back to indexDavid Ferrucci: IBM Watson, Jeopardy & Deep Conversations with AI | Lex Fridman Podcast #44
link |
The following is a conversation with David Feroci.
link |
He led the team that built Watson,
link |
the IBM question answering system
link |
that beat the top humans in the world
link |
at the game of Jeopardy.
link |
For spending a couple hours with David,
link |
I saw a genuine passion,
link |
not only for abstract understanding of intelligence,
link |
but for engineering it to solve real world problems
link |
under real world deadlines and resource constraints.
link |
Where science meets engineering
link |
is where brilliant, simple ingenuity emerges.
link |
People who work adjoining it to
link |
have a lot of wisdom earned
link |
through failures and eventual success.
link |
David is also the founder, CEO,
link |
and chief scientist of Elemental Cognition,
link |
a company working to engineer AI systems
link |
that understand the world the way people do.
link |
This is the Artificial Intelligence Podcast.
link |
If you enjoy it, subscribe on YouTube,
link |
give it five stars on iTunes,
link |
support it on Patreon,
link |
or simply connect with me on Twitter
link |
at Lex Friedman, spelled F R I D M A N.
link |
And now, here's my conversation with David Ferrucci.
link |
Your undergrad was in biology
link |
with an eye toward medical school
link |
before you went on for the PhD in computer science.
link |
So let me ask you an easy question.
link |
What is the difference between biological systems
link |
and computer systems?
link |
In your, when you sit back,
link |
look at the stars, and think philosophically.
link |
I often wonder whether or not
link |
there is a substantive difference.
link |
I mean, I think the thing that got me
link |
into computer science and into artificial intelligence
link |
was exactly this presupposition
link |
that if we can get machines to think,
link |
or I should say this question,
link |
this philosophical question,
link |
if we can get machines to think,
link |
to understand, to process information the way we do,
link |
so if we can describe a procedure, describe a process,
link |
even if that process were the intelligence process itself,
link |
then what would be the difference?
link |
So from a philosophical standpoint,
link |
I'm not sure I'm convinced that there is.
link |
I mean, you can go in the direction of spirituality,
link |
you can go in the direction of the soul,
link |
but in terms of what we can experience
link |
from an intellectual and physical perspective,
link |
I'm not sure there is.
link |
Clearly, there are different implementations,
link |
but if you were to say,
link |
is a biological information processing system
link |
fundamentally more capable
link |
than one we might be able to build out of silicon
link |
or some other substrate,
link |
I don't know that there is.
link |
How distant do you think is the biological implementation?
link |
So fundamentally, they may have the same capabilities,
link |
but is it really a far mystery
link |
where a huge number of breakthroughs are needed
link |
to be able to understand it,
link |
or is it something that, for the most part,
link |
in the important aspects,
link |
echoes of the same kind of characteristics?
link |
Yeah, that's interesting.
link |
I mean, so your question presupposes
link |
that there's this goal to recreate
link |
what we perceive as biological intelligence.
link |
I'm not sure that's the,
link |
I'm not sure that's how I would state the goal.
link |
I mean, I think that studying.
link |
Good, so I think there are a few goals.
link |
I think that understanding the human brain
link |
and how it works is important
link |
for us to be able to diagnose and treat issues,
link |
treat issues for us to understand our own strengths
link |
and weaknesses, both intellectual,
link |
psychological, and physical.
link |
So neuroscience and understanding the brain,
link |
from that perspective, there's a clear goal there.
link |
From the perspective of saying,
link |
I wanna mimic human intelligence,
link |
that one's a little bit more interesting.
link |
Human intelligence certainly has a lot of things we envy.
link |
It's also got a lot of problems, too.
link |
So I think we're capable of sort of stepping back
link |
and saying, what do we want out of an intelligence?
link |
How do we wanna communicate with that intelligence?
link |
How do we want it to behave?
link |
How do we want it to perform?
link |
Now, of course, it's somewhat of an interesting argument
link |
because I'm sitting here as a human
link |
with a biological brain,
link |
and I'm critiquing the strengths and weaknesses
link |
of human intelligence and saying
link |
that we have the capacity to step back
link |
and say, gee, what is intelligence
link |
and what do we really want out of it?
link |
And that, in and of itself, suggests that
link |
human intelligence is something quite enviable,
link |
that it can introspect that way.
link |
And the flaws, you mentioned the flaws.
link |
Humans have flaws.
link |
Yeah, but I think that flaws that human intelligence has
link |
is extremely prejudicial and biased
link |
in the way it draws many inferences.
link |
Do you think those are, sorry to interrupt,
link |
do you think those are features or are those bugs?
link |
Do you think the prejudice, the forgetfulness, the fear,
link |
what are the flaws?
link |
Maybe that's a flaw.
link |
You think those are all things that can be gotten,
link |
get in the way of intelligence
link |
or the essential components of intelligence?
link |
Well, again, if you go back and you define intelligence
link |
as being able to sort of accurately, precisely, rigorously,
link |
reason, develop answers,
link |
and justify those answers in an objective way,
link |
yeah, then human intelligence has these flaws
link |
in that it tends to be more influenced
link |
by some of the things you said.
link |
And it's largely an inductive process,
link |
meaning it takes past data,
link |
uses that to predict the future.
link |
Very advantageous in some cases,
link |
but fundamentally biased and prejudicial in other cases
link |
because it's gonna be strongly influenced by its priors,
link |
whether they're right or wrong
link |
from some objective reasoning perspective,
link |
you're gonna favor them because those are the decisions
link |
or those are the paths that succeeded in the past.
link |
And I think that mode of intelligence makes a lot of sense
link |
for when your primary goal is to act quickly
link |
and survive and make fast decisions.
link |
And I think those create problems
link |
when you wanna think more deeply
link |
and make more objective and reasoned decisions.
link |
Of course, humans capable of doing both.
link |
They do sort of one more naturally than they do the other,
link |
but they're capable of doing both.
link |
You're saying they do the one
link |
that responds quickly more naturally.
link |
Because that's the thing we kind of need
link |
to not be eaten by the predators in the world.
link |
For example, but then we've learned to reason through logic,
link |
we've developed science, we train people to do that.
link |
I think that's harder for the individual to do.
link |
I think it requires training and teaching.
link |
I think we are, the human mind certainly is capable of it,
link |
but we find it more difficult.
link |
And then there are other weaknesses, if you will,
link |
as you mentioned earlier, just memory capacity
link |
and how many chains of inference
link |
can you actually go through without like losing your way?
link |
So just focus and...
link |
So the way you think about intelligence,
link |
and we're really sort of floating
link |
in this philosophical space,
link |
but I think you're like the perfect person
link |
to talk about this,
link |
because we'll get to Jeopardy and beyond.
link |
That's like one of the most incredible accomplishments
link |
in AI, in the history of AI,
link |
but hence the philosophical discussion.
link |
So let me ask, you've kind of alluded to it,
link |
but let me ask again, what is intelligence?
link |
Underlying the discussions we'll have
link |
with Jeopardy and beyond,
link |
how do you think about intelligence?
link |
Is it a sufficiently complicated problem
link |
being able to reason your way through solving that problem?
link |
Is that kind of how you think about
link |
what it means to be intelligent?
link |
So I think of intelligence primarily two ways.
link |
One is the ability to predict.
link |
So in other words, if I have a problem,
link |
can I predict what's gonna happen next?
link |
Whether it's to predict the answer of a question
link |
or to say, look, I'm looking at all the market dynamics
link |
and I'm gonna tell you what's gonna happen next,
link |
or you're in a room and somebody walks in
link |
and you're gonna predict what they're gonna do next
link |
or what they're gonna say next.
link |
You're in a highly dynamic environment
link |
full of uncertainty, be able to predict.
link |
The more variables, the more complex.
link |
The more possibilities, the more complex.
link |
But can I take a small amount of prior data
link |
and learn the pattern and then predict
link |
what's gonna happen next accurately and consistently?
link |
That's certainly a form of intelligence.
link |
What do you need for that, by the way?
link |
You need to have an understanding
link |
of the way the world works
link |
in order to be able to unroll it into the future, right?
link |
What do you think is needed to predict?
link |
Depends what you mean by understanding.
link |
I need to be able to find that function.
link |
This is very much what deep learning does,
link |
machine learning does, is if you give me enough prior data
link |
and you tell me what the output variable is that matters,
link |
I'm gonna sit there and be able to predict it.
link |
And if I can predict it accurately
link |
so that I can get it right more often than not,
link |
I'm smart, if I can do that with less data
link |
and less training time, I'm even smarter.
link |
If I can figure out what's even worth predicting,
link |
I'm smarter, meaning I'm figuring out
link |
what path is gonna get me toward a goal.
link |
What about picking a goal?
link |
Sorry, you left again.
link |
Well, that's interesting about picking a goal,
link |
sort of an interesting thing.
link |
I think that's where you bring in
link |
what are you preprogrammed to do?
link |
We talk about humans,
link |
and well, humans are preprogrammed to survive.
link |
So it's sort of their primary driving goal.
link |
What do they have to do to do that?
link |
And that can be very complex, right?
link |
So it's not just figuring out that you need to run away
link |
from the ferocious tiger,
link |
but we survive in a social context as an example.
link |
So understanding the subtleties of social dynamics
link |
becomes something that's important for surviving,
link |
finding a mate, reproducing, right?
link |
So we're continually challenged with
link |
complex sets of variables, complex constraints,
link |
rules, if you will, or patterns.
link |
And we learn how to find the functions
link |
and predict the things.
link |
In other words, represent those patterns efficiently
link |
and be able to predict what's gonna happen.
link |
And that's a form of intelligence.
link |
That doesn't really require anything specific
link |
other than the ability to find that function
link |
and predict that right answer.
link |
That's certainly a form of intelligence.
link |
But then when we say, well, do we understand each other?
link |
In other words, would you perceive me as intelligent
link |
beyond that ability to predict?
link |
So now I can predict, but I can't really articulate
link |
how I'm going through that process,
link |
what my underlying theory is for predicting,
link |
and I can't get you to understand what I'm doing
link |
so that you can figure out how to do this yourself
link |
if you did not have, for example,
link |
the right pattern matching machinery that I did.
link |
And now we potentially have this breakdown
link |
where, in effect, I'm intelligent,
link |
but I'm sort of an alien intelligence relative to you.
link |
You're intelligent, but nobody knows about it, or I can't.
link |
Well, I can see the output.
link |
So you're saying, let's sort of separate the two things.
link |
One is you explaining why you were able
link |
to predict the future,
link |
and the second is me being able to,
link |
impressing me that you're intelligent,
link |
me being able to know
link |
that you successfully predicted the future.
link |
Do you think that's?
link |
Well, it's not impressing you that I'm intelligent.
link |
In other words, you may be convinced
link |
that I'm intelligent in some form.
link |
So how, what would convince?
link |
Because of my ability to predict.
link |
So I would look at the metrics.
link |
When you can't, I'd say, wow.
link |
You're right more times than I am.
link |
You're doing something interesting.
link |
That's a form of intelligence.
link |
But then what happens is, if I say, how are you doing that?
link |
And you can't communicate with me,
link |
and you can't describe that to me,
link |
now I may label you a savant.
link |
I may say, well, you're doing something weird,
link |
and it's just not very interesting to me,
link |
because you and I can't really communicate.
link |
And so now, so this is interesting, right?
link |
Because now this is, you're in this weird place
link |
where for you to be recognized
link |
as intelligent the way I'm intelligent,
link |
then you and I sort of have to be able to communicate.
link |
And then my, we start to understand each other,
link |
and then my respect and my appreciation,
link |
my ability to relate to you starts to change.
link |
So now you're not an alien intelligence anymore.
link |
You're a human intelligence now,
link |
because you and I can communicate.
link |
And so I think when we look at animals, for example,
link |
animals can do things we can't quite comprehend,
link |
we don't quite know how they do them,
link |
but they can't really communicate with us.
link |
They can't put what they're going through in our terms.
link |
And so we think of them as sort of,
link |
well, they're these alien intelligences,
link |
and they're not really worth necessarily what we're worth.
link |
We don't treat them the same way as a result of that.
link |
But it's hard because who knows what's going on.
link |
So just a quick elaboration on that,
link |
the explaining that you're intelligent,
link |
the explaining the reasoning that went into the prediction
link |
is not some kind of mathematical proof.
link |
If we look at humans,
link |
look at political debates and discourse on Twitter,
link |
it's mostly just telling stories.
link |
So your task is, sorry,
link |
your task is not to tell an accurate depiction
link |
of how you reason, but to tell a story, real or not,
link |
that convinces me that there was a mechanism by which you.
link |
Ultimately, that's what a proof is.
link |
I mean, even a mathematical proof is that.
link |
Because ultimately, the other mathematicians
link |
have to be convinced by your proof.
link |
Otherwise, in fact, there have been.
link |
That's the metric for success, yeah.
link |
There have been several proofs out there
link |
where mathematicians would study for a long time
link |
before they were convinced
link |
that it actually proved anything, right?
link |
You never know if it proved anything
link |
until the community of mathematicians decided that it did.
link |
So I mean, but it's a real thing, right?
link |
And that's sort of the point, right?
link |
Is that ultimately, this notion of understanding us,
link |
understanding something is ultimately a social concept.
link |
In other words, I have to convince enough people
link |
that I did this in a reasonable way.
link |
I did this in a way that other people can understand
link |
and replicate and that it makes sense to them.
link |
So human intelligence is bound together in that way.
link |
We're bound up in that sense.
link |
We sort of never really get away with it
link |
until we can sort of convince others
link |
that our thinking process makes sense.
link |
Did you think the general question of intelligence
link |
is then also a social construct?
link |
So if we ask questions of an artificial intelligence system,
link |
is this system intelligent?
link |
The answer will ultimately be a socially constructed.
link |
I think, so I think I'm making two statements.
link |
I'm saying we can try to define intelligence
link |
in this super objective way that says, here's this data.
link |
I wanna predict this type of thing, learn this function.
link |
And then if you get it right, often enough,
link |
we consider you intelligent.
link |
But that's more like a sub bond.
link |
It doesn't mean it's not useful.
link |
It could be incredibly useful.
link |
It could be solving a problem we can't otherwise solve
link |
and can solve it more reliably than we can.
link |
But then there's this notion of,
link |
can humans take responsibility
link |
for the decision that you're making?
link |
Can we make those decisions ourselves?
link |
Can we relate to the process that you're going through?
link |
And now you as an agent,
link |
whether you're a machine or another human, frankly,
link |
are now obliged to make me understand
link |
how it is that you're arriving at that answer
link |
and allow me, me or obviously a community
link |
or a judge of people to decide
link |
whether or not that makes sense.
link |
And by the way, that happens with the humans as well.
link |
You're sitting down with your staff, for example,
link |
and you ask for suggestions about what to do next.
link |
And someone says, oh, I think you should buy.
link |
And I actually think you should buy this much
link |
or whatever or sell or whatever it is.
link |
Or I think you should launch the product today or tomorrow
link |
or launch this product versus that product,
link |
whatever the decision may be.
link |
And the person says,
link |
I just have a good feeling about it.
link |
And you're not very satisfied.
link |
Now, that person could be,
link |
you might say, well, you've been right before,
link |
but I'm gonna put the company on the line.
link |
Can you explain to me why I should believe this?
link |
And that explanation may have nothing to do with the truth.
link |
You just, the ultimate.
link |
It's gotta convince the other person.
link |
Still be wrong, still be wrong.
link |
She's gotta be convincing.
link |
But it's ultimately gotta be convincing.
link |
And that's why I'm saying it's,
link |
we're bound together, right?
link |
Our intelligences are bound together in that sense.
link |
We have to understand each other.
link |
And if, for example, you're giving me an explanation,
link |
I mean, this is a very important point, right?
link |
You're giving me an explanation,
link |
and then I'm not good at reasoning well,
link |
and being objective,
link |
and following logical paths and consistent paths,
link |
and I'm not good at measuring
link |
and sort of computing probabilities across those paths.
link |
What happens is collectively,
link |
we're not gonna do well.
link |
How hard is that problem?
link |
So I think we'll talk quite a bit about the first
link |
on a specific objective metric benchmark performing well.
link |
But being able to explain the steps,
link |
the reasoning, how hard is that problem?
link |
I think that's very hard.
link |
I mean, I think that that's,
link |
well, it's hard for humans.
link |
The thing that's hard for humans, as you know,
link |
may not necessarily be hard for computers
link |
So, sorry, so how hard is that problem for computers?
link |
I think it's hard for computers,
link |
and the reason why I related to,
link |
or saying that it's also hard for humans
link |
is because I think when we step back
link |
and we say we wanna design computers to do that,
link |
one of the things we have to recognize
link |
is we're not sure how to do it well.
link |
I'm not sure we have a recipe for that.
link |
And even if you wanted to learn it,
link |
it's not clear exactly what data we use
link |
and what judgments we use to learn that well.
link |
And so what I mean by that is
link |
if you look at the entire enterprise of science,
link |
science is supposed to be at about
link |
objective reason and reason, right?
link |
So we think about, gee, who's the most intelligent person
link |
or group of people in the world?
link |
Do we think about the savants who can close their eyes
link |
and give you a number?
link |
We think about the think tanks,
link |
or the scientists or the philosophers
link |
who kind of work through the details
link |
and write the papers and come up with the thoughtful,
link |
logical proofs and use the scientific method.
link |
I think it's the latter.
link |
And my point is that how do you train someone to do that?
link |
And that's what I mean by it's hard.
link |
How do you, what's the process of training people
link |
That's a hard process.
link |
We work, as a society, we work pretty hard
link |
to get other people to understand our thinking
link |
and to convince them of things.
link |
Now we could persuade them,
link |
obviously you talked about this,
link |
like human flaws or weaknesses,
link |
we can persuade them through emotional means.
link |
But to get them to understand and connect to
link |
and follow a logical argument is difficult.
link |
We try it, we do it, we do it as scientists,
link |
we try to do it as journalists,
link |
we try to do it as even artists in many forms,
link |
as writers, as teachers.
link |
We go through a fairly significant training process
link |
And then we could ask, well, why is that so hard?
link |
And for humans, it takes a lot of work.
link |
And when we step back and say,
link |
well, how do we get a machine to do that?
link |
It's a vexing question.
link |
How would you begin to try to solve that?
link |
And maybe just a quick pause,
link |
because there's an optimistic notion
link |
in the things you're describing,
link |
which is being able to explain something through reason.
link |
But if you look at algorithms that recommend things
link |
that we'll look at next, whether it's Facebook, Google,
link |
advertisement based companies, their goal is to convince you
link |
to buy things based on anything.
link |
So that could be reason,
link |
because the best of advertisement is showing you things
link |
that you really do need and explain why you need it.
link |
But it could also be through emotional manipulation.
link |
The algorithm that describes why a certain decision
link |
was made, how hard is it to do it
link |
through emotional manipulation?
link |
And why is that a good or a bad thing?
link |
So you've kind of focused on reason, logic,
link |
really showing in a clear way why something is good.
link |
One, is that even a thing that us humans do?
link |
And two, how do you think of the difference
link |
in the reasoning aspect and the emotional manipulation?
link |
So you call it emotional manipulation,
link |
but more objectively is essentially saying,
link |
there are certain features of things
link |
that seem to attract your attention.
link |
I mean, it kind of give you more of that stuff.
link |
Manipulation is a bad word.
link |
Yeah, I mean, I'm not saying it's good right or wrong.
link |
It works to get your attention
link |
and it works to get you to buy stuff.
link |
And when you think about algorithms that look
link |
at the patterns of features
link |
that you seem to be spending your money on
link |
and say, I'm gonna give you something
link |
with a similar pattern.
link |
So I'm gonna learn that function
link |
because the objective is to get you to click on it
link |
or get you to buy it or whatever it is.
link |
I don't know, I mean, it is what it is.
link |
I mean, that's what the algorithm does.
link |
You can argue whether it's good or bad.
link |
It depends what your goal is.
link |
I guess this seems to be very useful
link |
for convincing, for telling a story.
link |
For convincing humans, it's good
link |
because again, this goes back to what is the human behavior
link |
like, how does the human brain respond to things?
link |
I think there's a more optimistic view of that too,
link |
which is that if you're searching
link |
for certain kinds of things,
link |
you've already reasoned that you need them.
link |
And these algorithms are saying, look, that's up to you
link |
to reason whether you need something or not.
link |
You may have an unhealthy addiction to this stuff
link |
or you may have a reasoned and thoughtful explanation
link |
for why it's important to you.
link |
And the algorithms are saying, hey, that's like, whatever.
link |
Like, that's your problem.
link |
All I know is you're buying stuff like that.
link |
You're interested in stuff like that.
link |
Could be a bad reason, could be a good reason.
link |
I'm gonna show you more of that stuff.
link |
And I think that it's not good or bad.
link |
It's not reasoned or not reasoned.
link |
The algorithm is doing what it does,
link |
which is saying, you seem to be interested in this.
link |
I'm gonna show you more of that stuff.
link |
And I think we're seeing this not just in buying stuff,
link |
but even in social media.
link |
You're reading this kind of stuff.
link |
I'm not judging on whether it's good or bad.
link |
I'm not reasoning at all.
link |
I'm just saying, I'm gonna show you other stuff
link |
with similar features.
link |
And like, and that's it.
link |
And I wash my hands from it and I say,
link |
that's all that's going on.
link |
You know, there is, people are so harsh on AI systems.
link |
So one, the bar of performance is extremely high.
link |
And yet we also ask them to, in the case of social media,
link |
to help find the better angels of our nature
link |
and help make a better society.
link |
What do you think about the role of AI there?
link |
So that, I agree with you.
link |
That's the interesting dichotomy, right?
link |
Because on one hand, we're sitting there
link |
and we're sort of doing the easy part,
link |
which is finding the patterns.
link |
We're not building, the system's not building a theory
link |
that is consumable and understandable to other humans
link |
that can be explained and justified.
link |
And so on one hand to say, oh, you know, AI is doing this.
link |
Why isn't doing this other thing?
link |
Well, this other thing's a lot harder.
link |
And it's interesting to think about why it's harder.
link |
And because you're interpreting the data
link |
in the context of prior models.
link |
In other words, understandings
link |
of what's important in the world, what's not important.
link |
What are all the other abstract features
link |
that drive our decision making?
link |
What's sensible, what's not sensible,
link |
what's good, what's bad, what's moral,
link |
what's valuable, what isn't?
link |
Where is that stuff?
link |
No one's applying the interpretation.
link |
So when I see you clicking on a bunch of stuff
link |
and I look at these simple features, the raw features,
link |
the features that are there in the data,
link |
like what words are being used or how long the material is
link |
or other very superficial features,
link |
what colors are being used in the material.
link |
Like, I don't know why you're clicking
link |
on this stuff you're clicking.
link |
Or if it's products, what the price is
link |
or what the categories and stuff like that.
link |
And I just feed you more of the same stuff.
link |
That's very different than kind of getting in there
link |
and saying, what does this mean?
link |
The stuff you're reading, like why are you reading it?
link |
What assumptions are you bringing to the table?
link |
Are those assumptions sensible?
link |
Does the material make any sense?
link |
Does it lead you to thoughtful, good conclusions?
link |
Again, there's interpretation and judgment involved
link |
in that process that isn't really happening in the AI today.
link |
That's harder because you have to start getting
link |
at the meaning of the stuff, of the content.
link |
You have to get at how humans interpret the content
link |
relative to their value system
link |
and deeper thought processes.
link |
So that's what meaning means is not just some kind
link |
of deep, timeless, semantic thing
link |
that the statement represents,
link |
but also how a large number of people
link |
are likely to interpret.
link |
So that's again, even meaning is a social construct.
link |
So you have to try to predict how most people
link |
would understand this kind of statement.
link |
Yeah, meaning is often relative,
link |
but meaning implies that the connections go beneath
link |
the surface of the artifacts.
link |
If I show you a painting, it's a bunch of colors on a canvas,
link |
what does it mean to you?
link |
And it may mean different things to different people
link |
because of their different experiences.
link |
It may mean something even different
link |
to the artist who painted it.
link |
As we try to get more rigorous with our communication,
link |
we try to really nail down that meaning.
link |
So we go from abstract art to precise mathematics,
link |
precise engineering drawings and things like that.
link |
We're really trying to say, I wanna narrow
link |
that space of possible interpretations
link |
because the precision of the communication
link |
ends up becoming more and more important.
link |
And so that means that I have to specify,
link |
and I think that's why this becomes really hard,
link |
because if I'm just showing you an artifact
link |
and you're looking at it superficially,
link |
whether it's a bunch of words on a page,
link |
or whether it's brushstrokes on a canvas
link |
or pixels on a photograph,
link |
you can sit there and you can interpret
link |
lots of different ways at many, many different levels.
link |
But when I wanna align our understanding of that,
link |
I have to specify a lot more stuff
link |
that's actually not directly in the artifact.
link |
Now I have to say, well, how are you interpreting
link |
this image and that image?
link |
And what about the colors and what do they mean to you?
link |
What perspective are you bringing to the table?
link |
What are your prior experiences with those artifacts?
link |
What are your fundamental assumptions and values?
link |
What is your ability to kind of reason,
link |
to chain together logical implication
link |
as you're sitting there and saying,
link |
well, if this is the case, then I would conclude this.
link |
And if that's the case, then I would conclude that.
link |
So your reasoning processes and how they work,
link |
your prior models and what they are,
link |
your values and your assumptions,
link |
all those things now come together into the interpretation.
link |
Getting in sync of that is hard.
link |
And yet humans are able to intuit some of that
link |
Because they have the shared experience.
link |
And we're not talking about shared,
link |
two people having shared experience.
link |
I mean, as a society.
link |
We have the shared experience and we have similar brains.
link |
So we tend to, in other words,
link |
part of our shared experiences are shared local experience.
link |
Like we may live in the same culture,
link |
we may live in the same society
link |
and therefore we have similar educations.
link |
We have some of what we like to call prior models
link |
about the word prior experiences.
link |
And we use that as a,
link |
think of it as a wide collection of interrelated variables
link |
and they're all bound to similar things.
link |
And so we take that as our background
link |
and we start interpreting things similarly.
link |
But as humans, we have a lot of shared experience.
link |
We do have similar brains, similar goals,
link |
similar emotions under similar circumstances.
link |
Because we're both humans.
link |
So now one of the early questions you asked,
link |
how is biological and computer information systems
link |
fundamentally different?
link |
Well, one is humans come with a lot of pre programmed stuff.
link |
A ton of program stuff.
link |
And they're able to communicate
link |
because they share that stuff.
link |
Do you think that shared knowledge,
link |
if we can maybe escape the hard work question,
link |
how much is encoded in the hardware?
link |
Just the shared knowledge in the software,
link |
the history, the many centuries of wars and so on
link |
that came to today, that shared knowledge.
link |
How hard is it to encode?
link |
Do you have a hope?
link |
Can you speak to how hard is it to encode that knowledge
link |
systematically in a way that could be used by a computer?
link |
So I think it is possible to learn to,
link |
for a machine to program a machine,
link |
to acquire that knowledge with a similar foundation.
link |
In other words, a similar interpretive foundation
link |
for processing that knowledge.
link |
What do you mean by that?
link |
So in other words, we view the world in a particular way.
link |
So in other words, we have a, if you will,
link |
as humans, we have a framework
link |
for interpreting the world around us.
link |
So we have multiple frameworks for interpreting
link |
the world around us.
link |
But if you're interpreting, for example,
link |
socio political interactions,
link |
you're thinking about where there's people,
link |
there's collections and groups of people,
link |
they have goals, goals largely built around survival
link |
and quality of life.
link |
There are fundamental economics around scarcity of resources.
link |
And when humans come and start interpreting
link |
a situation like that, because you brought up
link |
like historical events,
link |
they start interpreting situations like that.
link |
They apply a lot of this fundamental framework
link |
for interpreting that.
link |
Well, who are the people?
link |
What were their goals?
link |
What resources did they have?
link |
How much power influence did they have over the other?
link |
Like this fundamental substrate, if you will,
link |
for interpreting and reasoning about that.
link |
So I think it is possible to imbue a computer
link |
with that stuff that humans like take for granted
link |
when they go and sit down and try to interpret things.
link |
And then with that foundation, they acquire,
link |
they start acquiring the details,
link |
the specifics in a given situation,
link |
are then able to interpret it with regard to that framework.
link |
And then given that interpretation, they can do what?
link |
But not only can they predict,
link |
they can predict now with an explanation
link |
that can be given in those terms,
link |
in the terms of that underlying framework
link |
that most humans share.
link |
Now you could find humans that come and interpret events
link |
very differently than other humans
link |
because they're like using a different framework.
link |
The movie Matrix comes to mind
link |
where they decided humans were really just batteries,
link |
and that's how they interpreted the value of humans
link |
as a source of electrical energy.
link |
So, but I think that for the most part,
link |
we have a way of interpreting the events
link |
or the social events around us
link |
because we have this shared framework.
link |
It comes from, again, the fact that we're similar beings
link |
that have similar goals, similar emotions,
link |
and we can make sense out of these.
link |
These frameworks make sense to us.
link |
So how much knowledge is there, do you think?
link |
So you said it's possible.
link |
Well, there's a tremendous amount of detailed knowledge
link |
You could imagine effectively infinite number
link |
of unique situations and unique configurations
link |
But the knowledge that you need,
link |
what I refer to as like the frameworks,
link |
for you need for interpreting them, I don't think.
link |
I think those are finite.
link |
You think the frameworks are more important
link |
than the bulk of the knowledge?
link |
So it's like framing.
link |
Yeah, because what the frameworks do
link |
is they give you now the ability to interpret and reason,
link |
and to interpret and reason,
link |
to interpret and reason over the specifics
link |
in ways that other humans would understand.
link |
What about the specifics?
link |
You know, you acquire the specifics by reading
link |
and by talking to other people.
link |
So I'm mostly actually just even,
link |
if we can focus on even the beginning,
link |
the common sense stuff,
link |
the stuff that doesn't even require reading,
link |
or it almost requires playing around with the world
link |
or something, just being able to sort of manipulate objects,
link |
drink water and so on, all of that.
link |
Every time we try to do that kind of thing
link |
in robotics or AI, it seems to be like an onion.
link |
You seem to realize how much knowledge
link |
is really required to perform
link |
even some of these basic tasks.
link |
Do you have that sense as well?
link |
And if so, how do we get all those details?
link |
Are they written down somewhere?
link |
Do they have to be learned through experience?
link |
So I think when, like, if you're talking about
link |
sort of the physics, the basic physics around us,
link |
for example, acquiring information about,
link |
acquiring how that works.
link |
Yeah, I mean, I think there's a combination of things going,
link |
I think there's a combination of things going on.
link |
I think there is like fundamental pattern matching,
link |
like what we were talking about before,
link |
where you see enough examples,
link |
enough data about something and you start assuming that.
link |
And with similar input,
link |
I'm gonna predict similar outputs.
link |
You can't necessarily explain it at all.
link |
You may learn very quickly that when you let something go,
link |
it falls to the ground.
link |
But you can't necessarily explain that.
link |
But that's such a deep idea,
link |
that if you let something go, like the idea of gravity.
link |
I mean, people are letting things go
link |
and counting on them falling
link |
well before they understood gravity.
link |
But that seems to be, that's exactly what I mean,
link |
is before you take a physics class
link |
or study anything about Newton,
link |
just the idea that stuff falls to the ground
link |
and then you'd be able to generalize
link |
that all kinds of stuff falls to the ground.
link |
It just seems like a non, without encoding it,
link |
like hard coding it in,
link |
it seems like a difficult thing to pick up.
link |
It seems like you have to have a lot of different knowledge
link |
to be able to integrate that into the framework,
link |
sort of into everything else.
link |
So both know that stuff falls to the ground
link |
and start to reason about sociopolitical discourse.
link |
So both, like the very basic
link |
and the high level reasoning decision making.
link |
I guess my question is, how hard is this problem?
link |
And sorry to linger on it because again,
link |
and we'll get to it for sure,
link |
as what Watson with Jeopardy did is take on a problem
link |
that's much more constrained
link |
but has the same hugeness of scale,
link |
at least from the outsider's perspective.
link |
So I'm asking the general life question
link |
of to be able to be an intelligent being
link |
and reason in the world about both gravity and politics,
link |
how hard is that problem?
link |
So I think it's solvable.
link |
Okay, now beautiful.
link |
So what about time travel?
link |
Okay, I'm just saying the same answer.
link |
Not as convinced yet, okay.
link |
No, I think it is solvable.
link |
I mean, I think that it's a learn,
link |
first of all, it's about getting machines to learn.
link |
Learning is fundamental.
link |
And I think we're already in a place that we understand,
link |
for example, how machines can learn in various ways.
link |
Right now, our learning stuff is sort of primitive
link |
in that we haven't sort of taught machines
link |
to learn the frameworks.
link |
We don't communicate our frameworks
link |
because of how shared they are, in some cases we do,
link |
but we don't annotate, if you will,
link |
all the data in the world with the frameworks
link |
that are inherent or underlying our understanding.
link |
Instead, we just operate with the data.
link |
So if we wanna be able to reason over the data
link |
in similar terms in the common frameworks,
link |
we need to be able to teach the computer,
link |
or at least we need to program the computer
link |
to acquire, to have access to and acquire,
link |
learn the frameworks as well
link |
and connect the frameworks to the data.
link |
I think this can be done.
link |
I think we can start, I think machine learning,
link |
for example, with enough examples,
link |
can start to learn these basic dynamics.
link |
Will they relate them necessarily to the gravity?
link |
Not unless they can also acquire those theories as well
link |
and put the experiential knowledge
link |
and connect it back to the theoretical knowledge.
link |
I think if we think in terms of these class of architectures
link |
that are designed to both learn the specifics,
link |
find the patterns, but also acquire the frameworks
link |
and connect the data to the frameworks.
link |
If we think in terms of robust architectures like this,
link |
I think there is a path toward getting there.
link |
In terms of encoding architectures like that,
link |
do you think systems that are able to do this
link |
will look like neural networks or representing,
link |
if you look back to the 80s and 90s with the expert systems,
link |
they're more like graphs, systems that are based in logic,
link |
able to contain a large amount of knowledge
link |
where the challenge was the automated acquisition
link |
of that knowledge.
link |
I guess the question is when you collect both the frameworks
link |
and the knowledge from the data,
link |
what do you think that thing will look like?
link |
Yeah, so I mean, I think asking the question,
link |
they look like neural networks is a bit of a red herring.
link |
I mean, I think that they will certainly do inductive
link |
or pattern match based reasoning.
link |
And I've already experimented with architectures
link |
that combine both that use machine learning
link |
and neural networks to learn certain classes of knowledge,
link |
in other words, to find repeated patterns
link |
in order for it to make good inductive guesses,
link |
but then ultimately to try to take those learnings
link |
and marry them, in other words, connect them to frameworks
link |
so that it can then reason over that
link |
in terms other humans understand.
link |
So for example, at elemental cognition, we do both.
link |
We have architectures that do both, both those things,
link |
but also have a learning method
link |
for acquiring the frameworks themselves and saying,
link |
look, ultimately, I need to take this data.
link |
I need to interpret it in the form of these frameworks
link |
so they can reason over it.
link |
So there is a fundamental knowledge representation,
link |
like what you're saying,
link |
like these graphs of logic, if you will.
link |
There are also neural networks
link |
that acquire a certain class of information.
link |
Then they then align them with these frameworks,
link |
but there's also a mechanism
link |
to acquire the frameworks themselves.
link |
Yeah, so it seems like the idea of frameworks
link |
requires some kind of collaboration with humans.
link |
So do you think of that collaboration as direct?
link |
Well, and let's be clear.
link |
Only for the express purpose that you're designing,
link |
you're designing an intelligence
link |
that can ultimately communicate with humans
link |
in the terms of frameworks that help them understand things.
link |
So to be really clear,
link |
you can independently create a machine learning system,
link |
an intelligence that I might call an alien intelligence
link |
that does a better job than you with some things,
link |
but can't explain the framework to you.
link |
That doesn't mean it might be better than you at the thing.
link |
It might be that you cannot comprehend the framework
link |
that it may have created for itself that is inexplicable
link |
But you're more interested in a case where you can.
link |
My sort of approach to AI is because
link |
I've set the goal for myself.
link |
I want machines to be able to ultimately communicate,
link |
understanding with humans.
link |
I want them to be able to acquire and communicate,
link |
acquire knowledge from humans
link |
and communicate knowledge to humans.
link |
They should be using what inductive
link |
machine learning techniques are good at,
link |
which is to observe patterns of data,
link |
whether it be in language or whether it be in images
link |
or videos or whatever,
link |
to acquire these patterns,
link |
to induce the generalizations from those patterns,
link |
but then ultimately to work with humans
link |
to connect them to frameworks, interpretations, if you will,
link |
that ultimately make sense to humans.
link |
Of course, the machine is gonna have the strength
link |
that it has, the richer, longer memory,
link |
but it has the more rigorous reasoning abilities,
link |
the deeper reasoning abilities,
link |
so it'll be an interesting complementary relationship
link |
between the human and the machine.
link |
Do you think that ultimately needs explainability
link |
So if we look, we study, for example,
link |
Tesla autopilot a lot, where humans,
link |
I don't know if you've driven the vehicle,
link |
are aware of what it is.
link |
So you're basically the human and machine
link |
are working together there,
link |
and the human is responsible for their own life
link |
to monitor the system,
link |
and the system fails every few miles,
link |
and so there's hundreds,
link |
there's millions of those failures a day,
link |
and so that's like a moment of interaction.
link |
Yeah, that's exactly right.
link |
That's a moment of interaction
link |
where the machine has learned some stuff,
link |
it has a failure, somehow the failure's communicated,
link |
the human is now filling in the mistake, if you will,
link |
or maybe correcting or doing something
link |
that is more successful in that case,
link |
the computer takes that learning.
link |
So I believe that the collaboration
link |
between human and machine,
link |
I mean, that's sort of a primitive example
link |
and sort of a more,
link |
another example is where the machine's literally talking
link |
to you and saying, look, I'm reading this thing.
link |
I know that the next word might be this or that,
link |
but I don't really understand why.
link |
Can you help me understand the framework that supports this
link |
and then can kind of acquire that,
link |
take that and reason about it and reuse it
link |
the next time it's reading to try to understand something,
link |
not unlike a human student might do.
link |
I mean, I remember when my daughter was in first grade
link |
and she had a reading assignment about electricity
link |
and somewhere in the text it says,
link |
and electricity is produced by water flowing over turbines
link |
or something like that.
link |
And then there's a question that says,
link |
well, how is electricity created?
link |
And so my daughter comes to me and says,
link |
I mean, I could, you know,
link |
created and produced are kind of synonyms in this case.
link |
So I can go back to the text
link |
and I can copy by water flowing over turbines,
link |
but I have no idea what that means.
link |
Like I don't know how to interpret
link |
water flowing over turbines and what electricity even is.
link |
I mean, I can get the answer right by matching the text,
link |
but I don't have any framework for understanding
link |
what this means at all.
link |
And framework really is, I mean, it's a set of,
link |
not to be mathematical, but axioms of ideas
link |
that you bring to the table and interpreting stuff
link |
and then you build those up somehow.
link |
You build them up with the expectation
link |
that there's a shared understanding of what they are.
link |
Sure, yeah, it's the social, that us humans,
link |
do you have a sense that humans on earth in general
link |
share a set of, like how many frameworks are there?
link |
I mean, it depends on how you bound them, right?
link |
So in other words, how big or small,
link |
like their individual scope,
link |
but there's lots and there are new ones.
link |
I think the way I think about it is kind of in a layer.
link |
I think that the architectures are being layered in that.
link |
There's a small set of primitives.
link |
They allow you the foundation to build frameworks.
link |
And then there may be many frameworks,
link |
but you have the ability to acquire them.
link |
And then you have the ability to reuse them.
link |
I mean, one of the most compelling ways
link |
of thinking about this is a reasoning by analogy,
link |
where I can say, oh, wow,
link |
I've learned something very similar.
link |
I never heard of this game soccer,
link |
but if it's like basketball in the sense
link |
that the goal's like the hoop
link |
and I have to get the ball in the hoop
link |
and I have guards and I have this and I have that,
link |
like where are the similarities
link |
and where are the differences?
link |
And I have a foundation now
link |
for interpreting this new information.
link |
And then the different groups,
link |
like the millennials will have a framework.
link |
And then, you know, the Democrats and Republicans.
link |
Millennials, nobody wants that framework.
link |
Well, I mean, I think, right,
link |
I mean, you're talking about political and social ways
link |
of interpreting the world around them.
link |
And I think these frameworks are still largely,
link |
I think they differ in maybe
link |
what some fundamental assumptions and values are.
link |
Now, from a reasoning perspective,
link |
like the ability to process the framework,
link |
it might not be that different.
link |
The implications of different fundamental values
link |
or fundamental assumptions in those frameworks
link |
may reach very different conclusions.
link |
So from a social perspective,
link |
the conclusions may be very different.
link |
From an intelligence perspective,
link |
I just followed where my assumptions took me.
link |
Yeah, the process itself will look similar.
link |
But that's a fascinating idea
link |
that frameworks really help carve
link |
how a statement will be interpreted.
link |
I mean, having a Democrat and a Republican framework
link |
and then read the exact same statement
link |
and the conclusions that you derive
link |
will be totally different
link |
from an AI perspective is fascinating.
link |
What we would want out of the AI
link |
is to be able to tell you
link |
that this perspective, one perspective,
link |
one set of assumptions is gonna lead you here,
link |
another set of assumptions is gonna lead you there.
link |
And in fact, to help people reason and say,
link |
oh, I see where our differences lie.
link |
I have this fundamental belief about that.
link |
I have this fundamental belief about that.
link |
Yeah, that's quite brilliant.
link |
From my perspective, NLP,
link |
there's this idea that there's one way
link |
to really understand a statement,
link |
but that probably isn't.
link |
There's probably an infinite number of ways
link |
to understand a statement, depending on the question.
link |
There's lots of different interpretations,
link |
and the broader the content, the richer it is.
link |
And so you and I can have very different experiences
link |
with the same text, obviously.
link |
And if we're committed to understanding each other,
link |
we start, and that's the other important point,
link |
if we're committed to understanding each other,
link |
we start decomposing and breaking down our interpretation
link |
to its more and more primitive components
link |
until we get to that point where we say,
link |
oh, I see why we disagree.
link |
And we try to understand how fundamental
link |
that disagreement really is.
link |
But that requires a commitment
link |
to breaking down that interpretation
link |
in terms of that framework in a logical way.
link |
Otherwise, and this is why I think of AI
link |
as really complimenting and helping human intelligence
link |
to overcome some of its biases and its predisposition
link |
to be persuaded by more shallow reasoning
link |
in the sense that we get over this idea,
link |
well, I'm right because I'm Republican,
link |
or I'm right because I'm Democratic,
link |
and someone labeled this as Democratic point of view,
link |
or it has the following keywords in it.
link |
And if the machine can help us break that argument down
link |
and say, wait a second, what do you really think
link |
about this, right?
link |
So essentially holding us accountable
link |
to doing more critical thinking.
link |
We're gonna have to sit and think about this fast.
link |
That's, I love that.
link |
I think that's really empowering use of AI
link |
for the public discourse is completely disintegrating
link |
currently as we learn how to do it on social media.
link |
So one of the greatest accomplishments
link |
in the history of AI is Watson competing
link |
in the game of Jeopardy against humans.
link |
And you were a lead in that, a critical part of that.
link |
Let's start at the very basics.
link |
What is the game of Jeopardy?
link |
The game for us humans, human versus human.
link |
Right, so it's to take a question and answer it.
link |
The game of Jeopardy.
link |
It's just the opposite.
link |
Actually, well, no, but it's not, right?
link |
It's really to get a question and answer,
link |
but it's what we call a factoid question.
link |
So this notion of like, it really relates to some fact
link |
that two people would argue
link |
whether the facts are true or not.
link |
In fact, most people wouldn't.
link |
Jeopardy kind of counts on the idea
link |
that these statements have factual answers.
link |
And the idea is to, first of all,
link |
determine whether or not you know the answer,
link |
which is sort of an interesting twist.
link |
So first of all, understand the question.
link |
You have to understand the question.
link |
What is it asking?
link |
And that's a good point
link |
because the questions are not asked directly, right?
link |
the way the questions are asked is nonlinear.
link |
It's like, it's a little bit witty.
link |
It's a little bit playful sometimes.
link |
It's a little bit tricky.
link |
Yeah, they're asked in exactly numerous witty, tricky ways.
link |
Exactly what they're asking is not obvious.
link |
It takes inexperienced humans a while
link |
to go, what is it even asking?
link |
And it's sort of an interesting realization that you have
link |
when somebody says, oh, what's,
link |
Jeopardy is a question answering show.
link |
And then he's like, oh, like, I know a lot.
link |
And then you read it and you're still trying
link |
to process the question and the champions have answered
link |
There are three questions ahead
link |
by the time you figured out what the question even meant.
link |
So there's definitely an ability there
link |
to just parse out what the question even is.
link |
So that was certainly challenging.
link |
It's interesting historically though,
link |
if you look back at the Jeopardy games much earlier,
link |
you know, early games. Like 60s, 70s, that kind of thing.
link |
The questions were much more direct.
link |
They weren't quite like that.
link |
They got sort of more and more interesting,
link |
the way they asked them that sort of got more
link |
and more interesting and subtle and nuanced
link |
and humorous and witty over time,
link |
which really required the human
link |
to kind of make the right connections
link |
in figuring out what the question was even asking.
link |
So yeah, you have to figure out the questions even asking.
link |
Then you have to determine whether
link |
or not you think you know the answer.
link |
And because you have to buzz in really quickly,
link |
you sort of have to make that determination
link |
as quickly as you possibly can.
link |
Otherwise you lose the opportunity to buzz in.
link |
Even before you really know if you know the answer.
link |
I think a lot of humans will assume,
link |
they'll process it very superficially.
link |
In other words, what's the topic?
link |
What are some keywords?
link |
And just say, do I know this area or not
link |
before they actually know the answer?
link |
Then they'll buzz in and think about it.
link |
So it's interesting what humans do.
link |
Now, some people who know all things,
link |
like Ken Jennings or something,
link |
or the more recent big Jeopardy player,
link |
I mean, they'll just buzz in.
link |
They'll just assume they know all of Jeopardy
link |
and they'll just buzz in.
link |
Watson, interestingly, didn't even come close
link |
to knowing all of Jeopardy, right?
link |
Even at the peak, even at its best.
link |
Yeah, so for example, I mean,
link |
we had this thing called recall,
link |
which is like how many of all the Jeopardy questions,
link |
how many could we even find the right answer for anywhere?
link |
Like, can we come up with, we had a big body of knowledge,
link |
something in the order of several terabytes.
link |
I mean, from a web scale, it was actually very small,
link |
but from like a book scale,
link |
we're talking about millions of books, right?
link |
So the equivalent of millions of books,
link |
encyclopedias, dictionaries, books,
link |
it's still a ton of information.
link |
And I think it was only 85% was the answer
link |
anywhere to be found.
link |
So you're already down at that level
link |
just to get started, right?
link |
So, and so it was important to get a very quick sense
link |
of do you think you know the right answer to this question?
link |
So we had to compute that confidence
link |
as quickly as we possibly could.
link |
So in effect, we had to answer it
link |
and at least spend some time essentially answering it
link |
and then judging the confidence that our answer was right
link |
and then deciding whether or not
link |
we were confident enough to buzz in.
link |
And that would depend on what else was going on in the game.
link |
Because there was a risk.
link |
So like if you're really in a situation
link |
where I have to take a guess, I have very little to lose,
link |
then you'll buzz in with less confidence.
link |
So that was accounted for the financial standings
link |
of the different competitors.
link |
How much of the game was left?
link |
How much time was left?
link |
Where you were in the standing, things like that.
link |
How many hundreds of milliseconds
link |
that we're talking about here?
link |
Do you have a sense of what is?
link |
We targeted, yeah, we targeted.
link |
So, I mean, we targeted answering
link |
in under three seconds and.
link |
So the decision to buzz in and then the actual answering
link |
are those two different stages?
link |
Yeah, they were two different things.
link |
In fact, we had multiple stages,
link |
whereas like we would say, let's estimate our confidence,
link |
which was sort of a shallow answering process.
link |
And then ultimately decide to buzz in
link |
and then we may take another second or something
link |
to kind of go in there and do that.
link |
But by and large, we were saying like,
link |
we can't play the game.
link |
We can't even compete if we can't on average
link |
answer these questions in around three seconds or less.
link |
So you stepped in.
link |
So there's these three humans playing a game
link |
and you stepped in with the idea that IBM Watson
link |
would be one of, replace one of the humans
link |
and compete against two.
link |
Can you tell the story of Watson taking on this game?
link |
It seems exceptionally difficult.
link |
Yeah, so the story was that it was coming up,
link |
I think to the 10 year anniversary of Big Blue,
link |
not Big Blue, Deep Blue.
link |
IBM wanted to do sort of another kind of really
link |
fun challenge, public challenge that can bring attention
link |
to IBM research and the kind of the cool stuff
link |
that we were doing.
link |
I had been working in AI at IBM for some time.
link |
I had a team doing what's called
link |
open domain factoid question answering,
link |
which is, we're not gonna tell you what the questions are.
link |
We're not even gonna tell you what they're about.
link |
Can you go off and get accurate answers to these questions?
link |
And it was an area of AI research that I was involved in.
link |
And so it was a very specific passion of mine.
link |
Language understanding had always been a passion of mine.
link |
One sort of narrow slice on whether or not
link |
you could do anything with language
link |
was this notion of open domain and meaning
link |
I could ask anything about anything.
link |
Factoid meaning it essentially had an answer
link |
and being able to do that accurately and quickly.
link |
So that was a research area
link |
that my team had already been in.
link |
And so completely independently,
link |
several IBM executives, like what are we gonna do?
link |
What's the next cool thing to do?
link |
And Ken Jennings was on his winning streak.
link |
This was like, whatever it was, 2004, I think,
link |
was on his winning streak.
link |
And someone thought, hey, that would be really cool
link |
if the computer can play Jeopardy.
link |
And so this was like in 2004,
link |
they were shopping this thing around
link |
and everyone was telling the research execs, no way.
link |
Like, this is crazy.
link |
And we had some pretty senior people in the field
link |
and they're saying, no, this is crazy.
link |
And it would come across my desk and I was like,
link |
but that's kind of what I'm really interested in doing.
link |
But there was such this prevailing sense of this is nuts.
link |
We're not gonna risk IBM's reputation on this.
link |
We're just not doing it.
link |
And this happened in 2004, it happened in 2005.
link |
At the end of 2006, it was coming around again.
link |
And I was coming off of a,
link |
I was doing the open domain question answering stuff,
link |
but I was coming off a couple other projects.
link |
I had a lot more time to put into this.
link |
And I argued that it could be done.
link |
And I argue it would be crazy not to do this.
link |
Can I, you can be honest at this point.
link |
So even though you argued for it,
link |
what's the confidence that you had yourself privately
link |
that this could be done?
link |
Was, we just told the story,
link |
how you tell stories to convince others.
link |
How confident were you?
link |
What was your estimation of the problem at that time?
link |
So I thought it was possible.
link |
And a lot of people thought it was impossible.
link |
I thought it was possible.
link |
The reason why I thought it was possible
link |
was because I did some brief experimentation.
link |
I knew a lot about how we were approaching
link |
open domain factoid question answering.
link |
I've been doing it for some years.
link |
I looked at the Jeopardy stuff.
link |
I said, this is gonna be hard
link |
for a lot of the points that we mentioned earlier.
link |
Hard to interpret the question.
link |
Hard to do it quickly enough.
link |
Hard to compute an accurate confidence.
link |
None of this stuff had been done well enough before.
link |
But a lot of the technologies we're building
link |
were the kinds of technologies that should work.
link |
But more to the point, what was driving me was,
link |
I was in IBM research.
link |
I was a senior leader in IBM research.
link |
And this is the kind of stuff we were supposed to do.
link |
In other words, we were basically supposed to.
link |
This is the moonshot.
link |
We were supposed to take things and say,
link |
this is an active research area.
link |
It's our obligation to kind of,
link |
if we have the opportunity, to push it to the limits.
link |
And if it doesn't work,
link |
to understand more deeply why we can't do it.
link |
And so I was very committed to that notion saying,
link |
folks, this is what we do.
link |
It's crazy not to do this.
link |
This is an active research area.
link |
We've been in this for years.
link |
Why wouldn't we take this grand challenge
link |
and push it as hard as we can?
link |
At the very least, we'd be able to come out and say,
link |
here's why this problem is way hard.
link |
Here's what we tried and here's how we failed.
link |
So I was very driven as a scientist from that perspective.
link |
And then I also argued,
link |
based on what we did a feasibility study,
link |
why I thought it was hard but possible.
link |
And I showed examples of where it succeeded,
link |
where it failed, why it failed,
link |
and sort of a high level architecture approach
link |
for why we should do it.
link |
But for the most part, at that point,
link |
the execs really were just looking for someone crazy enough
link |
to say yes, because for several years at that point,
link |
everyone had said, no, I'm not willing to risk my reputation
link |
and my career on this thing.
link |
Clearly you did not have such fears.
link |
So you dived right in.
link |
And yet, for what I understand,
link |
it was performing very poorly in the beginning.
link |
So what were the initial approaches and why did they fail?
link |
Well, there were lots of hard aspects to it.
link |
I mean, one of the reasons why prior approaches
link |
that we had worked on in the past failed
link |
was because the questions were difficult to interpret.
link |
Like, what are you even asking for, right?
link |
Very often, like if the question was very direct,
link |
like what city, or what, even then it could be tricky,
link |
but what city or what person,
link |
is often when it would name it very clearly,
link |
you would know that.
link |
And if there were just a small set of them,
link |
in other words, we're gonna ask about these five types.
link |
Like, it's gonna be an answer,
link |
and the answer will be a city in this state
link |
or a city in this country.
link |
The answer will be a person of this type, right?
link |
Like an actor or whatever it is.
link |
But it turns out that in Jeopardy,
link |
there were like tens of thousands of these things.
link |
And it was a very, very long tail,
link |
meaning that it just went on and on.
link |
And so even if you focused on trying to encode the types
link |
at the very top, like there's five that were the most,
link |
let's say five of the most frequent,
link |
you still cover a very small percentage of the data.
link |
So you couldn't take that approach of saying,
link |
I'm just going to try to collect facts
link |
about these five or 10 types or 20 types
link |
or 50 types or whatever.
link |
So that was like one of the first things,
link |
like what do you do about that?
link |
And so we came up with an approach toward that.
link |
And the approach looked promising,
link |
and we continued to improve our ability
link |
to handle that problem throughout the project.
link |
The other issue was that right from the outside,
link |
I said, we're not going to,
link |
I committed to doing this in three to five years.
link |
So we did it in four.
link |
But one of the things that that,
link |
putting that like stake in the ground was,
link |
and I knew how hard the language understanding problem was.
link |
I said, we're not going to actually understand language
link |
to solve this problem.
link |
We are not going to interpret the question
link |
and the domain of knowledge that the question refers to
link |
and reason over that to answer these questions.
link |
Obviously we're not going to be doing that.
link |
simple search wasn't good enough to confidently answer
link |
with a single correct answer.
link |
First of all, that's like brilliant.
link |
That's such a great mix of innovation
link |
and practical engineering three, four, eight.
link |
So you're not trying to solve the general NLU problem.
link |
You're saying, let's solve this in any way possible.
link |
No, I was committed to saying, look,
link |
we're going to solving the open domain
link |
question answering problem.
link |
We're using Jeopardy as a driver for that.
link |
That's a big benchmark.
link |
Good enough, big benchmark, exactly.
link |
We could just like, whatever,
link |
like just figure out what works
link |
because I want to be able to go back
link |
to the academic science community
link |
and say, here's what we tried.
link |
Here's what worked.
link |
Here's what didn't work.
link |
I don't want to go in and say,
link |
oh, I only have one technology.
link |
I'm only going to use this.
link |
I'm going to do whatever it takes.
link |
I'm like, I'm going to think out of the box
link |
and do whatever it takes.
link |
One, and I also, there was another thing I believed.
link |
I believed that the fundamental NLP technologies
link |
and machine learning technologies would be adequate.
link |
And this was an issue of how do we enhance them?
link |
How do we integrate them?
link |
How do we advance them?
link |
So I had one researcher who came to me
link |
who had been working on question answering
link |
with me for a very long time,
link |
who had said, we're going to need Maxwell's equations
link |
for question answering.
link |
And I said, if we need some fundamental formula
link |
that breaks new ground in how we understand language,
link |
We're not going to get there from here.
link |
Like I am not counting.
link |
My assumption is I'm not counting
link |
on some brand new invention.
link |
What I'm counting on is the ability
link |
to take everything it has done before
link |
to figure out an architecture on how to integrate it well
link |
and then see where it breaks
link |
and make the necessary advances we need to make
link |
until this thing works.
link |
Push it hard to see where it breaks
link |
and then patch it up.
link |
I mean, that's how people change the world.
link |
I mean, that's the Elon Musk approach to the rockets,
link |
SpaceX, that's the Henry Ford and so on.
link |
And I happen to be, in this case, I happen to be right,
link |
but like we didn't know.
link |
But you kind of have to put a stake in the rest
link |
of how you're going to run the project.
link |
So yeah, and backtracking to search.
link |
So if you were to do, what's the brute force solution?
link |
What would you search over?
link |
So you have a question,
link |
how would you search the possible space of answers?
link |
Look, web search has come a long way even since then.
link |
But at the time, first of all,
link |
I mean, there were a couple of other constraints
link |
around the problem, which is interesting.
link |
So you couldn't go out to the web.
link |
You couldn't search the internet.
link |
In other words, the AI experiment was,
link |
we want a self contained device.
link |
If the device is as big as a room, fine,
link |
it's as big as a room,
link |
but we want a self contained device.
link |
You're not going out to the internet.
link |
You don't have a lifeline to anything.
link |
So it had to kind of fit in a shoe box, if you will,
link |
or at least a size of a few refrigerators,
link |
whatever it might be.
link |
See, but also you couldn't just get out there.
link |
You couldn't go off network, right, to kind of go.
link |
So there was that limitation.
link |
But then we did, but the basic thing was go do web search.
link |
Problem was, even when we went and did a web search,
link |
I don't remember exactly the numbers,
link |
but somewhere in the order of 65% of the time,
link |
the answer would be somewhere, you know,
link |
in the top 10 or 20 documents.
link |
So first of all, that's not even good enough to play Jeopardy.
link |
You know, the words, even if you could pull the,
link |
even if you could perfectly pull the answer
link |
out of the top 20 documents, top 10 documents,
link |
whatever it was, which we didn't know how to do.
link |
But even if you could do that, you'd be,
link |
and you knew it was right,
link |
unless you had enough confidence in it, right?
link |
So you'd have to pull out the right answer.
link |
You'd have to have confidence it was the right answer.
link |
And then you'd have to do that fast enough to now go buzz in
link |
and you'd still only get 65% of them right,
link |
which doesn't even put you in the winner's circle.
link |
Winner's circle, you have to be up over 70
link |
and you have to do it really quick
link |
and you have to do it really quickly.
link |
But now the problem is, well,
link |
even if I had somewhere in the top 10 documents,
link |
how do I figure out where in the top 10 documents
link |
that answer is and how do I compute a confidence
link |
of all the possible candidates?
link |
So it's not like I go in knowing the right answer
link |
and I have to pick it.
link |
I don't know the right answer.
link |
I have a bunch of documents,
link |
somewhere in there is the right answer.
link |
How do I, as a machine, go out
link |
and figure out which one's right?
link |
And then how do I score it?
link |
So, and now how do I deal with the fact
link |
that I can't actually go out to the web?
link |
First of all, if you pause on that, just think about it.
link |
If you could go to the web,
link |
do you think that problem is solvable
link |
if you just pause on it?
link |
Just thinking even beyond jeopardy,
link |
do you think the problem of reading text
link |
defined where the answer is?
link |
Well, we solved that in some definition of solves
link |
given the jeopardy challenge.
link |
How did you do it for jeopardy?
link |
So how do you take a body of work in a particular topic
link |
and extract the key pieces of information?
link |
So now forgetting about the huge volumes
link |
that are on the web, right?
link |
So now we have to figure out,
link |
we did a lot of source research.
link |
In other words, what body of knowledge
link |
is gonna be small enough,
link |
but broad enough to answer jeopardy?
link |
And we ultimately did find the body of knowledge
link |
I mean, it included Wikipedia and a bunch of other stuff.
link |
So like encyclopedia type of stuff.
link |
I don't know if you can speak to it.
link |
Encyclopedias, dictionaries,
link |
different types of semantic resources,
link |
like WordNet and other types of semantic resources like that,
link |
as well as like some web crawls.
link |
In other words, where we went out and took that content
link |
and then expanded it based on producing,
link |
statistically producing seeds,
link |
using those seeds for other searches and then expanding that.
link |
So using these like expansion techniques,
link |
we went out and had found enough content
link |
and we're like, okay, this is good.
link |
And even up until the end,
link |
we had a thread of research.
link |
It was always trying to figure out
link |
what content could we efficiently include.
link |
I mean, there's a lot of popular,
link |
like what is the church lady?
link |
Well, I think was one of the, like what,
link |
where do you, I guess that's probably an encyclopedia, so.
link |
So that was an encyclopedia,
link |
but then we would take that stuff
link |
and we would go out and we would expand.
link |
In other words, we'd go find other content
link |
that wasn't in the core resources and expand it.
link |
The amount of content, we grew it by an order of magnitude,
link |
but still, again, from a web scale perspective,
link |
this is very small amount of content.
link |
We then took all that content,
link |
we preanalyzed the crap out of it,
link |
meaning we parsed it,
link |
broke it down into all those individual words
link |
and then we did semantic,
link |
syntactic and semantic parses on it,
link |
had computer algorithms that annotated it
link |
and we indexed that in a very rich and very fast index.
link |
So we have a relatively huge amount of,
link |
let's say the equivalent of, for the sake of argument,
link |
two to 5 million bucks.
link |
We've now analyzed all that, blowing up its size even more
link |
because now we have all this metadata
link |
and then we richly indexed all of that
link |
and by the way, in a giant in memory cache.
link |
So Watson did not go to disk.
link |
So the infrastructure component there,
link |
if you could just speak to it, how tough it,
link |
I mean, I know 2000, maybe this is 2008, nine,
link |
that's kind of a long time ago.
link |
How hard is it to use multiple machines?
link |
How hard is the infrastructure component,
link |
the hardware component?
link |
So we used IBM hardware.
link |
We had something like, I forgot exactly,
link |
but close to 3000 cores completely connected.
link |
So you had a switch where every CPU
link |
was connected to every other CPU.
link |
And they were sharing memory in some kind of way.
link |
Large shared memory, right?
link |
And all this data was preanalyzed
link |
and put into a very fast indexing structure
link |
that was all in memory.
link |
And then we took that question,
link |
we would analyze the question.
link |
So all the content was now preanalyzed.
link |
So if I went and tried to find a piece of content,
link |
it would come back with all the metadata
link |
that we had precomputed.
link |
How do you shove that question?
link |
How do you connect the big knowledge base
link |
with the metadata and that's indexed
link |
to the simple little witty confusing question?
link |
So therein lies the Watson architecture, right?
link |
So we would take the question,
link |
we would analyze the question.
link |
So which means that we would parse it
link |
and interpret it a bunch of different ways.
link |
We'd try to figure out what is it asking about?
link |
So we had multiple strategies
link |
to kind of determine what was it asking for.
link |
That might be represented as a simple string,
link |
a character string,
link |
or something we would connect back
link |
to different semantic types
link |
that were from existing resources.
link |
So anyway, the bottom line is
link |
we would do a bunch of analysis in the question.
link |
And question analysis had to finish and had to finish fast.
link |
So we do the question analysis
link |
because then from the question analysis,
link |
we would now produce searches.
link |
So we would, and we had built
link |
using open source search engines,
link |
but we had a number of different search engines
link |
we would use that had different characteristics.
link |
We went in there and engineered
link |
and modified those search engines,
link |
ultimately to now take our question analysis,
link |
produce multiple queries
link |
based on different interpretations of the question
link |
and fire out a whole bunch of searches in parallel.
link |
And they would come back with passages.
link |
So these are passive search algorithms.
link |
They would come back with passages.
link |
And so now let's say you had a thousand passages.
link |
Now for each passage, you parallelize again.
link |
So you went out and you parallelize the search.
link |
Each search would now come back
link |
with a whole bunch of passages.
link |
Maybe you had a total of a thousand
link |
or 5,000 whatever passages.
link |
For each passage now,
link |
you'd go and figure out whether or not
link |
there was a candidate,
link |
we'd call it candidate answer in there.
link |
So you had a whole bunch of other algorithms
link |
that would find candidate answers,
link |
possible answers to the question.
link |
And so you had candidate answer,
link |
called candidate answer generators,
link |
a whole bunch of those.
link |
So for every one of these components,
link |
the team was constantly doing research coming up,
link |
better ways to generate search queries from the questions,
link |
better ways to analyze the question,
link |
better ways to generate candidates.
link |
And speed, so better is accuracy and speed.
link |
Correct, so right, speed and accuracy
link |
for the most part were separated.
link |
We handle that sort of in separate ways.
link |
Like I focus purely on accuracy, end to end accuracy.
link |
Are we ultimately getting more questions
link |
and producing more accurate confidences?
link |
And then a whole nother team
link |
that was constantly analyzing the workflow
link |
to find the bottlenecks.
link |
And then figuring out how to both parallelize
link |
and drive the algorithm speed.
link |
But anyway, so now think of it like,
link |
you have this big fan out now, right?
link |
Because you had multiple queries,
link |
now you have thousands of candidate answers.
link |
For each candidate answer, you're gonna score it.
link |
So you're gonna use all the data that built up.
link |
You're gonna use the question analysis,
link |
you're gonna use how the query was generated,
link |
you're gonna use the passage itself,
link |
and you're gonna use the candidate answer
link |
that was generated, and you're gonna score that.
link |
So now we have a group of researchers
link |
coming up with scores.
link |
There are hundreds of different scores.
link |
So now you're getting a fan out of it again
link |
from however many candidate answers you have
link |
to all the different scores.
link |
So if you have 200 different scores
link |
and you have a thousand candidates,
link |
now you have 200,000 scores.
link |
And so now you gotta figure out,
link |
how do I now rank these answers
link |
based on the scores that came back?
link |
And I wanna rank them based on the likelihood
link |
that they're a correct answer to the question.
link |
So every scorer was its own research project.
link |
What do you mean by scorer?
link |
So is that the annotation process
link |
of basically a human being saying that this answer
link |
Think of it, if you wanna think of it,
link |
what you're doing, you know,
link |
if you wanna think about what a human would be doing,
link |
human would be looking at a possible answer,
link |
they'd be reading the, you know, Emily Dickinson,
link |
they'd be reading the passage in which that occurred,
link |
they'd be looking at the question,
link |
and they'd be making a decision of how likely it is
link |
that Emily Dickinson, given this evidence in this passage,
link |
is the right answer to that question.
link |
So that's the annotation task.
link |
That's the annotation process.
link |
That's the scoring task.
link |
But scoring implies zero to one kind of continuous.
link |
You give it a zero to one score.
link |
So it's not a binary.
link |
No, you give it a score.
link |
Give it a zero to, yeah, exactly, zero to one score.
link |
But humans give different scores,
link |
so you have to somehow normalize and all that kind of stuff
link |
that deal with all that complexity.
link |
It depends on what your strategy is.
link |
It could be relative, too.
link |
We actually looked at the raw scores
link |
as well as standardized scores,
link |
because humans are not involved in this.
link |
Humans are not involved.
link |
Sorry, so I'm misunderstanding the process here.
link |
Where is the ground truth coming from?
link |
Ground truth is only the answers to the questions.
link |
So it's end to end.
link |
So I was always driving end to end performance.
link |
It's a very interesting, a very interesting
link |
engineering approach,
link |
and ultimately scientific research approach,
link |
always driving end to end.
link |
Now, that's not to say
link |
we wouldn't make hypotheses
link |
that individual component performance
link |
was related in some way to end to end performance.
link |
Of course we would,
link |
because people would have to build individual components.
link |
But ultimately, to get your component integrated
link |
to the system, you have to show impact
link |
on end to end performance, question answering performance.
link |
So there's many very smart people working on this,
link |
and they're basically trying to sell their ideas
link |
as a component that should be part of the system.
link |
And they would do research on their component,
link |
and they would say things like,
link |
I'm gonna improve this as a candidate generator,
link |
or I'm gonna improve this as a question score,
link |
or as a passive scorer,
link |
I'm gonna improve this, or as a parser,
link |
and I can improve it by 2% on its component metric,
link |
like a better parse, or a better candidate,
link |
or a better type estimation, whatever it is.
link |
And then I would say,
link |
I need to understand how the improvement
link |
on that component metric
link |
is gonna affect the end to end performance.
link |
If you can't estimate that,
link |
and can't do experiments to demonstrate that,
link |
it doesn't get in.
link |
That's like the best run AI project I've ever heard.
link |
Okay, what breakthrough would you say,
link |
like, I'm sure there's a lot of day to day breakthroughs,
link |
but was there like a breakthrough
link |
that really helped improve performance?
link |
Like where people began to believe,
link |
or is it just a gradual process?
link |
Well, I think it was a gradual process,
link |
but one of the things that I think gave people confidence
link |
that we can get there was that,
link |
as we follow this procedure of different ideas,
link |
build different components,
link |
plug them into the architecture,
link |
run the system, see how we do,
link |
do the error analysis,
link |
start off new research projects to improve things.
link |
And the very important idea
link |
that the individual component work
link |
did not have to deeply understand everything
link |
that was going on with every other component.
link |
And this is where we leverage machine learning
link |
in a very important way.
link |
So while individual components
link |
could be statistically driven machine learning components,
link |
some of them were heuristic,
link |
some of them were machine learning components,
link |
the system has a whole combined all the scores
link |
using machine learning.
link |
This was critical because that way
link |
you can divide and conquer.
link |
So you can say, okay, you work on your candidate generator,
link |
or you work on this approach to answer scoring,
link |
you work on this approach to type scoring,
link |
you work on this approach to passage search
link |
or to pass a selection and so forth.
link |
But when we just plug it in,
link |
and we had enough training data to say,
link |
now we can train and figure out
link |
how do we weigh all the scores relative to each other
link |
based on the predicting the outcome,
link |
which is right or wrong on Jeopardy.
link |
And we had enough training data to do that.
link |
So this enabled people to work independently
link |
and to let the machine learning do the integration.
link |
Beautiful, so yeah, the machine learning
link |
is doing the fusion,
link |
and then it's a human orchestrated ensemble
link |
of different approaches.
link |
Still impressive that you were able
link |
to get it done in a few years.
link |
That's not obvious to me that it's doable,
link |
if I just put myself in that mindset.
link |
But when you look back at the Jeopardy challenge,
link |
again, when you're looking up at the stars,
link |
what are you most proud of, looking back at those days?
link |
I'm most proud of my,
link |
my commitment and my team's commitment
link |
to be true to the science,
link |
to not be afraid to fail.
link |
That's beautiful because there's so much pressure,
link |
because it is a public event, it is a public show,
link |
that you were dedicated to the idea.
link |
Do you think it was a success?
link |
In the eyes of the world, it was a success.
link |
By your, I'm sure, exceptionally high standards,
link |
is there something you regret you would do differently?
link |
It was a success for our goal.
link |
Our goal was to build the most advanced
link |
open domain question answering system.
link |
We went back to the old problems
link |
that we used to try to solve,
link |
and we did dramatically better on all of them,
link |
as well as we beat Jeopardy.
link |
So we won at Jeopardy.
link |
So it was a success.
link |
It was, I worry that the community
link |
or the world would not understand it as a success
link |
because it came down to only one game.
link |
And I knew statistically speaking,
link |
this can be a huge technical success,
link |
and we could still lose that one game.
link |
And that's a whole nother theme of this, of the journey.
link |
But it was a success.
link |
It was not a success in natural language understanding,
link |
but that was not the goal.
link |
Yeah, that was, but I would argue,
link |
I understand what you're saying
link |
in terms of the science,
link |
but I would argue that the inspiration of it, right?
link |
The, not a success in terms of solving
link |
natural language understanding.
link |
There was a success of being an inspiration
link |
to future challenges.
link |
That drive future efforts.
link |
What's the difference between how human being
link |
compete in Jeopardy and how Watson does it?
link |
That's important in terms of intelligence.
link |
Yeah, so that actually came up very early on
link |
in the project also.
link |
In fact, I had people who wanted to be on the project
link |
who were early on, who sort of approached me
link |
once I committed to do it,
link |
had wanted to think about how humans do it.
link |
And they were, from a cognition perspective,
link |
like human cognition and how that should play.
link |
And I would not take them on the project
link |
because another assumption or another stake
link |
I put in the ground was,
link |
I don't really care how humans do this.
link |
At least in the context of this project.
link |
I need to build in the context of this project.
link |
In NLU and in building an AI that understands
link |
how it needs to ultimately communicate with humans,
link |
So it wasn't that I didn't care in general.
link |
In fact, as an AI scientist, I care a lot about that,
link |
but I'm also a practical engineer
link |
and I committed to getting this thing done
link |
and I wasn't gonna get distracted.
link |
I had to kind of say, like, if I'm gonna get this done,
link |
I'm gonna chart this path.
link |
And this path says, we're gonna engineer a machine
link |
that's gonna get this thing done.
link |
And we know what search and NLP can do.
link |
We have to build on that foundation.
link |
If I come in and take a different approach
link |
and start wondering about how the human mind
link |
might or might not do this,
link |
I'm not gonna get there from here in the timeframe.
link |
I think that's a great way to lead the team.
link |
But now that it's done and there's one,
link |
when you look back, analyze what's the difference actually.
link |
So I was a little bit surprised actually
link |
to discover over time, as this would come up
link |
from time to time and we'd reflect on it,
link |
and talking to Ken Jennings a little bit
link |
and hearing Ken Jennings talk about
link |
how he answered questions,
link |
that it might've been closer to the way humans
link |
answer questions than I might've imagined previously.
link |
Because humans are probably in the game of Jeopardy!
link |
at the level of Ken Jennings,
link |
are probably also cheating their way to winning, right?
link |
Not cheating, but shallow.
link |
Well, they're doing shallow analysis.
link |
They're doing the fastest possible.
link |
They're doing shallow analysis.
link |
So they are very quickly analyzing the question
link |
and coming up with some key vectors or cues, if you will.
link |
And they're taking those cues
link |
and they're very quickly going through
link |
like their library of stuff,
link |
not deeply reasoning about what's going on.
link |
And then sort of like a lots of different,
link |
like what we would call these scores,
link |
would kind of score that in a very shallow way
link |
and then say, oh, boom, you know, that's what it is.
link |
And so it's interesting as we reflected on that.
link |
So we may be doing something that's not too far off
link |
from the way humans do it,
link |
but we certainly didn't approach it by saying,
link |
how would a human do this?
link |
Now in elemental cognition,
link |
like the project I'm leading now,
link |
we ask those questions all the time
link |
because ultimately we're trying to do something
link |
that is to make the intelligence of the machine
link |
and the intelligence of the human very compatible.
link |
Well, compatible in the sense
link |
they can communicate with one another
link |
and they can reason with this shared understanding.
link |
So how they think about things and how they build answers,
link |
how they build explanations
link |
becomes a very important question to consider.
link |
So what's the difference between this open domain,
link |
but cold constructed question answering of Jeopardy
link |
and more something that requires understanding
link |
for shared communication with humans and machines?
link |
Yeah, well, this goes back to the interpretation
link |
of what we were talking about before.
link |
Jeopardy, the system's not trying to interpret the question
link |
and it's not interpreting the content it's reusing
link |
with regard to any particular framework.
link |
I mean, it is parsing it and parsing the content
link |
and using grammatical cues and stuff like that.
link |
So if you think of grammar as a human framework
link |
in some sense, it has that,
link |
but when you get into the richer semantic frameworks,
link |
what do people, how do they think, what motivates them,
link |
what are the events that are occurring
link |
and why are they occurring
link |
and what causes what else to happen
link |
and where are things in time and space?
link |
And like when you start thinking about how humans formulate
link |
and structure the knowledge that they acquire in their head
link |
and wasn't doing any of that.
link |
What do you think are the essential challenges
link |
of like free flowing communication, free flowing dialogue
link |
versus question answering even with the framework
link |
of the interpretation dialogue?
link |
Do you see free flowing dialogue
link |
as a fundamentally more difficult than question answering
link |
even with shared interpretation?
link |
So dialogue is important in a number of different ways.
link |
I mean, it's a challenge.
link |
So first of all, when I think about the machine that,
link |
when I think about a machine that understands language
link |
and ultimately can reason in an objective way
link |
that can take the information that it perceives
link |
through language or other means
link |
and connect it back to these frameworks,
link |
reason and explain itself,
link |
that system ultimately needs to be able to talk to humans
link |
or it needs to be able to interact with humans.
link |
So in some sense it needs to dialogue.
link |
That doesn't mean that it,
link |
sometimes people talk about dialogue and they think,
link |
you know, how do humans talk to like,
link |
talk to each other in a casual conversation
link |
and you can mimic casual conversations.
link |
We're not trying to mimic casual conversations.
link |
We're really trying to produce a machine
link |
whose goal is to help you think
link |
and help you reason about your answers and explain why.
link |
So instead of like talking to your friend down the street
link |
about having a small talk conversation
link |
with your friend down the street,
link |
this is more about like you would be communicating
link |
to the computer on Star Trek
link |
where like, what do you wanna think about?
link |
Like, what do you wanna reason about?
link |
I'm gonna tell you the information I have.
link |
I'm gonna have to summarize it.
link |
I'm gonna ask you questions.
link |
You're gonna answer those questions.
link |
I'm gonna go back and forth with you.
link |
I'm gonna figure out what your mental model is.
link |
I'm gonna now relate that to the information I have
link |
and present it to you in a way that you can understand it
link |
and then we could ask followup questions.
link |
So it's that type of dialogue that you wanna construct.
link |
It's more structured, it's more goal oriented,
link |
but it needs to be fluid.
link |
In other words, it has to be engaging and fluid.
link |
It has to be productive and not distracting.
link |
So there has to be a model of,
link |
in other words, the machine has to have a model
link |
of how humans think through things and discuss them.
link |
So basically a productive, rich conversation
link |
unlike this podcast.
link |
I'd like to think it's more similar to this podcast.
link |
I'll ask you about humor as well, actually.
link |
But what's the hardest part of that?
link |
Because it seems we're quite far away
link |
as a community from that still to be able to,
link |
so one is having a shared understanding.
link |
That's, I think, a lot of the stuff you said
link |
with frameworks is quite brilliant.
link |
But just creating a smooth discourse.
link |
It feels clunky right now.
link |
Which aspects of this whole problem
link |
that you just specified of having
link |
a productive conversation is the hardest?
link |
And that we're, or maybe any aspect of it
link |
you can comment on because it's so shrouded in mystery.
link |
So I think to do this you kind of have to be creative
link |
in the following sense.
link |
If I were to do this as purely a machine learning approach
link |
and someone said learn how to have a good,
link |
fluent, structured knowledge acquisition conversation,
link |
I'd go out and say, okay, I have to collect
link |
a bunch of data of people doing that.
link |
People reasoning well, having a good, structured
link |
conversation that both acquires knowledge efficiently
link |
as well as produces answers and explanations
link |
as part of the process.
link |
To collect the data.
link |
To collect the data because I don't know
link |
how much data is like that.
link |
Okay, there's one, there's a humorous commentary
link |
on the lack of rational discourse.
link |
But also even if it's out there, say it was out there,
link |
how do you actually annotate, like how do you collect
link |
an accessible example?
link |
Right, so I think any problem like this
link |
where you don't have enough data to represent
link |
the phenomenon you want to learn,
link |
in other words you want, if you have enough data
link |
you could potentially learn the pattern.
link |
In an example like this it's hard to do.
link |
This is sort of a human sort of thing to do.
link |
What recently came out at IBM was the debater projects
link |
and it's interesting, right, because now you do have
link |
these structured dialogues, these debate things
link |
where they did use machine learning techniques
link |
to generate these debates.
link |
Dialogues are a little bit tougher in my opinion
link |
than generating a structured argument
link |
where you have lots of other structured arguments
link |
like this, you could potentially annotate that data
link |
and you could say this is a good response,
link |
this is a bad response in a particular domain.
link |
Here I have to be responsive and I have to be opportunistic
link |
with regard to what is the human saying.
link |
So I'm goal oriented in saying I want to solve the problem,
link |
I want to acquire the knowledge necessary,
link |
but I also have to be opportunistic and responsive
link |
to what the human is saying.
link |
So I think that it's not clear that we could just train
link |
on the body of data to do this, but we could bootstrap it.
link |
In other words, we can be creative and we could say,
link |
what do we think the structure of a good dialogue is
link |
that does this well?
link |
And we can start to create that.
link |
If we can create that more programmatically,
link |
at least to get this process started
link |
and I can create a tool that now engages humans effectively,
link |
I could start generating data,
link |
I could start the human learning process
link |
and I can update my machine,
link |
but I could also start the automatic learning process
link |
as well, but I have to understand
link |
what features to even learn over.
link |
So I have to bootstrap the process a little bit first.
link |
And that's a creative design task
link |
that I could then use as input
link |
into a more automatic learning task.
link |
So some creativity in bootstrapping.
link |
What elements of a conversation
link |
do you think you would like to see?
link |
So one of the benchmarks for me is humor, right?
link |
That seems to be one of the hardest.
link |
And to me, the biggest contrast is sort of Watson.
link |
So one of the greatest sketches,
link |
comedy sketches of all time, right,
link |
is the SNL celebrity Jeopardy
link |
with Alex Trebek and Sean Connery
link |
and Burt Reynolds and so on,
link |
with Sean Connery commentating on Alex Trebek's
link |
while they're alive.
link |
And I think all of them are in the negative pointwise.
link |
So they're clearly all losing
link |
in terms of the game of Jeopardy,
link |
but they're winning in terms of comedy.
link |
So what do you think about humor in this whole interaction
link |
in the dialogue that's productive?
link |
Or even just what humor represents to me
link |
is the same idea that you're saying about framework,
link |
because humor only exists
link |
within a particular human framework.
link |
So what do you think about humor?
link |
What do you think about things like humor
link |
that connect to the kind of creativity
link |
you mentioned that's needed?
link |
I think there's a couple of things going on there.
link |
So I sort of feel like,
link |
and I might be too optimistic this way,
link |
but I think that there are,
link |
we did a little bit about with puns in Jeopardy.
link |
We literally sat down and said,
link |
And it's like wordplay,
link |
and you could formalize these things.
link |
So I think there's a lot aspects of humor
link |
that you could formalize.
link |
You could also learn humor.
link |
You could just say, what do people laugh at?
link |
And if you have enough, again,
link |
if you have enough data to represent the phenomenon,
link |
you might be able to weigh the features
link |
and figure out what humans find funny
link |
and what they don't find funny.
link |
The machine might not be able to explain
link |
why the human is funny unless we sit back
link |
and think about that more formally.
link |
I think, again, I think you do a combination of both.
link |
And I'm always a big proponent of that.
link |
I think robust architectures and approaches
link |
are always a little bit combination of us reflecting
link |
and being creative about how things are structured,
link |
how to formalize them,
link |
and then taking advantage of large data and doing learning
link |
and figuring out how to combine these two approaches.
link |
I think there's another aspect to humor though,
link |
which goes to the idea that I feel like I can relate
link |
to the person telling the story.
link |
And I think that's an interesting theme
link |
in the whole AI theme,
link |
which is, do I feel differently when I know it's a robot?
link |
And when I imagine that the robot is not conscious
link |
the way I'm conscious,
link |
when I imagine the robot does not actually
link |
have the experiences that I experience,
link |
do I find it funny?
link |
Or do, because it's not as related,
link |
I don't imagine that the person's relating it to it
link |
the way I relate to it.
link |
I think this also, you see this in the arts
link |
and in entertainment where,
link |
sometimes you have savants who are remarkable at a thing,
link |
whether it's sculpture or it's music or whatever,
link |
but the people who get the most attention
link |
are the people who can evoke a similar emotional response,
link |
who can get you to emote, right?
link |
About the way they are.
link |
In other words, who can basically make the connection
link |
from the artifact, from the music or the painting
link |
of the sculpture to the emotion
link |
and get you to share that emotion with them.
link |
And then, and that's when it becomes compelling.
link |
So they're communicating at a whole different level.
link |
They're just not communicating the artifact.
link |
They're communicating their emotional response
link |
And then you feel like, oh wow,
link |
I can relate to that person, I can connect to that,
link |
I can connect to that person.
link |
So I think humor has that aspect as well.
link |
So the idea that you can connect to that person,
link |
person being the critical thing,
link |
but we're also able to anthropomorphize objects pretty,
link |
robots and AI systems pretty well.
link |
So we're almost looking to make them human.
link |
So maybe from your experience with Watson,
link |
maybe you can comment on, did you consider that as part,
link |
well, obviously the problem of jeopardy
link |
doesn't require anthropomorphization, but nevertheless.
link |
Well, there was some interest in doing that.
link |
And that's another thing I didn't want to do
link |
because I didn't want to distract
link |
from the actual scientific task.
link |
But you're absolutely right.
link |
I mean, humans do anthropomorphize
link |
and without necessarily a lot of work.
link |
I mean, you just put some eyes
link |
and a couple of eyebrow movements
link |
and you're getting humans to react emotionally.
link |
And I think you can do that.
link |
So I didn't mean to suggest that,
link |
that that connection cannot be mimicked.
link |
I think that connection can be mimicked
link |
and can produce that emotional response.
link |
I just wonder though, if you're told what's really going on,
link |
if you know that the machine is not conscious,
link |
not having the same richness of emotional reactions
link |
and understanding that it doesn't really
link |
share the understanding,
link |
but it's essentially just moving its eyebrow
link |
or drooping its eyes or making them bigger,
link |
whatever it's doing, just getting the emotional response,
link |
will you still feel it?
link |
I think you probably would for a while.
link |
And then when it becomes more important
link |
that there's a deeper share of understanding,
link |
it may run flat, but I don't know.
link |
I'm pretty confident that majority of the world,
link |
even if you tell them how it works,
link |
well, it will not matter,
link |
especially if the machine herself says that she is conscious.
link |
That's very possible.
link |
So you, the scientist that made the machine is saying
link |
that this is how the algorithm works.
link |
Everybody will just assume you're lying
link |
and that there's a conscious being there.
link |
So you're deep into the science fiction genre now,
link |
I don't think it's, it's actually psychology.
link |
I think it's not science fiction.
link |
I think it's reality.
link |
I think it's a really powerful one
link |
that we'll have to be exploring in the next few decades.
link |
It's a very interesting element of intelligence.
link |
So what do you think,
link |
we've talked about social constructs of intelligences
link |
and frameworks and the way humans
link |
kind of interpret information.
link |
What do you think is a good test of intelligence
link |
So there's the Alan Turing with the Turing test.
link |
Watson accomplished something very impressive with Jeopardy.
link |
What do you think is a test
link |
that would impress the heck out of you
link |
that you saw that a computer could do?
link |
They would say, this is crossing a kind of threshold
link |
that gives me pause in a good way.
link |
My expectations for AI are generally high.
link |
What does high look like by the way?
link |
So not the threshold, test is a threshold.
link |
What do you think is the destination?
link |
What do you think is the ceiling?
link |
I think machines will in many measures
link |
will be better than us, will become more effective.
link |
In other words, better predictors about a lot of things
link |
than ultimately we can do.
link |
I think where they're gonna struggle
link |
is what we talked about before,
link |
which is relating to communicating with
link |
and understanding humans in deeper ways.
link |
And so I think that's a key point,
link |
like we can create the super parrot.
link |
What I mean by the super parrot is given enough data,
link |
a machine can mimic your emotional response,
link |
can even generate language that will sound smart
link |
and what someone else might say under similar circumstances.
link |
Like I would just pause on that,
link |
like that's the super parrot, right?
link |
So given similar circumstances,
link |
moves its faces in similar ways,
link |
changes its tone of voice in similar ways,
link |
produces strings of language that would similar
link |
that a human might say,
link |
not necessarily being able to produce
link |
a logical interpretation or understanding
link |
that would ultimately satisfy a critical interrogation
link |
or a critical understanding.
link |
I think you just described me in a nutshell.
link |
So I think philosophically speaking,
link |
you could argue that that's all we're doing
link |
as human beings to work super parrots.
link |
So I was gonna say, it's very possible,
link |
you know, humans do behave that way too.
link |
And so upon deeper probing and deeper interrogation,
link |
you may find out that there isn't a shared understanding
link |
because I think humans do both.
link |
Like humans are statistical language model machines
link |
and they are capable reasoners.
link |
You know, they're both.
link |
And you don't know which is going on, right?
link |
So, and I think it's an interesting problem.
link |
We talked earlier about like where we are
link |
in our social and political landscape.
link |
Can you distinguish someone who can string words together
link |
and sound like they know what they're talking about
link |
from someone who actually does?
link |
Can you do that without dialogue,
link |
without interrogative or probing dialogue?
link |
So it's interesting because humans are really good
link |
in their own mind, justifying or explaining what they hear
link |
because they project their understanding onto yours.
link |
So you could say, you could put together a string of words
link |
and someone will sit there and interpret it
link |
in a way that's extremely biased
link |
to the way they wanna interpret it.
link |
They wanna assume that you're an idiot
link |
and they'll interpret it one way.
link |
They will assume you're a genius
link |
and they'll interpret it another way that suits their needs.
link |
So this is tricky business.
link |
So I think to answer your question,
link |
as AI gets better and better, better and better mimic,
link |
you recreate the super parrots,
link |
we're challenged just as we are with,
link |
we're challenged with humans.
link |
Do you really know what you're talking about?
link |
Do you have a meaningful interpretation,
link |
a powerful framework that you could reason over
link |
and justify your answers, justify your predictions
link |
and your beliefs, why you think they make sense.
link |
Can you convince me what the implications are?
link |
So can you reason intelligently and make me believe
link |
that the implications of your prediction and so forth?
link |
So what happens is it becomes reflective.
link |
My standard for judging your intelligence
link |
depends a lot on mine.
link |
But you're saying there should be a large group of people
link |
with a certain standard of intelligence
link |
that would be convinced by this particular AI system.
link |
Then they'll pass.
link |
There should be, but I think depending on the content,
link |
one of the problems we have there
link |
is that if that large community of people
link |
are not judging it with regard to a rigorous standard
link |
of objective logic and reason, you still have a problem.
link |
Like masses of people can be persuaded.
link |
The millennials, yeah.
link |
To turn their brains off.
link |
By the way, I have nothing against the millennials.
link |
No, I don't, I'm just, just.
link |
So you're a part of one of the great benchmarks,
link |
challenges of AI history.
link |
What do you think about AlphaZero, OpenAI5,
link |
AlphaStar accomplishments on video games recently,
link |
which are also, I think, at least in the case of Go,
link |
with AlphaGo and AlphaZero playing Go,
link |
was a monumental accomplishment as well.
link |
What are your thoughts about that challenge?
link |
I think it was a giant landmark for AI.
link |
I think it was phenomenal.
link |
I mean, it was one of those other things
link |
nobody thought like solving Go was gonna be easy,
link |
particularly because it's hard for,
link |
particularly hard for humans.
link |
Hard for humans to learn, hard for humans to excel at.
link |
And so it was another measure, a measure of intelligence.
link |
I mean, it's very interesting what they did.
link |
And I loved how they solved the data problem,
link |
which again, they bootstrapped it
link |
and got the machine to play itself,
link |
to generate enough data to learn from.
link |
I think that was brilliant.
link |
I think that was great.
link |
And of course, the result speaks for itself.
link |
I think it makes us think about,
link |
again, it is, okay, what's intelligence?
link |
What aspects of intelligence are important?
link |
Can the Go machine help me make me a better Go player?
link |
Is it an alien intelligence?
link |
Am I even capable of,
link |
like again, if we put in very simple terms,
link |
it found the function, it found the Go function.
link |
Can I even comprehend the Go function?
link |
Can I talk about the Go function?
link |
Can I conceptualize the Go function,
link |
like whatever it might be?
link |
So one of the interesting ideas of that system
link |
is that it plays against itself, right?
link |
But there's no human in the loop there.
link |
So like you're saying, it could have by itself
link |
created an alien intelligence.
link |
Toward a Go, imagine you're sentencing,
link |
you're a judge and you're sentencing people,
link |
or you're setting policy,
link |
or you're making medical decisions,
link |
and you can't explain,
link |
you can't get anybody to understand
link |
what you're doing or why.
link |
So it's an interesting dilemma
link |
for the applications of AI.
link |
Do we hold AI to this accountability
link |
that says humans have to be willing
link |
to take responsibility for the decision?
link |
In other words, can you explain why you would do the thing?
link |
Will you get up and speak to other humans
link |
and convince them that this was a smart decision?
link |
Is the AI enabling you to do that?
link |
Can you get behind the logic that was made there?
link |
Do you think, sorry to land on this point,
link |
because it's a fascinating one.
link |
It's a great goal for AI.
link |
Do you think it's achievable in many cases?
link |
Or, okay, there's two possible worlds
link |
that we have in the future.
link |
One is where AI systems do like medical diagnosis
link |
or things like that, or drive a car
link |
without ever explaining to you why it fails when it does.
link |
That's one possible world and we're okay with it.
link |
Or the other where we are not okay with it
link |
and we really hold back the technology
link |
from getting too good before it's able to explain.
link |
Which of those worlds are more likely, do you think,
link |
and which are concerning to you or not?
link |
I think the reality is it's gonna be a mix.
link |
I'm not sure I have a problem with that.
link |
I mean, I think there are tasks that are perfectly fine
link |
with machines show a certain level of performance
link |
and that level of performance is already better than humans.
link |
So for example, I don't know that I take driverless cars.
link |
If driverless cars learn how to be more effective drivers
link |
than humans but can't explain what they're doing,
link |
but bottom line, statistically speaking,
link |
they're 10 times safer than humans,
link |
I don't know that I care.
link |
I think when we have these edge cases
link |
when something bad happens and we wanna decide
link |
who's liable for that thing and who made that mistake
link |
and what do we do about that?
link |
And I think those edge cases are interesting cases.
link |
And now do we go to designers of the AI
link |
and the AI says, I don't know if that's what it learned
link |
to do and it says, well, you didn't train it properly.
link |
You were negligent in the training data
link |
that you gave that machine.
link |
Like, how do we drive down the reliability?
link |
So I think those are interesting questions.
link |
So the optimization problem there, sorry,
link |
is to create an AI system that's able
link |
to explain the lawyers away.
link |
I think it's gonna be interesting.
link |
I mean, I think this is where technology
link |
and social discourse are gonna get like deeply intertwined
link |
and how we start thinking about problems, decisions
link |
and problems like that.
link |
I think in other cases it becomes more obvious
link |
where it's like, why did you decide
link |
to give that person a longer sentence or deny them parole?
link |
Again, policy decisions or why did you pick that treatment?
link |
Like that treatment ended up killing that guy.
link |
Like, why was that a reasonable choice to make?
link |
And people are gonna demand explanations.
link |
Now there's a reality though here.
link |
And the reality is that it's not,
link |
I'm not sure humans are making reasonable choices
link |
when they do these things.
link |
They are using statistical hunches, biases,
link |
or even systematically using statistical averages
link |
This is what happened to my dad
link |
and if you saw the talk I gave about that.
link |
But they decided that my father was brain dead.
link |
He had went into cardiac arrest
link |
and it took a long time for the ambulance to get there
link |
and he was not resuscitated right away and so forth.
link |
And they came and they told me he was brain dead
link |
and why was he brain dead?
link |
Because essentially they gave me
link |
a purely statistical argument under these conditions
link |
with these four features, 98% chance he's brain dead.
link |
I said, but can you just tell me not inductively,
link |
but deductively go there and tell me
link |
his brain's not functioning is the way for you to do that.
link |
And the protocol in response was,
link |
no, this is how we make this decision.
link |
I said, this is inadequate for me.
link |
I understand the statistics and I don't know how,
link |
there's a 2% chance he's still alive.
link |
I just don't know the specifics.
link |
I need the specifics of this case
link |
and I want the deductive logical argument
link |
about why you actually know he's brain dead.
link |
So I wouldn't sign the do not resuscitate.
link |
And I don't know, it was like they went through
link |
lots of procedures, it was a big long story,
link |
but the bottom was a fascinating story by the way,
link |
but how I reasoned and how the doctors reasoned
link |
through this whole process.
link |
But I don't know, somewhere around 24 hours later
link |
or something, he was sitting up in bed
link |
with zero brain damage.
link |
I mean, what lessons do you draw from that story,
link |
That the data that's being used
link |
to make statistical inferences
link |
doesn't adequately reflect the phenomenon.
link |
So in other words, you're getting shit wrong,
link |
I'm sorry, but you're getting stuff wrong
link |
because your model is not robust enough
link |
and you might be better off not using statistical inference
link |
and statistical averages in certain cases
link |
when you know the model's insufficient
link |
and that you should be reasoning about the specific case
link |
more logically and more deductibly
link |
and hold yourself responsible
link |
and hold yourself accountable to doing that.
link |
And perhaps AI has a role to say the exact thing
link |
what you just said, which is perhaps this is a case
link |
you should think for yourself,
link |
you should reason deductively.
link |
Well, so it's hard because it's hard to know that.
link |
You'd have to go back and you'd have to have enough data
link |
to essentially say, and this goes back to how do we,
link |
this goes back to the case of how do we decide
link |
whether the AI is good enough to do a particular task
link |
and regardless of whether or not
link |
it produces an explanation.
link |
And what standard do we hold for that?
link |
So if you look more broadly, for example,
link |
as my father, as a medical case,
link |
the medical system ultimately helped him a lot
link |
throughout his life, without it,
link |
he probably would have died much sooner.
link |
So overall, it sort of worked for him
link |
in sort of a net, net kind of way.
link |
Actually, I don't know that that's fair.
link |
But maybe not in that particular case, but overall,
link |
like the medical system overall does more good than bad.
link |
Yeah, the medical system overall
link |
was doing more good than bad.
link |
Now, there's another argument that suggests
link |
that wasn't the case, but for the sake of argument,
link |
let's say like that's, let's say a net positive.
link |
And I think you have to sit there and there
link |
and take that into consideration.
link |
Now you look at a particular use case,
link |
like for example, making this decision,
link |
have you done enough studies to know
link |
how good that prediction really is?
link |
And have you done enough studies to compare it,
link |
to say, well, what if we dug in in a more direct,
link |
let's get the evidence, let's do the deductive thing
link |
and not use statistics here,
link |
how often would that have done better?
link |
So you have to do the studies
link |
to know how good the AI actually is.
link |
And it's complicated because it depends how fast
link |
you have to make the decision.
link |
So if you have to make the decision super fast,
link |
you have no choice.
link |
If you have more time, right?
link |
But if you're ready to pull the plug,
link |
and this is a lot of the argument that I had with a doctor,
link |
I said, what's he gonna do if you do it,
link |
what's gonna happen to him in that room if you do it my way?
link |
You know, well, he's gonna die anyway.
link |
So let's do it my way then.
link |
I mean, it raises questions for our society
link |
to struggle with, as the case with your father,
link |
but also when things like race and gender
link |
start coming into play when certain,
link |
when judgments are made based on things
link |
that are complicated in our society,
link |
at least in the discourse.
link |
And it starts, you know, I think I'm safe to say
link |
that most of the violent crimes committed
link |
by males, so if you discriminate based,
link |
you know, it's a male versus female saying that
link |
if it's a male, more likely to commit the crime.
link |
This is one of my very positive and optimistic views
link |
of why the study of artificial intelligence,
link |
the process of thinking and reasoning logically
link |
and statistically, and how to combine them
link |
is so important for the discourse today,
link |
because it's causing a, regardless of what state AI devices
link |
are or not, it's causing this dialogue to happen.
link |
This is one of the most important dialogues
link |
that in my view, the human species can have right now,
link |
which is how to think well, how to reason well,
link |
how to understand our own cognitive biases
link |
and what to do about them.
link |
That has got to be one of the most important things
link |
we as a species can be doing, honestly.
link |
We are, we've created an incredibly complex society.
link |
We've created amazing abilities to amplify noise faster
link |
than we can amplify signal.
link |
We are challenged.
link |
We are deeply, deeply challenged.
link |
We have, you know, big segments of the population
link |
getting hit with enormous amounts of information.
link |
Do they know how to do critical thinking?
link |
Do they know how to objectively reason?
link |
Do they understand what they are doing,
link |
nevermind what their AI is doing?
link |
This is such an important dialogue to be having.
link |
And, you know, we are fundamentally,
link |
our thinking can be and easily becomes fundamentally bias.
link |
And there are statistics and we shouldn't blind our,
link |
we shouldn't discard statistical inference,
link |
but we should understand the nature
link |
of statistical inference.
link |
As a society, as you know,
link |
we decide to reject statistical inference
link |
to favor understanding and deciding on the individual.
link |
We consciously make that choice.
link |
So even if the statistics said,
link |
even if the statistics said males are more likely to have,
link |
you know, to be violent criminals,
link |
we still take each person as an individual
link |
and we treat them based on the logic
link |
and the knowledge of that situation.
link |
We purposefully and intentionally
link |
reject the statistical inference.
link |
We do that out of respect for the individual.
link |
For the individual.
link |
Yeah, and that requires reasoning and thinking.
link |
Looking forward, what grand challenges
link |
would you like to see in the future?
link |
Because the Jeopardy challenge, you know,
link |
captivated the world.
link |
AlphaGo, AlphaZero captivated the world.
link |
Deep Blue certainly beating Kasparov.
link |
Gary's bitterness aside captivated the world.
link |
What do you think, do you have ideas
link |
for next grand challenges for future challenges of that?
link |
You know, look, I mean, I think there are lots
link |
of really great ideas for grand challenges.
link |
I'm particularly focused on one right now,
link |
which is, you know, can you demonstrate
link |
that they understand, that they could read and understand,
link |
that they can acquire these frameworks
link |
and communicate, you know,
link |
reason and communicate with humans.
link |
So it is kind of like the Turing test,
link |
but it's a little bit more demanding than the Turing test.
link |
It's not enough to convince me that you might be human
link |
because you could, you know, you can parrot a conversation.
link |
I think, you know, the standard is a little bit higher,
link |
is for example, can you, you know, the standard is higher.
link |
And I think one of the challenges
link |
of devising this grand challenge is that we're not sure
link |
what intelligence is, we're not sure how to determine
link |
whether or not two people actually understand each other
link |
and in what depth they understand it, you know,
link |
to what depth they understand each other.
link |
So the challenge becomes something along the lines of,
link |
can you satisfy me that we have a shared understanding?
link |
So if I were to probe and probe and you probe me,
link |
can machines really act like thought partners
link |
where they can satisfy me that we have a shared,
link |
our understanding is shared enough
link |
that we can collaborate and produce answers together
link |
and that, you know, they can help me explain
link |
and justify those answers.
link |
So maybe here's an idea.
link |
So we'll have AI system run for president and convince.
link |
I'm sorry, go ahead.
link |
Well, no, you have to convince the voters
link |
that they should vote.
link |
So like, I guess what does winning look like?
link |
Again, that's why I think this is such a challenge
link |
because we go back to the emotional persuasion.
link |
We go back to, you know, now we're checking off an aspect
link |
of human cognition that is in many ways weak or flawed,
link |
right, we're so easily manipulated.
link |
Our minds are drawn for often the wrong reasons, right?
link |
Not the reasons that ultimately matter to us,
link |
but the reasons that can easily persuade us.
link |
I think we can be persuaded to believe one thing or another
link |
for reasons that ultimately don't serve us well
link |
And a good benchmark should not play with those elements
link |
of emotional manipulation.
link |
And I think that's where we have to set the higher standard
link |
for ourselves of what, you know, what does it mean?
link |
This goes back to rationality
link |
and it goes back to objective thinking.
link |
And can you produce, can you acquire information
link |
and produce reasoned arguments
link |
and to those reasoned arguments
link |
pass a certain amount of muster and is it,
link |
and can you acquire new knowledge?
link |
You know, can you, for example, can you reason,
link |
I have acquired new knowledge,
link |
can you identify where it's consistent or contradictory
link |
with other things you've learned?
link |
And can you explain that to me
link |
and get me to understand that?
link |
So I think another way to think about it perhaps
link |
is can a machine teach you, can it help you understand
link |
something that you didn't really understand before
link |
where it's taking you, so you're not,
link |
again, it's almost like can it teach you,
link |
can it help you learn and in an arbitrary space
link |
so it can open those domain space?
link |
So can you tell the machine, and again,
link |
this borrows from some science fiction,
link |
but can you go off and learn about this topic
link |
that I'd like to understand better
link |
and then work with me to help me understand it?
link |
That's quite brilliant.
link |
What, the machine that passes that kind of test,
link |
do you think it would need to have self awareness
link |
or even consciousness?
link |
What do you think about consciousness
link |
and the importance of it maybe in relation to having a body,
link |
having a presence, an entity?
link |
Do you think that's important?
link |
You know, people used to ask me if Watson was conscious
link |
and I used to say, he's conscious of what exactly?
link |
I mean, I think, you know, maybe it depends
link |
what it is that you're conscious of.
link |
I mean, like, so, you know, did it, if you, you know,
link |
it's certainly easy for it to answer questions
link |
about, it would be trivial to program it
link |
to answer questions about whether or not
link |
it was playing Jeopardy.
link |
I mean, it could certainly answer questions
link |
that would imply that it was aware of things.
link |
Exactly, what does it mean to be aware
link |
and what does it mean to be conscious of?
link |
It's sort of interesting.
link |
I mean, I think that we differ from one another
link |
based on what we're conscious of.
link |
But wait, wait a minute, yes, for sure.
link |
There's degrees of consciousness in there, so.
link |
Well, and there's just areas.
link |
Like, it's not just degrees, what are you aware of?
link |
Like, what are you not aware of?
link |
But nevertheless, there's a very subjective element
link |
to our experience.
link |
Let me even not talk about consciousness.
link |
Let me talk about another, to me,
link |
really interesting topic of mortality, fear of mortality.
link |
Watson, as far as I could tell,
link |
did not have a fear of death.
link |
Most, most humans do.
link |
Wasn't conscious of death.
link |
So there's an element of finiteness to our existence
link |
that I think, like you mentioned, survival,
link |
that adds to the whole thing.
link |
I mean, consciousness is tied up with that,
link |
that we are a thing.
link |
It's a subjective thing that ends.
link |
And that seems to add a color and flavor
link |
to our motivations in a way
link |
that seems to be fundamentally important for intelligence,
link |
or at least the kind of human intelligence.
link |
Well, I think for generating goals, again,
link |
I think you could have,
link |
you could have an intelligence capability
link |
and a capability to learn, a capability to predict.
link |
But I think without,
link |
I mean, again, you get fear,
link |
but essentially without the goal to survive.
link |
So you think you can just encode that
link |
without having to really?
link |
I think you could encode.
link |
I mean, you could create a robot now,
link |
and you could say, you know, plug it in,
link |
and say, protect your power source, you know,
link |
and give it some capabilities,
link |
and it'll sit there and operate
link |
to try to protect its power source and survive.
link |
I mean, so I don't know
link |
that that's philosophically a hard thing to demonstrate.
link |
It sounds like a fairly easy thing to demonstrate
link |
that you can give it that goal.
link |
Will it come up with that goal by itself?
link |
I think you have to program that goal in.
link |
But there's something,
link |
because I think, as we touched on,
link |
intelligence is kind of like a social construct.
link |
The fact that a robot will be protecting its power source
link |
would add depth and grounding to its intelligence
link |
in terms of us being able to respect it.
link |
I mean, ultimately, it boils down to us acknowledging
link |
that it's intelligent.
link |
And the fact that it can die,
link |
I think, is an important part of that.
link |
The interesting thing to reflect on
link |
is how trivial that would be.
link |
And I don't think, if you knew how trivial that was,
link |
you would associate that with being intelligence.
link |
I mean, I literally put in a statement of code
link |
that says you have the following actions you can take.
link |
You give it a bunch of actions,
link |
like maybe you mount a laser gun on it,
link |
or you give it the ability to scream or screech or whatever.
link |
And you say, if you see your power source threatened,
link |
then you could program that in,
link |
and you're gonna take these actions to protect it.
link |
You know, you could train it on a bunch of things.
link |
So, and now you're gonna look at that and you say,
link |
well, you know, that's intelligence,
link |
which is protecting its power source?
link |
Maybe, but that's, again, this human bias that says,
link |
the thing I identify, my intelligence and my conscious,
link |
so fundamentally with the desire,
link |
or at least the behaviors associated
link |
with the desire to survive,
link |
that if I see another thing doing that,
link |
I'm going to assume it's intelligent.
link |
What timeline, year,
link |
will society have something that would,
link |
that you would be comfortable calling
link |
an artificial general intelligence system?
link |
Well, what's your intuition?
link |
Nobody can predict the future,
link |
certainly not the next few months or 20 years away,
link |
but what's your intuition?
link |
How far away are we?
link |
It's hard to make these predictions.
link |
I mean, I would be guessing,
link |
and there's so many different variables,
link |
including just how much we want to invest in it
link |
and how important we think it is,
link |
what kind of investment we're willing to make in it,
link |
what kind of talent we end up bringing to the table,
link |
the incentive structure, all these things.
link |
So I think it is possible to do this sort of thing.
link |
I think it's, I think trying to sort of
link |
ignore many of the variables and things like that,
link |
is it a 10 year thing, is it a 23 year?
link |
Probably closer to a 20 year thing, I guess.
link |
But not several hundred years.
link |
No, I don't think it's several hundred years.
link |
I don't think it's several hundred years.
link |
But again, so much depends on how committed we are
link |
to investing and incentivizing this type of work.
link |
And it's sort of interesting.
link |
Like, I don't think it's obvious how incentivized we are.
link |
I think from a task perspective,
link |
if we see business opportunities to take this technique
link |
or that technique to solve that problem,
link |
I think that's the main driver for many of these things.
link |
From a general intelligence,
link |
it's kind of an interesting question.
link |
Are we really motivated to do that?
link |
And like, we just struggled ourselves right now
link |
to even define what it is.
link |
So it's hard to incentivize
link |
when we don't even know what it is
link |
we're incentivized to create.
link |
And if you said mimic a human intelligence,
link |
I just think there are so many challenges
link |
with the significance and meaning of that.
link |
That there's not a clear directive.
link |
There's no clear directive to do precisely that thing.
link |
So assistance in a larger and larger number of tasks.
link |
a system that's particularly able to operate my microwave
link |
and making a grilled cheese sandwich.
link |
I don't even know how to make one of those.
link |
And then the same system will be doing the vacuum cleaning.
link |
And then the same system would be teaching
link |
my kids that I don't have math.
link |
I think that when you get into a general intelligence
link |
for learning physical tasks,
link |
and again, I wanna go back to your body question
link |
because I think your body question was interesting,
link |
but you wanna go back to learning the abilities
link |
to physical tasks.
link |
You might have, we might get,
link |
I imagine in that timeframe,
link |
we will get better and better at learning these kinds
link |
of tasks, whether it's mowing your lawn
link |
or driving a car or whatever it is.
link |
I think we will get better and better at that
link |
where it's learning how to make predictions
link |
over large bodies of data.
link |
I think we're gonna continue to get better
link |
and better at that.
link |
And machines will outpace humans
link |
in a variety of those things.
link |
The underlying mechanisms for doing that may be the same,
link |
meaning that maybe these are deep nets,
link |
there's infrastructure to train them,
link |
reusable components to get them to do different classes
link |
of tasks, and we get better and better
link |
at building these kinds of machines.
link |
You could argue that the general learning infrastructure
link |
in there is a form of a general type of intelligence.
link |
I think what starts getting harder is this notion of,
link |
can we effectively communicate and understand
link |
and build that shared understanding?
link |
Because of the layers of interpretation that are required
link |
to do that, and the need for the machine to be engaged
link |
with humans at that level in a continuous basis.
link |
So how do you get the machine in the game?
link |
How do you get the machine in the intellectual game?
link |
Yeah, and to solve AGI,
link |
you probably have to solve that problem.
link |
You have to get the machine,
link |
so it's a little bit of a bootstrapping thing.
link |
Can we get the machine engaged in the intellectual game,
link |
but in the intellectual dialogue with the humans?
link |
Are the humans sufficiently in intellectual dialogue
link |
with each other to generate enough data in this context?
link |
And how do you bootstrap that?
link |
Because every one of those conversations,
link |
every one of those conversations,
link |
those intelligent interactions,
link |
require so much prior knowledge
link |
that it's a challenge to bootstrap it.
link |
So the question is, and how committed?
link |
So I think that's possible, but when I go back to,
link |
are we incentivized to do that?
link |
I know we're incentivized to do the former.
link |
Are we incentivized to do the latter significantly enough?
link |
Do people understand what the latter really is well enough?
link |
Part of the elemental cognition mission
link |
is to try to articulate that better and better
link |
through demonstrations
link |
and through trying to craft these grand challenges
link |
and get people to say, look,
link |
this is a class of intelligence.
link |
This is a class of AI.
link |
What is the potential of this?
link |
What's the business potential?
link |
What's the societal potential to that?
link |
And to build up that incentive system around that.
link |
Yeah, I think if people don't understand yet,
link |
I think they will.
link |
I think there's a huge business potential here.
link |
So it's exciting that you're working on it.
link |
We kind of skipped over,
link |
but I'm a huge fan of physical presence of things.
link |
Do you think Watson had a body?
link |
Do you think having a body adds to the interactive element
link |
between the AI system and a human,
link |
or just in general to intelligence?
link |
So I think going back to that shared understanding bit,
link |
humans are very connected to their bodies.
link |
I mean, one of the challenges in getting an AI
link |
to kind of be a compatible human intelligence
link |
is that our physical bodies are generating a lot of features
link |
that make up the input.
link |
So in other words, our bodies are the tool
link |
we use to affect output,
link |
but they also generate a lot of input for our brains.
link |
So we generate emotion, we generate all these feelings,
link |
we generate all these signals that machines don't have.
link |
So machines don't have this as the input data
link |
and they don't have the feedback that says,
link |
I've gotten this emotion or I've gotten this idea,
link |
I now want to process it,
link |
and then it then affects me as a physical being,
link |
and I can play that out.
link |
In other words, I could realize the implications of that,
link |
implications again, on my mind body complex,
link |
I then process that, and the implications again,
link |
our internal features are generated, I learn from them,
link |
they have an effect on my mind body complex.
link |
So it's interesting when we think,
link |
do we want a human intelligence?
link |
Well, if we want a human compatible intelligence,
link |
probably the best thing to do
link |
is to embed it in a human body.
link |
Just to clarify, and both concepts are beautiful,
link |
is humanoid robots, so robots that look like humans is one,
link |
or did you mean actually sort of what Elon Musk
link |
was working with Neuralink,
link |
really embedding intelligence systems
link |
to ride along human bodies?
link |
No, I mean riding along is different.
link |
I meant like if you want to create an intelligence
link |
that is human compatible,
link |
meaning that it can learn and develop
link |
a shared understanding of the world around it,
link |
you have to give it a lot of the same substrate.
link |
Part of that substrate is the idea
link |
that it generates these kinds of internal features,
link |
like sort of emotional stuff, it has similar senses,
link |
it has to do a lot of the same things
link |
with those same senses, right?
link |
So I think if you want that,
link |
again, I don't know that you want that.
link |
That's not my specific goal,
link |
I think that's a fascinating scientific goal,
link |
I think it has all kinds of other implications.
link |
That's sort of not the goal.
link |
I want to create, I think of it
link |
as I create intellectual thought partners for humans,
link |
so that kind of intelligence.
link |
I know there are other companies
link |
that are creating physical thought partners,
link |
physical partners for humans,
link |
but that's kind of not where I'm at.
link |
But the important point is that a big part
link |
of what we process is that physical experience
link |
of the world around us.
link |
On the point of thought partners,
link |
what role does an emotional connection,
link |
or forgive me, love, have to play
link |
in that thought partnership?
link |
Is that something you're interested in,
link |
put another way, sort of having a deep connection,
link |
beyond intellectual?
link |
Yeah, with the AI, between human and AI.
link |
Is that something that gets in the way
link |
of the rational discourse?
link |
Is that something that's useful?
link |
I worry about biases, obviously.
link |
So in other words, if you develop an emotional relationship
link |
with a machine, all of a sudden you start,
link |
are more likely to believe what it's saying,
link |
even if it doesn't make any sense.
link |
So I worry about that.
link |
But at the same time,
link |
I think the opportunity to use machines
link |
to provide human companionship is actually not crazy.
link |
And intellectual and social companionship
link |
is not a crazy idea.
link |
Do you have concerns, as a few people do,
link |
Elon Musk, Sam Harris,
link |
about long term existential threats of AI,
link |
and perhaps short term threats of AI?
link |
We talked about bias,
link |
we talked about different misuses,
link |
but do you have concerns about thought partners,
link |
systems that are able to help us make decisions
link |
together as humans,
link |
somehow having a significant negative impact
link |
on society in the long term?
link |
I think there are things to worry about.
link |
I think giving machines too much leverage is a problem.
link |
And what I mean by leverage is,
link |
is too much control over things that can hurt us,
link |
whether it's socially, psychologically, intellectually,
link |
And if you give the machines too much control,
link |
I think that's a concern.
link |
You forget about the AI,
link |
just once you give them too much control,
link |
human bad actors can hack them and produce havoc.
link |
So that's a problem.