:42.800
that you have to be okay with failure
link |
02:11:46.880
and almost engineer failure into the system.
link |
02:11:49.720
Mathematicians hate it too.
link |
02:11:50.960
But I think it was Turing who said something
link |
02:11:53.640
along the lines of,
link |
02:11:55.040
I can give you an intelligent system
link |
02:11:57.080
or I can give you a flawless system,
link |
02:11:59.520
but I can't give you both.
link |
02:12:00.880
And it's in sort of creativity and abstract thinking
link |
02:12:04.120
seem to rely somewhat on stochasticity
link |
02:12:07.840
and not having components
link |
02:12:11.320
that perform exactly the same way every time.
link |
02:12:13.760
This is where like the disagreement I have with,
link |
02:12:16.000
not disagreement, but a different view on the world.
link |
02:12:18.560
I'm with Turing,
link |
02:12:19.680
but when I talk to robotic, robot colleagues,
link |
02:12:24.600
that sounds like I'm talking to robots,
link |
02:12:26.440
colleagues that are roboticists,
link |
02:12:29.760
the goal is perfection.
link |
02:12:31.920
And to me is like, no,
link |
02:12:33.960
I think the goal should be imperfection
link |
02:12:38.880
that's communicated.
link |
02:12:40.960
And through the interaction between humans and robots,
link |
02:12:44.160
that imperfection becomes a feature, not a bug.
link |
02:12:49.600
Like together, seen as a system,
link |
02:12:52.160
the human and the robot together
link |
02:12:53.680
are better than either of them individually,
link |
02:12:56.400
but the robot itself is not perfect in any way.
link |
02:13:00.520
Of course, there's a bunch of disagreements,
link |
02:13:02.760
including with Mr. Elon about,
link |
02:13:06.360
to me, autonomous driving is fundamentally
link |
02:13:08.640
a human robot interaction problem,
link |
02:13:10.760
not a robotics problem.
link |
02:13:12.360
To Elon, it's a robotics problem.
link |
02:13:14.320
That's actually an open and fascinating question,
link |
02:13:18.560
whether humans can be removed from the loop completely.
link |
02:13:24.400
We've talked about a lot of fascinating chemistry
link |
02:13:27.680
and physics and engineering,
link |
02:13:31.240
and we're always running up against this issue
link |
02:13:33.680
that nature seems to dictate what's easy and what's hard.
link |
02:13:37.560
So you have this cool little paper
link |
02:13:40.080
that I'd love to just ask you about.
link |
02:13:43.680
It's titled,
link |
02:13:44.520
Does Cosmological Evolution Select for Technology?
link |
02:13:48.200
So in physics, there's parameters
link |
02:13:53.240
that seem to define the way our universe works,
link |
02:13:56.200
that physics works, that if it worked any differently,
link |
02:13:59.320
we would get a very different world.
link |
02:14:01.720
So it seems like the parameters are very fine tuned
link |
02:14:04.240
to the kind of physics that we see.
link |
02:14:06.480
All the beautiful E equals MC squared,
link |
02:14:08.560
they would get these nice, beautiful laws.
link |
02:14:10.440
It seems like very fine tuned for that.
link |
02:14:13.160
So what you argue in this article
link |
02:14:15.960
is it may be that the universe has also fine tuned
link |
02:14:20.600
its parameters that enable the kind of technological
link |
02:14:25.400
innovation that we see, the technology that we see.
link |
02:14:29.600
Can you explain this idea?
link |
02:14:31.520
Yeah, I think you've introduced it nicely.
link |
02:14:33.440
Let me just try to say a few things in my language layout.
link |
02:14:39.560
What is this fine tuning problem?
link |
02:14:41.680
So physicists have spent centuries trying to understand
link |
02:14:46.560
the system of equations that govern the way nature behaves,
link |
02:14:51.640
the way particles move and interact with each other.
link |
02:14:55.120
And as that understanding has become more clear over time,
link |
02:15:00.120
it became sort of evident that it's all well adjusted
link |
02:15:07.640
to allow a universe like we see, very complex,
link |
02:15:13.040
this large, long lived universe.
link |
02:15:16.480
And so one answer to that is, well, of course it is
link |
02:15:19.920
because we wouldn't be here otherwise.
link |
02:15:21.520
But I don't know, that's not very satisfying.
link |
02:15:24.400
That's sort of, that's what's known
link |
02:15:25.560
as the weak anthropic principle.
link |
02:15:27.240
It's a statement of selection bias.
link |
02:15:29.200
We can only observe a universe that is fit for us to live in.
link |
02:15:33.640
So what does it mean for a universe
link |
02:15:34.960
to be fit for us to live in?
link |
02:15:36.120
Well, the pursuit of physics,
link |
02:15:38.400
it is based partially on coming up with equations
link |
02:15:42.600
that describe how things behave
link |
02:15:44.640
and interact with each other.
link |
02:15:46.280
But in all those equations you have,
link |
02:15:48.480
so there's the form of the equation,
link |
02:15:49.960
sort of how different fields or particles
link |
02:15:54.200
move in space and time.
link |
02:15:56.480
But then there are also the parameters
link |
02:15:58.480
that just tell you sort of the strength
link |
02:16:01.120
of different couplings.
link |
02:16:02.840
How strongly does a charged particle
link |
02:16:05.160
couple to the electromagnetic field or masses?
link |
02:16:07.600
How strongly does a particle couple
link |
02:16:10.760
to the Higgs field or something like that?
link |
02:16:12.960
And those parameters that define,
link |
02:16:16.960
not the general structure of the equations,
link |
02:16:19.760
but the relative importance of different terms,
link |
02:16:23.600
they seem to be every bit as important
link |
02:16:25.240
as the structure of the equations themselves.
link |
02:16:27.760
And so I forget who it was.
link |
02:16:29.400
Somebody, when they were working through this
link |
02:16:31.200
and trying to see, okay, if I adjust the parameter,
link |
02:16:34.000
this parameter over here,
link |
02:16:34.840
call it the, say the fine structure constant,
link |
02:16:36.800
which tells us the strength
link |
02:16:37.920
of the electromagnetic interaction.
link |
02:16:40.400
Oh boy, I can't change it very much.
link |
02:16:42.320
Otherwise nothing works.
link |
02:16:43.720
The universe sort of doesn't,
link |
02:16:45.360
it just pops into existence and goes away
link |
02:16:47.240
in a nanosecond or something like that.
link |
02:16:48.920
And somebody had the phrase,
link |
02:16:51.040
this looks like a put up job,
link |
02:16:52.920
meaning every one of these parameters was dialed in.
link |
02:16:57.080
It's arguable how precisely they have to be dialed in,
link |
02:17:00.680
but dialed in to some extent,
link |
02:17:03.040
not just in order to enable our existence,
link |
02:17:05.360
that's a very anthropocentric view,
link |
02:17:07.120
but to enable a universe like this one.
link |
02:17:10.000
So, okay, maybe I think the majority position
link |
02:17:14.040
of working physicists in the field is,
link |
02:17:17.000
it has to be that way in order for us to exist.
link |
02:17:18.960
We're here, we shouldn't be surprised
link |
02:17:20.400
that that's the way the universe is.
link |
02:17:22.800
And I don't know, for a while,
link |
02:17:24.520
that never sat well with me,
link |
02:17:26.120
but I just kind of moved on
link |
02:17:28.040
because there are things to do
link |
02:17:29.440
and a lot of exciting work.
link |
02:17:31.160
It doesn't depend on resolving this puzzle,
link |
02:17:33.760
but as I started working more with technology,
link |
02:17:39.320
getting into the more recent years of my career,
link |
02:17:41.840
particularly when I started,
link |
02:17:43.720
after having worked with silicon for a long time,
link |
02:17:46.520
which was kind of eerie on its own,
link |
02:17:49.000
but then when I switched over to superconductors,
link |
02:17:51.120
I was just like, this is crazy.
link |
02:17:53.560
It's just absolutely astonishing
link |
02:17:57.360
that our universe gives us superconductivity.
link |
02:18:00.560
It's one of the most beautiful physical phenomena
link |
02:18:02.440
and it's also extraordinarily useful for technology.
link |
02:18:06.560
So you can argue that the universe
link |
02:18:07.920
has to have the parameters it does for us to exist
link |
02:18:11.280
because we couldn't be here otherwise,
link |
02:18:13.000
but why does it give us technology?
link |
02:18:14.760
Why does it give us silicon that has this ideal oxide
link |
02:18:18.840
that allows us to make a transistor
link |
02:18:20.920
without trying that hard?
link |
02:18:23.640
That can't be explained by the same anthropic reasoning.
link |
02:18:27.800
Yeah, so it's asking the why question.
link |
02:18:30.360
I mean, a slight natural extension of that question is,
link |
02:18:34.680
I wonder if the parameters were different
link |
02:18:39.440
if we would simply have just another set of paint brushes
link |
02:18:44.240
to create totally other things
link |
02:18:46.880
that wouldn't look like anything
link |
02:18:49.240
like the technology of today,
link |
02:18:50.880
but would nevertheless have incredible complexity,
link |
02:18:54.520
which is if you sort of zoom out and start defining things,
link |
02:18:57.160
not by like how many batteries it needs
link |
02:19:01.400
and whether it can make toast,
link |
02:19:03.560
but more like how much complexity is within the system
link |
02:19:06.440
or something like that.
link |
02:19:07.280
Well, yeah, you can start to quantify things.
link |
02:19:10.000
You're exactly right.
link |
02:19:10.840
So nowhere am I arguing that
link |
02:19:13.720
in all of the vast parameter space
link |
02:19:15.840
of everything that could conceivably exist
link |
02:19:18.000
in the multiverse of nature,
link |
02:19:20.640
there's this one point in parameter space
link |
02:19:23.280
where complexity arises.
link |
02:19:25.160
I doubt it.
link |
02:19:26.640
That would be a shameful waste of resources, it seems.
link |
02:19:31.120
But it might be that we reside
link |
02:19:33.880
at one place in parameter space
link |
02:19:35.640
that has been adapted through an evolutionary process
link |
02:19:40.040
to allow us to make certain technologies
link |
02:19:43.440
that allow our particular kind of universe to arise
link |
02:19:47.080
and sort of achieve the things it does.
link |
02:19:49.760
See, I wonder if nature in this kind of discussion,
link |
02:19:52.720
if nature is a catalyst for innovation
link |
02:19:55.720
or if it's a ceiling for innovation.
link |
02:19:57.680
So like, is it going to always limit us?
link |
02:20:00.800
Like you're talking about silicon.
link |
02:20:04.000
Is it just make it super easy to do awesome stuff
link |
02:20:06.640
in a certain dimension,
link |
02:20:08.000
but we could still do awesome stuff in other ways,
link |
02:20:10.240
it'll just be harder?
link |
02:20:11.560
Or does it really set like the maximum we can do?
link |
02:20:15.440
That's a good thing to,
link |
02:20:17.840
that's a good subject to discuss.
link |
02:20:19.400
I guess I feel like we need to lay
link |
02:20:20.960
a little bit more groundwork.
link |
02:20:23.160
So I want to make sure that
link |
02:20:27.560
I introduce this in the context
link |
02:20:29.240
of Lee Smolin's previous idea.
link |
02:20:31.800
So who's Lee Smolin and what kind of ideas does he have?
link |
02:20:35.640
Okay, Lee Smolin is a theoretical physicist
link |
02:20:39.000
who back in the late 1980s published a paper
link |
02:20:42.440
in the early 1990s introduced this idea
link |
02:20:45.040
of cosmological natural selection,
link |
02:20:47.000
which argues that the universe did evolve.
link |
02:20:51.440
So his paper was called, did the universe evolve?
link |
02:20:54.480
And I gave myself the liberty of titling my paper
link |
02:20:59.480
does cosmological selection
link |
02:21:01.440
or does cosmological evolution select for technology
link |
02:21:03.960
in reference to that.
link |
02:21:05.000
So he introduced that idea decades ago.
link |
02:21:08.200
Now he primarily works on quantum gravity,
link |
02:21:12.200
loop quantum gravity, other approaches to
link |
02:21:14.640
unifying quantum mechanics with general relativity,
link |
02:21:19.280
as you can read about in his most recent book, I believe,
link |
02:21:22.360
and he's been on your show as well.
link |
02:21:24.280
So, but I want to introduce this idea
link |
02:21:27.760
of cosmological natural selection,
link |
02:21:29.360
because I think that is one of the core ideas
link |
02:21:32.640
that could change our understanding
link |
02:21:35.320
of how the universe got here, our role in it,
link |
02:21:37.800
what technology is doing here.
link |
02:21:39.840
But there's a couple more pieces
link |
02:21:41.240
that need to be set up first.
link |
02:21:42.360
So the beginning of our universe is largely accepted
link |
02:21:46.320
to be the big bang.
link |
02:21:47.360
And what that means is if you look back in time
link |
02:21:49.920
by looking far away in space,
link |
02:21:52.640
you see that everything used to be at one point
link |
02:21:56.960
and it expanded away from there.
link |
02:21:58.920
There was an era in the evolutionary process of our universe
link |
02:22:04.040
that was called inflation.
link |
02:22:05.520
And this idea was developed primarily by Alan Guth
link |
02:22:08.880
and others, Andre Linde and others in the 80s.
link |
02:22:13.120
And this idea of inflation is basically that
link |
02:22:16.040
when a singularity begins this process of growth,
link |
02:22:25.240
there can be a temporary stage
link |
02:22:27.560
where it just accelerates incredibly rapidly.
link |
02:22:30.880
And based on quantum field theory,
link |
02:22:33.760
this tells us that this should produce matter
link |
02:22:35.720
in precisely the proportions that we find
link |
02:22:37.840
of hydrogen and helium in the big bang,
link |
02:22:39.960
lithium too, lithium also, and other things too.
link |
02:22:44.800
So the predictions that come out of big bang
link |
02:22:47.120
inflationary cosmology have stood up extremely well
link |
02:22:50.720
to empirical verification,
link |
02:22:52.520
the cosmic microwave background, things like this.
link |
02:22:55.720
So most scientists working in the field
link |
02:22:59.520
think that the origin of our universe is the big bang.
link |
02:23:03.720
And I base all my thinking on that as well.
link |
02:23:08.040
I'm just laying this out there so that people understand
link |
02:23:11.440
that where I'm coming from is an extension,
link |
02:23:14.160
not a replacement of existing well founded ideas.
link |
02:23:19.080
In a paper, I believe it was 1986 with Alan Guth
link |
02:23:23.800
and another author Farhi,
link |
02:23:26.560
they wrote that a big bang,
link |
02:23:30.240
I don't remember the exact quote,
link |
02:23:31.520
a big bang is inextricably linked with a black hole.
link |
02:23:35.280
The singularity that we call our origin
link |
02:23:39.240
is mathematically indistinguishable from a black hole.
link |
02:23:42.000
They're the same thing.
link |
02:23:44.360
And Lee Smolin based his thinking on that idea,
link |
02:23:48.880
I believe, I don't mean to speak for him,
link |
02:23:50.720
but this is my reading of it.
link |
02:23:52.080
So what Lee Smolin will say is that
link |
02:23:56.040
a black hole in one universe is a big bang
link |
02:23:58.600
in another universe.
link |
02:24:00.720
And this allows us to have progeny, offspring.
link |
02:24:04.680
So a universe can be said to have come
link |
02:24:08.400
before another universe.
link |
02:24:10.480
And very crucially, Smolin argues,
link |
02:24:14.200
I think this is potentially one of the great ideas
link |
02:24:16.920
of all time, that's my opinion,
link |
02:24:18.640
that when a black hole forms, it's not a classical entity,
link |
02:24:22.400
it's a quantum gravitational entity.
link |
02:24:24.520
So it is subject to the fluctuations
link |
02:24:27.240
that are inherent in quantum mechanics, the properties,
link |
02:24:34.080
what we're calling the parameters
link |
02:24:35.440
that describe the physics of that system
link |
02:24:38.600
are subject to slight mutations
link |
02:24:40.680
so that the offspring universe
link |
02:24:42.600
does not have the exact same parameters
link |
02:24:45.200
defining its physics as its parent universe.
link |
02:24:48.440
They're close, but they're a little bit different.
link |
02:24:50.440
And so now you have a mechanism for evolution,
link |
02:24:55.160
for natural selection.
link |
02:24:57.280
So there's mutation, so there's,
link |
02:24:59.680
and then if you think about the DNA of the universe
link |
02:25:03.280
are the basic parameters that govern its laws.
link |
02:25:05.880
Exactly, so what Smolin said is our universe results
link |
02:25:11.560
from an evolutionary process that can be traced back
link |
02:25:14.520
some, he estimated, 200 million generations.
link |
02:25:17.640
Initially, there was something like a vacuum fluctuation
link |
02:25:20.640
that produced through random chance a universe
link |
02:25:25.800
that was able to reproduce just one.
link |
02:25:27.280
So now it had one offspring.
link |
02:25:28.640
And then over time, it was able to make more and more
link |
02:25:30.960
until it evolved into a highly structured universe
link |
02:25:35.280
with a very long lifetime, with a great deal of complexity
link |
02:25:40.160
and importantly, especially importantly for Lee Smolin,
link |
02:25:44.080
stars, stars make black holes.
link |
02:25:47.160
Therefore, we should expect our universe
link |
02:25:49.720
to be optimized, have its physical parameters optimized
link |
02:25:53.160
to make very large numbers of stars
link |
02:25:55.600
because that's how you make black holes
link |
02:25:57.840
and black holes make offspring.
link |
02:25:59.440
So we expect the physics of our universe to have evolved
link |
02:26:03.720
to maximize fecundity, the number of offspring.
link |
02:26:06.720
And the way Lee Smolin argues you do that
link |
02:26:09.200
is through stars that the biggest ones die
link |
02:26:12.160
in these core collapse supernova
link |
02:26:13.440
that make a black hole and a child.
link |
02:26:15.720
Okay, first of all, I agree with you
link |
02:26:19.200
that this is back to our fractal view of everything
link |
02:26:24.760
from intelligence to our universe.
link |
02:26:27.240
That is very compelling and a very powerful idea
link |
02:26:31.120
that unites the origin of life
link |
02:26:36.120
and perhaps the origin of ideas and intelligence.
link |
02:26:39.760
So from a Dawkins perspective here on earth,
link |
02:26:42.200
the evolution of those and then the evolution
link |
02:26:45.360
of the laws of physics that led to us.
link |
02:26:51.000
I mean, it's beautiful.
link |
02:26:52.280
And then you stacking on top of that,
link |
02:26:54.840
that maybe we are one of the offspring.
link |
02:26:57.480
Right, okay, so before getting into where I'd like
link |
02:27:02.320
to take that idea, let me just a little bit more groundwork.
link |
02:27:05.160
There is this concept of the multiverse
link |
02:27:07.080
and it can be confusing.
link |
02:27:08.600
Different people use the word multiverse in different ways.
link |
02:27:11.760
In the multiverse that I think is relevant to picture
link |
02:27:16.920
when trying to grasp Lee Smolin's idea,
link |
02:27:20.840
essentially every vacuum fluctuation
link |
02:27:24.280
can be referred to as a universe.
link |
02:27:25.920
It occurs, it borrows energy from the vacuum
link |
02:27:28.840
for some finite amount of time
link |
02:27:30.200
and it evanesces back into the quantum vacuum.
link |
02:27:33.960
And ideas of Guth before that and Andrei Linde
link |
02:27:38.960
with eternal inflation aren't that different
link |
02:27:42.480
that you would expect nature
link |
02:27:44.480
due to the quantum properties of the vacuum,
link |
02:27:46.920
which we know exist, they're measurable
link |
02:27:49.480
through things like the Casimir effect and others.
link |
02:27:52.200
You know that there are these fluctuations
link |
02:27:54.200
that are occurring.
link |
02:27:55.120
What Smolin is arguing is that there is
link |
02:27:58.080
this extensive multiverse, that this universe,
link |
02:28:01.400
what we can measure and interact with
link |
02:28:04.520
is not unique in nature.
link |
02:28:07.040
It's just our residents, it's where we reside.
link |
02:28:10.800
And there are countless, potentially infinity
link |
02:28:13.640
other universes, other entire evolutionary trajectories
link |
02:28:17.280
that have evolved into things like
link |
02:28:19.360
what you were mentioning a second ago
link |
02:28:21.000
with different parameters and different ways
link |
02:28:24.080
of achieving complexity and reproduction
link |
02:28:26.000
and all that stuff.
link |
02:28:27.040
So it's not that the evolutionary process
link |
02:28:30.480
is a funnel towards this end point, not at all.
link |
02:28:34.240
Just like the biological evolutionary process
link |
02:28:37.000
that has occurred within our universe
link |
02:28:39.320
is not a unique route toward achieving
link |
02:28:42.800
one specific chosen kind of species.
link |
02:28:45.000
No, we have extraordinary diversity around us.
link |
02:28:49.160
That's what evolution does.
link |
02:28:50.520
And for any one species like us,
link |
02:28:52.200
you might feel like we're at the center of this process.
link |
02:28:54.840
We're the destination of this process,
link |
02:28:57.160
but we're just one of the many
link |
02:28:59.480
nearly infinite branches of this process.
link |
02:29:02.240
And I suspect it is exactly infinite.
link |
02:29:04.240
I mean, I just can't understand how with this idea,
link |
02:29:09.080
you can ever draw a boundary around it and say,
link |
02:29:11.040
no, the universe, I mean, the multiverse
link |
02:29:13.480
has 10 to the one quadrillion components,
link |
02:29:17.720
but not infinity.
link |
02:29:18.880
I don't know that.
link |
02:29:20.200
Well, yeah, I have cognitively incapable
link |
02:29:24.080
as I think all of us are
link |
02:29:25.680
and truly understanding the concept of infinity.
link |
02:29:29.080
And the concept of nothing as well.
link |
02:29:31.000
And nothing, but also the concept of a lot
link |
02:29:34.320
is pretty difficult.
link |
02:29:35.360
I can just, I can count.
link |
02:29:37.880
I run out of fingers at a certain point
link |
02:29:39.920
and then you're screwed.
link |
02:29:40.760
And when you're wearing shoes
link |
02:29:41.760
and you can't even get down to your toes, it's like.
link |
02:29:44.400
It's like, all right, a thousand fine, a million.
link |
02:29:47.040
Is that what?
link |
02:29:48.040
And then it gets crazier and crazier.
link |
02:29:50.040
Right, right.
link |
02:29:51.720
So this particular, so when we say technology, by the way,
link |
02:29:55.720
I mean, there's some, not to over romanticize the thing,
link |
02:30:00.640
but there is some aspect about this branch of ours
link |
02:30:04.160
that allows us to, for the universe to know itself.
link |
02:30:08.040
Yes, yes.
link |
02:30:08.960
So to be, to have like little conscious cognitive fingers
link |
02:30:15.080
that are able to feel like to scratch the head.
link |
02:30:18.320
Right, right, right.
link |
02:30:19.640
To be able to construct E equals MC squared
link |
02:30:22.040
and to introspect, to start to gain some understanding
link |
02:30:25.640
of the laws that govern it.
link |
02:30:27.400
Isn't that, isn't that kind of amazing?
link |
02:30:32.000
Okay, I'm just human, but it feels like that,
link |
02:30:35.600
if I were to build a system that does this kind of thing,
link |
02:30:39.200
that evolves laws of physics, that evolves life,
link |
02:30:42.080
that evolves intelligence, that my goal would be
link |
02:30:45.440
to come up with things that are able to think about itself.
link |
02:30:48.960
Right, aren't we kind of close to the design specs,
link |
02:30:53.360
the destination?
link |
02:30:54.680
We're pretty close, I don't know.
link |
02:30:56.040
I mean, I'm spending my career designing things
link |
02:30:58.360
that I hope will think about themselves,
link |
02:30:59.800
so you and I aren't too far apart on that one.
link |
02:31:02.680
But then maybe that problem is a lot harder
link |
02:31:05.520
than we imagine.
link |
02:31:06.360
Maybe we need to.
link |
02:31:07.800
Let's not get, let's not get too far
link |
02:31:09.480
because I want to emphasize something that,
link |
02:31:12.040
what you're saying is, isn't it fascinating
link |
02:31:14.600
that the universe evolved something
link |
02:31:16.880
that can be conscious, reflect on itself?
link |
02:31:19.880
But Lee Smolin's idea didn't take us there, remember?
link |
02:31:23.760
It took us to stars.
link |
02:31:25.720
Lee Smolin has argued, I think,
link |
02:31:29.160
right on almost every single way
link |
02:31:32.080
that cosmological natural selection
link |
02:31:35.760
could lead to a universe with rich structure.
link |
02:31:38.720
And he argued that the structure,
link |
02:31:41.120
the physics of our universe is designed
link |
02:31:43.160
to make a lot of stars so that they can make black holes.
link |
02:31:46.080
But that doesn't explain what we're doing here.
link |
02:31:48.200
In order for that to be an explanation of us,
link |
02:31:51.320
what you have to assume is that once you made that universe
link |
02:31:55.360
that was capable of producing stars,
link |
02:31:58.040
life, planets, all these other things,
link |
02:32:00.280
we're along for the ride.
link |
02:32:01.280
They got lucky.
link |
02:32:02.160
We're kind of arising, growing up in the cracks,
link |
02:32:05.560
but the universe isn't here for us.
link |
02:32:06.880
We're still kind of a fluke in that picture.
link |
02:32:09.200
And I can't, I don't necessarily have
link |
02:32:12.400
like a philosophical opposition to that stance.
link |
02:32:14.840
It's just not, okay, so I don't think it's complete.
link |
02:32:20.200
So it seems like whatever we got going on here to you,
link |
02:32:22.960
it seems like whatever we have here on earth
link |
02:32:25.640
seems like a thing you might want to select for
link |
02:32:28.440
in this whole big process.
link |
02:32:29.680
Exactly.
link |
02:32:30.520
So if what you are truly,
link |
02:32:32.080
if your entire evolutionary process
link |
02:32:34.840
only cares about fecundity,
link |
02:32:36.760
it only cares about making offspring universes
link |
02:32:39.760
because then there's gonna be the most of them
link |
02:32:41.720
in that local region of hyperspace,
link |
02:32:45.120
which is the set of all possible universes, let's say.
link |
02:32:50.240
You don't care how those universes are made.
link |
02:32:52.800
You know they have to be made by black holes.
link |
02:32:54.840
This is what inflationary theory tells us.
link |
02:32:57.920
The big bang tells us that black holes make universes.
link |
02:33:02.040
But what if there was a technological means
link |
02:33:04.200
to make universes?
link |
02:33:05.920
Stars require a ton of matter
link |
02:33:09.280
because they're not thinking very carefully
link |
02:33:11.640
about how you make a black hole.
link |
02:33:12.800
They're just using gravity, you know?
link |
02:33:16.040
But if we devise technologies
link |
02:33:19.160
that can efficiently compress matter into a singularity,
link |
02:33:23.280
it turns out that if you can compress about 10 kilograms
link |
02:33:26.200
into a very small volume,
link |
02:33:28.560
that will make a black hole
link |
02:33:29.720
that is likely highly probable to inflate
link |
02:33:32.360
into its own offspring universe.
link |
02:33:34.720
This is according to calculations done by other people
link |
02:33:37.080
who are professional quantum theorists,
link |
02:33:38.720
quantum field theorists,
link |
02:33:40.160
and I hope I am grasping what they're telling me correctly.
link |
02:33:44.480
I am somewhat of a translator here.
link |
02:33:47.520
But so that's the position
link |
02:33:50.400
that is particularly intriguing to me,
link |
02:33:52.680
which is that what might have happened is that,
link |
02:33:56.240
okay, this particular branch on the vast tree of evolution,
link |
02:34:01.120
cosmological evolution that we're talking about,
link |
02:34:03.200
not biological evolution within our universe,
link |
02:34:05.680
but cosmological evolution,
link |
02:34:07.760
went through exactly the process
link |
02:34:09.560
that Elise Mullen described,
link |
02:34:10.920
got to the stage where stars were making lots of black holes
link |
02:34:15.800
but then continued to evolve and somehow bridged that gap
link |
02:34:19.480
and made intelligence and intelligence
link |
02:34:22.040
capable of devising technologies
link |
02:34:24.080
because technologies, intelligent species
link |
02:34:27.840
working in conjunction with technologies
link |
02:34:29.640
could then produce even more.
link |
02:34:32.080
Yeah, more efficiently, more faster and better
link |
02:34:35.800
and more different.
link |
02:34:36.840
Then you start to have different kind of mechanisms
link |
02:34:38.800
and mutation perhaps, all that kind of stuff.
link |
02:34:40.800
And so if you do a simple calculation that says,
link |
02:34:43.080
all right, if I want to,
link |
02:34:44.960
we know roughly how many core collapse supernovae
link |
02:34:50.920
have resulted in black holes in our galaxy
link |
02:34:54.440
since the beginning of the universe
link |
02:34:55.960
and it's something like a billion.
link |
02:34:57.720
So then you would have to estimate
link |
02:35:00.600
that it would be possible for a technological civilization
link |
02:35:04.200
to produce more than a billion black holes
link |
02:35:07.360
with the energy and matter at their disposal.
link |
02:35:09.920
And so one of the calculations in that paper,
link |
02:35:12.640
back of the envelope,
link |
02:35:14.000
but I think revealing nonetheless is that
link |
02:35:16.320
if you take a relatively common asteroid,
link |
02:35:20.720
something that's about a kilometer in diameter,
link |
02:35:23.640
what I'm thinking of is just scrap material
link |
02:35:26.680
laying around in our solar system
link |
02:35:28.960
and break it up into 10 kilogram chunks
link |
02:35:31.360
and turn each of those into a universe,
link |
02:35:33.320
then you would have made at least a trillion black holes
link |
02:35:38.080
outpacing the star production rate
link |
02:35:41.480
by some three orders of magnitude.
link |
02:35:43.320
That's one asteroid.
link |
02:35:44.760
So now if you envision an intelligent species
link |
02:35:46.840
that would potentially have been devised initially
link |
02:35:50.920
by humans, but then based on superconducting
link |
02:35:53.120
optoelectronic networks, no doubt,
link |
02:35:55.160
and they go out and populate,
link |
02:35:57.080
they don't have to fill the galaxy.
link |
02:35:58.920
They just have to get out to the asteroid belt.
link |
02:36:01.640
They could potentially dramatically outpace
link |
02:36:05.240
the rate at which stars are producing offspring universes.
link |
02:36:07.840
And then wouldn't you expect that
link |
02:36:10.600
that's where we came from instead of a star?
link |
02:36:13.080
Yeah, so you have to somehow become masters of gravity,
link |
02:36:16.520
so like, or generate.
link |
02:36:17.360
John, this is really gravity.
link |
02:36:18.680
So stars make black holes with gravity,
link |
02:36:20.600
but any force that can make the energy density
link |
02:36:26.160
can compactify matter to produce
link |
02:36:28.320
a great enough energy density can form a singularity.
link |
02:36:31.120
It doesn't, it would not likely be gravity.
link |
02:36:33.480
It's the weakest force.
link |
02:36:34.480
You're more likely to use something like the technologies
link |
02:36:38.480
that we're developing for fusion, for example.
link |
02:36:40.480
So I don't know, the Large Ignition Facility
link |
02:36:44.240
recently blasted a pellet with 100 really bright lasers
link |
02:36:50.520
and caused that to get dense enough
link |
02:36:53.360
to engage in nuclear fusion.
link |
02:36:55.280
So something more like that,
link |
02:36:56.560
or a tokamak with a really hot plasma, I'm not sure.
link |
02:36:59.200
Something, I don't know exactly how it would be done.
link |
02:37:02.160
I do like the idea of that,
link |
02:37:04.560
especially just been reading a lot about gravitational waves
link |
02:37:07.120
and the fact that us humans with our technological
link |
02:37:10.600
capabilities, one of the most impressive
link |
02:37:14.120
technological accomplishments of human history is LIGO,
link |
02:37:17.360
being able to precisely detect gravitational waves.
link |
02:37:20.720
I'm particularly find appealing the idea
link |
02:37:25.000
that other alien civilizations from very far distances
link |
02:37:29.520
communicate with gravity, with gravitational waves,
link |
02:37:34.360
because as you become greater and greater master of gravity,
link |
02:37:37.800
which seems way out of reach for us right now,
link |
02:37:40.600
maybe that seems like a effective way of sending signals,
link |
02:37:44.280
especially if your job is to manufacture black holes.
link |
02:37:48.440
Right.
link |
02:37:49.280
So that, so let me ask there,
link |
02:37:53.360
whatever, I mean, broadly thinking,
link |
02:37:56.440
because we tend to think other alien civilizations
link |
02:37:58.920
would be very human like,
link |
02:38:00.000
but if we think of alien civilizations out there
link |
02:38:04.080
as basically generators of black holes,
link |
02:38:07.640
however they do it, because they got stars,
link |
02:38:10.960
do you think there's a lot of them
link |
02:38:12.800
in our particular universe out there?
link |
02:38:17.760
In our universe?
link |
02:38:20.480
Well, okay, let me ask, okay, this is great.
link |
02:38:23.400
Let me ask a very generic question
link |
02:38:26.680
and then let's see how you answer it,
link |
02:38:29.200
which is how many alien civilizations are out there?
link |
02:38:35.080
If the hypothesis that I just described
link |
02:38:38.400
is on the right track,
link |
02:38:40.840
it would mean that the parameters of our universe
link |
02:38:43.880
have been selected so that intelligent civilizations
link |
02:38:48.480
will occur in sufficient numbers
link |
02:38:51.120
so that if they reach something
link |
02:38:54.720
like supreme technological maturity,
link |
02:38:56.480
let's define that as the ability to produce black holes,
link |
02:39:00.080
then that's not a highly improbable event.
link |
02:39:02.640
It doesn't need to happen often
link |
02:39:05.440
because as I just described,
link |
02:39:06.680
if you get one of them in a galaxy,
link |
02:39:09.200
you're gonna make more black holes
link |
02:39:10.520
than the stars in that galaxy.
link |
02:39:12.720
But there's also not a super strong motivation,
link |
02:39:16.560
well, it's not obvious that you need them
link |
02:39:21.600
to be ubiquitous throughout the galaxy.
link |
02:39:23.920
Right.
link |
02:39:24.760
One of the things that I try to emphasize in that paper
link |
02:39:27.520
is that given this idea
link |
02:39:30.640
of how our parameters might've been selected,
link |
02:39:35.120
it's clear that it's a series of trade offs, right?
link |
02:39:39.280
If you make, I mean, in order for intelligent life
link |
02:39:42.040
of our variety or anything resembling us to occur,
link |
02:39:45.760
you need a bunch of stuff, you need stars.
link |
02:39:47.600
So that's right back to Smolin's roots of this idea,
link |
02:39:51.080
but you also need water to have certain properties.
link |
02:39:54.440
You need things like the rocky planets,
link |
02:39:58.760
like the Earth to be within the habitable zone,
link |
02:40:00.600
all these things that you start talking about
link |
02:40:02.680
in the field of astrobiology,
link |
02:40:06.760
trying to understand life in the universe,
link |
02:40:08.800
but you can't over emphasize,
link |
02:40:10.480
you can't tune the parameters so precisely
link |
02:40:13.600
to maximize the number of stars
link |
02:40:15.160
or to give water exactly the properties
link |
02:40:18.960
or to make rocky planets like Earth the most numerous.
link |
02:40:22.280
You have to compromise on all these things.
link |
02:40:24.480
And so I think the way to test this idea
link |
02:40:27.360
is to look at what parameters are necessary
link |
02:40:30.240
for each of these different subsystems,
link |
02:40:32.760
and I've laid out a few that I think are promising,
link |
02:40:35.080
there could be countless others,
link |
02:40:36.640
and see how changing the parameters
link |
02:40:40.840
makes it more or less likely that stars would form
link |
02:40:43.600
and have long lifetimes or that rocky planets
link |
02:40:46.320
in the habitable zone are likely to form,
link |
02:40:48.280
all these different things.
link |
02:40:49.400
So we can test how much these things are in a tug of war
link |
02:40:53.360
with each other, and the prediction would be
link |
02:40:56.280
that we kind of sit at this central point
link |
02:40:58.160
where if you move the parameters too much,
link |
02:41:02.040
stars aren't stable, or life doesn't form,
link |
02:41:05.680
or technology's infeasible,
link |
02:41:07.720
because life alone, at least the kind of life
link |
02:41:10.760
that we know of, cannot make black holes.
link |
02:41:14.160
We don't have this, well, I'm speaking for myself,
link |
02:41:16.400
you're a very fit and strong person,
link |
02:41:18.560
but it might be possible for you,
link |
02:41:20.600
but not for me to compress matter.
link |
02:41:22.360
So we need these technologies, but we don't know,
link |
02:41:25.960
we have not been able to quantify yet
link |
02:41:28.960
how finely adjusted the parameters would need to be
link |
02:41:33.840
in order for silicon to have the properties it does.
link |
02:41:35.600
Okay, this is not directly speaking to what you're saying,
link |
02:41:37.840
you're getting to the Fermi paradox,
link |
02:41:39.520
which is where are they, where are the life forms out there,
link |
02:41:42.840
how numerous are they, that sort of thing.
link |
02:41:44.560
What I'm trying to argue is that
link |
02:41:46.200
if this framework is on the right track,
link |
02:41:50.800
a potentially correct explanation for our existence,
link |
02:41:53.720
we, it doesn't necessarily predict
link |
02:41:56.320
that intelligent civilizations are just everywhere,
link |
02:41:59.280
because even if you just get one of them in a galaxy,
link |
02:42:02.520
which is quite rare, it could be enough
link |
02:42:05.760
to dramatically increase the fecundity
link |
02:42:08.920
of the universe as a whole.
link |
02:42:10.160
Yeah, and I wonder, once you start generating
link |
02:42:12.440
the offspring for universes, black holes,
link |
02:42:15.400
how that has effect on the,
link |
02:42:18.360
what kind of effect does it have
link |
02:42:19.920
on the other candidate's civilizations
link |
02:42:24.920
within that universe?
link |
02:42:26.160
Maybe it has a destructive aspect,
link |
02:42:28.400
or there could be some arguments
link |
02:42:29.840
about once you have a lot of offspring,
link |
02:42:32.040
that that just quickly accelerates
link |
02:42:34.040
to where the other ones can't even catch up.
link |
02:42:35.760
It could, but I guess if you want me
link |
02:42:39.160
to put my chips on the table or whatever,
link |
02:42:42.520
I think I come down more on the side
link |
02:42:46.320
that intelligent life civilizations are rare.
link |
02:42:52.400
And I guess I follow Max Tegmark here.
link |
02:42:57.160
And also there's a lot of papers coming out recently
link |
02:43:01.000
in the field of astrobiology that are seeming to say,
link |
02:43:04.360
all right, you just work through the numbers
link |
02:43:06.000
on some modified Drake equation or something like that.
link |
02:43:09.680
And it looks like it's not improbable.
link |
02:43:13.040
You shouldn't be surprised that an intelligent species
link |
02:43:16.280
has arisen in our galaxy,
link |
02:43:18.040
but if you think there's one the next solar system over,
link |
02:43:20.360
it's highly improbable.
link |
02:43:21.600
So I can see that the number,
link |
02:43:23.960
the probability of finding a civilization in a galaxy,
link |
02:43:28.280
maybe it's most likely that you're gonna find
link |
02:43:31.080
one to a hundred or something.
link |
02:43:32.880
But okay, now it's really important
link |
02:43:34.960
to put a time window on that, I think,
link |
02:43:36.720
because does that mean in the entire lifetime of the galaxy
link |
02:43:40.720
before it, so for in our case, before we run into Andromeda,
link |
02:43:49.760
I think it's highly probable, I shouldn't say I think,
link |
02:43:53.960
it's tempting to believe that it's highly probable
link |
02:43:56.760
that in that entire lifetime of your galaxy,
link |
02:44:00.000
you're gonna get at least one intelligent species,
link |
02:44:02.600
maybe thousands or something like that.
link |
02:44:05.160
But it's also, I think, a little bit naive to think
link |
02:44:10.400
that they're going to coincide in time
link |
02:44:13.200
and we'll be able to observe them.
link |
02:44:14.960
And also, if you look at the span of life on Earth,
link |
02:44:20.080
the Earth history, it was surprising to me
link |
02:44:24.480
to kind of look at the amount of time,
link |
02:44:27.880
first of all, the short amount of time,
link |
02:44:29.680
there's no life, it's surprising.
link |
02:44:31.520
Life sprang up pretty quickly.
link |
02:44:33.440
It's single cell.
link |
02:44:35.280
But that's the point I'm trying to make
link |
02:44:36.960
is like so much of life on Earth
link |
02:44:42.040
was just like single cell organisms, like most of it.
link |
02:44:45.840
Most of it was like boring bacteria type of stuff.
link |
02:44:48.640
Well, bacteria are fascinating, but I take your point.
link |
02:44:50.640
No, I get it.
link |
02:44:51.520
I mean, no offense to them.
link |
02:44:52.880
But this kind of speaking from the perspective
link |
02:44:56.840
of your paper of something that's able
link |
02:44:58.720
to generate technology as we kind of understand it,
link |
02:45:01.440
that's a very short moment in time
link |
02:45:03.400
relative to that full history of life on Earth.
link |
02:45:08.640
And maybe our universe is just saturated
link |
02:45:12.200
with bacteria like humans.
link |
02:45:15.880
Right.
link |
02:45:17.480
But not the special extra AGI super humans,
link |
02:45:24.200
that those are very rare.
link |
02:45:25.560
And once those spring up, everything just goes to like,
link |
02:45:30.400
it accelerates very quickly.
link |
02:45:33.360
Yeah, we just don't have enough data to really say,
link |
02:45:36.520
but I find this whole subject extremely engaging.
link |
02:45:40.000
I mean, there's this concept,
link |
02:45:41.680
I think it's called the Rare Earth Hypothesis,
link |
02:45:43.880
which is that basically stating that,
link |
02:45:46.920
okay, microbes were here right away
link |
02:45:49.040
after the Hadian era where we were being bombarded.
link |
02:45:52.120
Well, after, yeah, bombarded by comets, asteroids,
link |
02:45:54.920
things like that, and also after the moon formed.
link |
02:45:57.080
So once things settled down a little bit,
link |
02:45:59.560
in a few hundred million years,
link |
02:46:02.280
you have microbes everywhere.
link |
02:46:03.680
And it could have been, we don't know exactly
link |
02:46:05.200
when it could have been remarkably brief that that took.
link |
02:46:08.080
So it does indicate that, okay,
link |
02:46:10.200
life forms relatively easily.
link |
02:46:12.160
I think that alone is sort of a checker on the scale
link |
02:46:15.920
for the argument that the parameters that allow
link |
02:46:21.480
even microbial life to form are not just a fluke.
link |
02:46:24.360
But anyway, that aside, yes,
link |
02:46:27.520
then there was this long dormant period,
link |
02:46:29.880
not dormant, things were happening,
link |
02:46:31.600
but important things were happening
link |
02:46:34.120
for some two and a half billion years or something
link |
02:46:37.320
after the metabolic process
link |
02:46:40.160
that releases oxygen was developed.
link |
02:46:42.840
Then basically the planet's just sitting there,
link |
02:46:46.120
getting more and more oxygenated,
link |
02:46:47.560
more and more oxygenated until it's enough
link |
02:46:50.200
that you can build these large, complex organisms.
link |
02:46:54.160
And so the Rare Earth Hypothesis would argue
link |
02:46:56.480
that the microbes are common everywhere
link |
02:47:01.000
in any planet that's roughly in the habitable zone
link |
02:47:04.160
and has some water on it, it's probably gonna have those.
link |
02:47:06.640
But then getting to this Cambrian explosion
link |
02:47:09.320
that happened some between 500 and 600 million years ago,
link |
02:47:13.760
that's rare, you know?
link |
02:47:16.360
And I buy that, I think that is rare.
link |
02:47:19.080
So if you say how much life is in our galaxy,
link |
02:47:21.880
I think that's probably the right answer
link |
02:47:24.120
is that microbes are everywhere.
link |
02:47:26.440
Cambrian explosion is extremely rare.
link |
02:47:29.360
And then, but the Cambrian explosion kind of went like that
link |
02:47:32.760
where within a couple of tens or a hundred million years,
link |
02:47:38.600
all of these body plans came into existence.
link |
02:47:40.960
And basically all of the body plans
link |
02:47:43.120
that are now in existence on the planet
link |
02:47:46.080
were formed in that brief window
link |
02:47:48.720
and we've just been shuffling around since then.
link |
02:47:51.640
So then what caused humans to pop out of that?
link |
02:47:54.840
I mean, that could be another extremely rare threshold
link |
02:48:01.920
that a planet roughly in the habitable zone with water
link |
02:48:06.200
is not guaranteed to cross, you know?
link |
02:48:08.400
To me, it's fascinating for being humble,
link |
02:48:10.200
like the humans cannot possibly be the most amazing thing
link |
02:48:13.080
that such, if you look at the entirety of the system
link |
02:48:15.880
that Lee Smolin and you paint,
link |
02:48:17.800
that cannot possibly be the most amazing thing
link |
02:48:20.080
that process generates.
link |
02:48:21.480
So like, if you look at the evolution,
link |
02:48:23.720
what's the equivalent in the cosmological evolution
link |
02:48:27.040
and its selection for technology,
link |
02:48:29.000
the equivalent of the human eye or the human brain?
link |
02:48:32.440
Universes that are able to do some like,
link |
02:48:35.600
they don't need the damn stars.
link |
02:48:37.760
They're able to just do some incredible generation
link |
02:48:42.120
of complexity fast, like much more than,
link |
02:48:46.640
if you think about it,
link |
02:48:47.520
it's like most of our universe is pretty freaking boring.
link |
02:48:50.920
There's not much going on, there's a few rocks flying around
link |
02:48:53.400
and there's some like apes
link |
02:48:54.520
that are just like doing podcasts on some weird planet.
link |
02:49:00.520
It just seems very inefficient.
link |
02:49:02.840
If you think about like the amazing thing in the human eye,
link |
02:49:05.960
the visual cortex can do, the brain, the nervous,
link |
02:49:09.440
everything that makes us more powerful
link |
02:49:12.840
than single cell organisms.
link |
02:49:15.480
Like if there's an equivalent of that for universes,
link |
02:49:19.320
like the richness of physics
link |
02:49:21.680
that could be expressed
link |
02:49:24.640
through a particular set of parameters.
link |
02:49:26.840
Like, I mean, like for me,
link |
02:49:31.040
I'm a sort of from a computer science perspective,
link |
02:49:33.760
huge fan of cellular automata,
link |
02:49:35.600
which is a nice sort of pretty visual way
link |
02:49:39.320
to illustrate how different laws
link |
02:49:42.160
can result in drastically different levels of complexity.
link |
02:49:46.400
So like, it's like, yeah, okay.
link |
02:49:49.000
So we're all like celebrating,
link |
02:49:50.320
look, our little cellular automata
link |
02:49:52.120
is able to generate pretty triangles and squares
link |
02:49:54.560
and therefore we achieve general intelligence.
link |
02:49:57.600
And then there'll be like some badass Chuck Norris type,
link |
02:50:01.840
like universal Turing machine type of cellular automata.
link |
02:50:06.560
They're able to generate other cellular automata
link |
02:50:09.560
that does any arbitrary level of computation off the bat.
link |
02:50:14.160
Like those have to then exist.
link |
02:50:16.480
And then we're just like, we'll be forgotten.
link |
02:50:19.840
This story, this podcast just entertains
link |
02:50:23.800
a few other apes for a few months.
link |
02:50:26.880
Well, I'm kind of surprised to hear your cynicism.
link |
02:50:30.240
No, I'm very up.
link |
02:50:32.080
I usually think of you as like one who celebrates humanity
link |
02:50:36.120
and all its forms and things like that.
link |
02:50:37.600
And I guess I just, I don't,
link |
02:50:39.240
I see it the way you just described.
link |
02:50:41.000
I mean, okay, we've been here for 13.7 billion years
link |
02:50:44.520
and you're saying, gosh, that's a long time.
link |
02:50:47.240
Let's get on with the show already.
link |
02:50:48.480
Some other universe could have kicked our butt by now,
link |
02:50:51.520
but that's putting a characteristic time.
link |
02:50:55.360
I mean, why is 13.7 billion a long time?
link |
02:50:58.400
I mean, compared to what?
link |
02:51:00.440
I guess, so when I look at our universe,
link |
02:51:02.320
I see this extraordinary hierarchy
link |
02:51:05.400
that has developed over that time.
link |
02:51:08.080
So at the beginning, it was a chaotic mess of some plasma
link |
02:51:13.720
and nothing interesting going on there.
link |
02:51:16.080
And even for the first stars to form,
link |
02:51:18.880
that a lot of really interesting evolutionary processes
link |
02:51:23.880
had to occur, by evolutionary in that sense,
link |
02:51:26.240
I just mean taking place over extended periods of time
link |
02:51:30.520
and structures are forming then.
link |
02:51:32.000
And then it took that first generation of stars
link |
02:51:34.640
in order to produce the metals
link |
02:51:38.760
that then can more efficiently produce
link |
02:51:41.280
another generation of stars.
link |
02:51:42.440
We're only the third generation of stars.
link |
02:51:44.720
So we might still be pretty quick to the game here.
link |
02:51:47.920
So, but I don't think, I don't, okay.
link |
02:51:51.000
So then you have these stars
link |
02:51:52.440
and then you have solar systems on those solar systems.
link |
02:51:54.640
You have rocky worlds, you have gas giants,
link |
02:51:58.840
like all this complexity.
link |
02:51:59.800
And then you start getting life
link |
02:52:01.120
and the complexity that's evolved
link |
02:52:03.800
through the evolutionary process in life forms
link |
02:52:06.440
is just, it's not a let down to me.
link |
02:52:09.880
Just seeing that.
link |
02:52:10.720
Some of it is like some of the planets is like icy,
link |
02:52:14.600
it's like different flavors of ice cream.
link |
02:52:16.080
They're icy, but there might be water underneath.
link |
02:52:18.560
All kinds of life forms with some volcanoes,
link |
02:52:21.120
all kinds of weird stuff.
link |
02:52:22.320
No, no, I don't, I think it's beautiful.
link |
02:52:24.600
I think our life is beautiful.
link |
02:52:25.960
And I think it was designed that by design,
link |
02:52:29.560
the scarcity of the whole thing.
link |
02:52:31.200
I think mortality, as terrifying as it is,
link |
02:52:33.840
is fundamental to the whole reason we enjoy everything.
link |
02:52:37.280
No, I think it's beautiful.
link |
02:52:38.200
I just think that all of us conscious beings
link |
02:52:42.400
in the grand scheme of basically every scale
link |
02:52:45.520
will be completely forgotten.
link |
02:52:46.960
Well, that's true.
link |
02:52:47.800
I think everything is transient
link |
02:52:49.320
and that would go back to maybe something more like Lao Tzu,
link |
02:52:52.520
the Tao Te Ching or something where it's like,
link |
02:52:55.440
yes, there is nothing but change.
link |
02:52:57.720
There is nothing but emergence and dissolve and that's it.
link |
02:53:00.720
But I just, in this picture,
link |
02:53:03.000
this hierarchy that's developed,
link |
02:53:04.520
I don't mean to say that now it gets to us
link |
02:53:06.600
and that's the pinnacle.
link |
02:53:07.440
In fact, I think at a high level,
link |
02:53:10.520
the story I'm trying to tease out in my research is about,
link |
02:53:14.240
okay, well, so then what's the next level of hierarchy?
link |
02:53:17.000
And if it's, okay, we're kind of pretty smart.
link |
02:53:21.520
I mean, talking about people like Lee Small
link |
02:53:23.840
and Alan Guth, Max Tegmark, okay, we're really smart.
link |
02:53:26.240
Talking about me, okay, we're kind of,
link |
02:53:28.480
we can find our way to the grocery store or whatever,
link |
02:53:30.760
but what's next?
link |
02:53:33.000
I mean, what if there's another level of hierarchy
link |
02:53:36.120
that grows on top of us
link |
02:53:37.760
that is even more profoundly capable?
link |
02:53:40.280
And I mean, we've talked a lot
link |
02:53:42.080
about superconducting sensors.
link |
02:53:43.560
Imagine these cognitive systems far more capable than us
link |
02:53:48.920
residing somewhere else in the solar system
link |
02:53:52.320
off of the surface of the earth,
link |
02:53:53.640
where it's much darker, much colder,
link |
02:53:55.400
much more naturally suited to them.
link |
02:53:57.080
And they have these sensors that can detect single photons
link |
02:54:00.440
of light from radio waves out to all across the spectrum
link |
02:54:04.560
of the gamma rays and just see the whole universe.
link |
02:54:07.400
And they just live in space
link |
02:54:08.960
with these massive collection optics so that they,
link |
02:54:12.960
what do they do?
link |
02:54:13.800
They just look out and experience that vast array
link |
02:54:18.800
of what's being developed.
link |
02:54:22.520
And if you're such a system,
link |
02:54:25.120
presumably you would do some things for fun.
link |
02:54:28.920
And the kind of fun thing I would do
link |
02:54:31.720
as somebody who likes video games
link |
02:54:33.840
is I would create and maintain
link |
02:54:37.040
and observe something like earth.
link |
02:54:42.760
So in some sense, we're like all what players on a stage
link |
02:54:47.320
for this superconducting cold computing system out there.
link |
02:54:54.160
I mean, all of this is fascinating to think.
link |
02:54:56.680
The fact that you're actually designing systems
link |
02:54:59.360
here on earth that are trying to push this technological
link |
02:55:01.560
at the very cutting edge and also thinking about
link |
02:55:04.800
how does the like the evolution of physical laws
link |
02:55:09.760
lead us to the way we are is fascinating.
link |
02:55:14.240
That coupling is fascinating.
link |
02:55:15.920
It's like the ultimate rigorous application of philosophy
link |
02:55:20.800
to the rigorous application of engineering.
link |
02:55:23.680
So Jeff, you're one of the most fascinating.
link |
02:55:26.400
I'm so glad I did not know much about you
link |
02:55:29.000
except through your work.
link |
02:55:30.440
And I'm so glad we got this chance to talk.
link |
02:55:34.200
You're one of the best explainers
link |
02:55:37.800
of exceptionally difficult concepts.
link |
02:55:40.940
And you're also, speaking of like fractal,
link |
02:55:44.600
you're able to function intellectually
link |
02:55:46.680
at all levels of the stack, which I deeply appreciate.
link |
02:55:50.240
This was really fun.
link |
02:55:51.640
You're a great educator, a great scientist.
link |
02:55:53.600
It's an honor that you would spend
link |
02:55:56.080
your valuable time with me.
link |
02:55:57.280
It's an honor that you would spend your time with me as well.
link |
02:56:00.120
Thanks, Jeff.
link |
02:56:01.760
Thanks for listening to this conversation
link |
02:56:03.560
with Jeff Schoenlein.
link |
02:56:05.240
To support this podcast,
link |
02:56:06.680
please check out our sponsors in the description.
link |
02:56:09.580
And now let me leave you with some words
link |
02:56:12.060
from the great John Carmack,
link |
02:56:14.320
who surely will be a guest on this podcast soon.
link |
02:56:18.220
Because of the nature of Moore's Law,
link |
02:56:20.100
anything that an extremely clever graphics programmer
link |
02:56:22.800
can do at one point can be replicated
link |
02:56:26.040
by a merely competent programmer
link |
02:56:27.920
some number of years later.
link |
02:56:30.520
Thank you for listening and hope to see you next time.