back to indexDavid Ferrucci: IBM Watson, Jeopardy & Deep Conversations with AI | Lex Fridman Podcast #44
link |
The following is a conversation with David Ferrochi.
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 and joining it to have a lot of wisdom
link |
earned 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 or simply connect with me on Twitter.
link |
Alex Friedman spelled F R I D M A N.
link |
And now here's my conversation with David Ferrochi.
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 there is a substantive difference.
link |
And I think the thing that got me into computer science
link |
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, 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,
link |
where the intelligence process itself,
link |
then what would be the difference?
link |
So from a philosophical standpoint,
link |
I'm not trying to convince 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's 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 for us
link |
to be able to diagnose and treat issues
link |
for us to understand our own strengths and weaknesses,
link |
both intellectual, psychological, and physical.
link |
So neuroscience and understanding the brain
link |
from that perspective, there's a clear, clear goal there.
link |
From the perspective of saying,
link |
I want to 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 want to 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 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 even in and of itself suggests
link |
that human intelligence is something quite enviable,
link |
that it can introspect that way.
link |
And the flaws, you mentioned the flaws,
link |
the humans have flaws.
link |
Yeah, I think that flaws that human intelligence has
link |
is extremely prejudicial and bias
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,
link |
the fear, what are the flaws?
link |
List them all, what love, maybe that's a flaw.
link |
You think those are all things that can be gotten,
link |
get in the weight 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,
link |
rigorously reason, develop answers
link |
and justify those answers in an objective way,
link |
yeah, then human intelligence has these flaws
link |
and 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, 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
link |
makes a lot of sense for when your primary goal
link |
is to act quickly 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 that responds quickly
link |
and more naturally?
link |
Because that's the thing we kinda need to not be eaten
link |
by the predators in the world.
link |
For example, but then we've learned
link |
to reason through logic, we've developed science,
link |
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, 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 can you actually
link |
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 because we'll get to Jeopardy and Beyond.
link |
That's like an incredible, one of the most incredible
link |
accomplishments 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 two, 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 |
So in a highly dynamic environment full of uncertainty,
link |
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 |
If I can do that with less data and less training time,
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 to interrupt again.
link |
Well, that's interesting about picking a goal,
link |
sort of an interesting thing.
link |
And I think that's where you bring in,
link |
what do you pre program to do?
link |
We talk about humans and well,
link |
humans are pre programmed 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
link |
with complex, excessive 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 |
We're 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.
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 to predict
link |
the future, and the second is me being able to,
link |
like impressing me that you're intelligent,
link |
me being able to know that you successfully predicted
link |
the future, 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 me?
link |
Because of my ability to predict.
link |
So I would look at the metrics and I'd say, wow,
link |
you're right, 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 this is interesting, right?
link |
Because now you're in this weird place
link |
where for you to be recognized as intelligent
link |
the way I'm intelligent,
link |
then you and I sort of have to be able to communicate.
link |
And then 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,
link |
for example, animals can do things,
link |
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 low.
link |
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, look at political debates
link |
and discourse on Twitter,
link |
it's mostly just telling stories.
link |
So 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
link |
Well, 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, otherwise.
link |
In fact, there have been...
link |
That's the metric of success, yeah.
link |
There have been several proofs out there
link |
where mathematicians would study for a long time
link |
before they were convinced that it actually proved anything.
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.
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 could do this in a way that other people
link |
can understand and replicate and that it makes sense to them.
link |
So our 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 |
So do you think the general question of intelligence
link |
is then also a social construct?
link |
So if we ask questions
link |
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 a super objective way that says, here's this data.
link |
I want to 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 subordinate.
link |
I think it is, 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 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, whether you're a machine
link |
or another human, frankly, are now obliged
link |
to make me understand how it is that you're arriving
link |
at that answer and allow me, me or obviously a community
link |
or a judge of people, to decide whether or not
link |
And by the way, that happens with 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, well, I think you should buy, and I
link |
should think you should buy this much, or have or sell
link |
or whatever it is, or I think you should launch the product
link |
today or tomorrow or launch this product versus that product,
link |
whatever the decision may be.
link |
And the person said, I just have a good feeling about it.
link |
And you're not very satisfied.
link |
Now, that person could be, you might say, well, you've
link |
been right before, but I'm going to 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 have to convince the other person.
link |
You'll still be wrong.
link |
You'll still be wrong.
link |
You just got to be convincing.
link |
But it's ultimately got to be convincing.
link |
And that's why I'm saying we're bound together.
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 |
and this is a very important point,
link |
you're giving me an explanation.
link |
And I'm not good at reasoning well and being objective
link |
and following logical paths and consistent paths.
link |
And I'm not good at measuring and computing probabilities
link |
across those paths.
link |
What happens is, collectively, we're not going to do well.
link |
How hard is that problem, the second one?
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, the reasoning,
link |
how hard is that problem?
link |
I think that's very hard.
link |
I mean, I think that's, 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 and vice versa.
link |
So how hard is that problem for computers?
link |
I think it's hard for computers.
link |
And the reason why I related to saying that it's also
link |
hard for humans is because I think when we step back
link |
and we say we want to 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, if you
link |
look at the entire enterprise of science,
link |
science is supposed to be at about objective reason.
link |
So we think about, 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 or the scientists
link |
or the philosophers who kind of work through the details
link |
and write the papers and come up with the thoughtful, logical
link |
proves and use the scientific method.
link |
And 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 |
What's the process of training people to do that well?
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 wade them.
link |
Obviously, we talked about this, like human flaws
link |
or weaknesses, we can persuade them through emotional means.
link |
But to get them to understand and connect to and follow
link |
a logical argument is difficult.
link |
We do it as scientists, 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, well, how
link |
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, because there's
link |
an optimistic notion 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 look at next, whether it's Facebook, Google,
link |
advertisement based companies, their goal
link |
is to convince you to buy things based on anything.
link |
So that could be reason, because the best of advertisement
link |
is showing you things that you really do need
link |
and explain why you need it.
link |
But it could also be through emotional manipulation.
link |
The algorithm that describes why a certain reason,
link |
a certain decision was made.
link |
How hard is it to do it through emotional manipulation?
link |
And why is that a good or a bad thing?
link |
So you've kind of focused on reason, logic, really
link |
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 in the reasoning
link |
aspect and the emotional manipulation?
link |
So you call it emotional manipulation,
link |
but more objectively, it's essentially
link |
saying 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'm not saying it's good, right, or wrong.
link |
It works to get your attention, and it
link |
works to get you to buy stuff.
link |
And when you think about algorithms
link |
that look at the patterns of features
link |
that you seem to be spending your money on,
link |
and say, I'm going to give you something
link |
with a similar pattern, I'm going to 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 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 for convincing.
link |
For telling us the story.
link |
For convincing humans, it's good because, again, this
link |
goes back to, what is the human behavior like?
link |
What 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, you've already
link |
reasoned that you need them.
link |
And these algorithms are saying, look, that's up to you.
link |
The reason whether you need something or not, that's your job.
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 |
It could be a bad reason.
link |
It could be a good reason.
link |
I'm going to 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 |
And the algorithm is doing what it does,
link |
which is saying, you seem to be interested in this.
link |
I'm going to 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 going to show you other stuff
link |
with similar features.
link |
And I wash my hands from it.
link |
And I say, that's all that's going on.
link |
People are so harsh on AI systems.
link |
So one, the bar of performance is extremely high.
link |
And yet, we also ask them, in the case of social media,
link |
to help find the better angels of our nature
link |
and help make a better society.
link |
So what do you think about the role of AI?
link |
So that's, 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, which
link |
is finding the patterns.
link |
We're not building a, the system's not building a theory
link |
that is consumable and understandable
link |
to other humans that can be explained and justified.
link |
And so on one hand, to say, oh, AI is doing this.
link |
Why isn't doing this other thing?
link |
Well, this other thing is 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, in other words,
link |
understandings of what's important in the world,
link |
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, what's good,
link |
what's bad, what's moral, 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 on this stuff
link |
you're clicking, 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 and deeper thought processes.
link |
So that's what meaning means, is not just some kind of deep,
link |
timeless, semantic thing that the statement represents,
link |
but also how a large number of people
link |
are likely to interpret.
link |
So it's, again, even meaning is a social construct.
link |
So you have to try to predict how most people would
link |
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 to the artist
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 want
link |
to narrow 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, you can sit there
link |
and you can interpret lots of different ways
link |
at many, many different levels.
link |
But when I want to align our understanding of that,
link |
I have to specify a lot more stuff that's actually not
link |
directly in the artifact.
link |
Now, I have to say, well, how are you
link |
interpreting this image and that image?
link |
And what about the colors?
link |
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, well, if this is
link |
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 and thinking 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, as a society.
link |
We have the shared experience.
link |
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 similar, what we like to call prior models
link |
about the word prior experiences.
link |
And we use that as a, think of it as a wide collection
link |
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 programmed stuff.
link |
And they're able to communicate
link |
because they have a lot of,
link |
because they share that stuff.
link |
Do you think that shared knowledge,
link |
if we can maybe escape the hardware question,
link |
how much is encoded in the hardware?
link |
Just the shared knowledge in the software,
link |
the history of the many centuries of wars
link |
and so on 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 for a machine,
link |
to program a machine, to acquire that knowledge
link |
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 |
And so in other words, we have, if you will, as humans,
link |
we have a framework for interpreting the world around us.
link |
So we have multiple frameworks
link |
for interpreting the world around us.
link |
But if you're interpreting, for example,
link |
social political interactions,
link |
you're thinking about whether there's people,
link |
there's collections in groups of people,
link |
goals are largely built around survival and quality of life.
link |
There are fundamental economics around scarcity of resources.
link |
And when humans come and start interpreting a situation
link |
like that, because you brought up historical events,
link |
they start interpreting situations like that.
link |
They apply a lot of this,
link |
a lot of this fundamental framework for interpreting that.
link |
Well, who are the people?
link |
What were their goals?
link |
What reasons 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 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 and a given situation,
link |
are then able to interpret it with regards to that framework.
link |
And then given that interpretation,
link |
they can do what they can predict.
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
link |
and interpret events 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 |
It's tremendous amount of detailed knowledge in the world.
link |
You can imagine effectively infinite number
link |
of unique situations and unique configurations
link |
But the knowledge that you need,
link |
what I referred to as like the frameworks,
link |
for you need for interpreting them, I don't think.
link |
I think that those are finite.
link |
You think the frameworks are more important
link |
than the bulk of the knowledge.
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 it,
link |
to interpret and reason over the specifics
link |
in ways that other humans would understand.
link |
What about the specifics?
link |
Were you required the specifics by reading
link |
and by talking to other people?
link |
So I'm mostly actually just even,
link |
if you 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 is really required
link |
to perform even some of these basic tasks.
link |
Do you have that sense as well?
link |
And 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 you're talking about sort of the physics,
link |
the basic physics around us,
link |
for example, acquiring information
link |
about acquiring how that works.
link |
Yeah, I think there's a combination of things going on.
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 you start assuming that
link |
and with similar input,
link |
I'm gonna predict similar outputs.
link |
You don't 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 and counting
link |
on them falling 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, if 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 social political 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 |
Sorry to linger on it because again,
link |
and we'll get to it for sure as Watson with Jeopardy did
link |
is take on a problem 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 reasoning in the world about both gravity
link |
and politics, 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 not as convinced yet.
link |
No, I think it is solvable.
link |
I mean, I think that it's, first of all,
link |
it's about getting machines to learn.
link |
Learning is fundamental.
link |
And I think we're already in a place
link |
that we understand, for example,
link |
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 there are 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 want to 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
link |
and acquire, 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 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
link |
or representing, if you look back to the 80s and 90s
link |
of the expert systems,
link |
so 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,
link |
when you collect both the frameworks
link |
and the knowledge from the data,
link |
what do you think that thing will look like?
link |
I think asking the question,
link |
they look like neural networks is a bit of a red herring.
link |
I think that they will certainly do inductive
link |
or pattern match based reasoning.
link |
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 order 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 |
For example, at Elemental Cognition, we do both.
link |
We have architectures that do both,
link |
but both those things,
link |
but also have a learning method
link |
for acquiring the frameworks themselves
link |
and saying, 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 |
There is a fundamental knowledge representation,
link |
like you saying, like these graphs of logic, if you will.
link |
There are also neural networks
link |
that acquire certain class of information.
link |
Then they align them with these frameworks,
link |
but there's also a mechanism to acquire the frameworks themselves.
link |
It seems like the idea of frameworks
link |
requires some kind of collaboration with humans.
link |
Do you think of that collaboration as different?
link |
Only for the express purpose
link |
that you're designing machine,
link |
you're designing an intelligence
link |
that can ultimately communicate with humans
link |
in the terms of frameworks
link |
that help them understand things.
link |
To be really clear,
link |
you can independently create
link |
a machine learning system
link |
that I might call an alien intelligence
link |
that does a better job than you would 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
link |
that is inexplicable to you.
link |
But you're more interested in a case where you can.
link |
My sort of approach to AI
link |
is because 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
link |
machine learning techniques are good at,
link |
which is to observe
link |
whether it be in language or whether it be in images
link |
or videos or whatever,
link |
to acquire these patterns
link |
the generalizations from those patterns,
link |
but then ultimately to work with humans
link |
to connect them to frameworks,
link |
interpretations, if you will,
link |
that ultimately make sense to humans.
link |
Of course, the machine is going to have the strength
link |
that it has, the richer and longer memory,
link |
the more rigorous reasoning abilities,
link |
the deeper reasoning abilities,
link |
so it'll be an interesting
link |
complementary relationship
link |
between the human and the machine.
link |
Do you think that ultimately needs explainability,
link |
If you study, for example, Tesla autopilot a lot,
link |
I don't know if you've driven the vehicle
link |
or are aware of...
link |
and machine are working together there,
link |
and the human is responsible for their own life
link |
to monitor the system,
link |
and the system fails
link |
there's millions of those failures
link |
and so that's like a moment
link |
of interaction. Do you see...
link |
That's exactly right. That's a moment of interaction
link |
the machine has learned some stuff,
link |
somehow the failure is communicated,
link |
the human is now filling in
link |
the mistake, if you will, or maybe correcting
link |
or doing something that is more successful in that case,
link |
the computer takes that learning.
link |
that the collaboration between human
link |
that's sort of a primitive example
link |
and sort of a more...
link |
Another example is where the machine
link |
is literally talking to you and saying,
link |
look, I'm reading this thing.
link |
the next word might be this or that,
link |
but I don't really understand why.
link |
Can you help me understand the framework
link |
that supports this
link |
and then can kind of acquire that,
link |
take that and reason about it and reuse it?
link |
Try to understand something.
link |
a human student might do.
link |
I remember when my daughter was in first grade
link |
reading assignment about electricity
link |
somewhere in the text it says
link |
an electricity is produced by water
link |
flowing over turbines or something like that.
link |
And then there's a question that says,
link |
well, how is the electricity created?
link |
And so my daughter comes to me and says,
link |
I mean, I could create it and produce
link |
or kind of send it in this case.
link |
So I can go back to the text and I can copy
link |
by water flowing over turbines.
link |
But I have no idea what that means.
link |
Like, I don't know how to
link |
interpret water flowing over turbines
link |
and what electricity even is. I mean, I can get the
link |
answer right by matching the text.
link |
But I don't have any framework
link |
for understanding what this means at all.
link |
And framework, really,
link |
I mean, it's a set of not to be mathematical,
link |
ideas that you bring to the table
link |
and interpreting stuff and then you build those up
link |
You build them up with the expectation that
link |
there's a shared understanding of what
link |
Yeah, it's the social that us humans
link |
have a sense that humans on earth
link |
in general share a set of
link |
like how many frameworks are there?
link |
I mean, it depends on how
link |
you bound them, right? So in other words, how
link |
big or small like their individual scope.
link |
and there are new ones. I think
link |
the way I think about is kind of in a
link |
layer. I think that the architect is being layered
link |
in that there's a small
link |
that allow you the foundation to build
link |
frameworks. And then there may be
link |
many frameworks, but you have the ability
link |
to acquire them. And then you have the ability
link |
I mean, one of the most compelling ways of thinking
link |
about this is reasoning by analogy
link |
where I can say, oh, wow, I've learned something very
link |
I never heard of this. I never heard of this
link |
But if it's like basketball
link |
in the sense that the goals like the hoop
link |
and I have to get the ball in the hoop and I
link |
have guards and I have this and I have that
link |
like where does the
link |
where the similarities and where the differences
link |
and I have a foundation now for
link |
interpreting this new information.
link |
And then the different groups
link |
like the millennials will have a framework
link |
well, that you know, yeah, well
link |
Democrats and Republicans
link |
millennials, nobody wants that framework.
link |
Well, I mean, I think
link |
right. I mean, we're talking about political
link |
and social ways of interpreting the world around
link |
them. And I think these frameworks are
link |
still largely, largely similar. I think they
link |
differ in maybe what some fundamental
link |
assumptions and values are.
link |
Now, from a reasoning
link |
perspective, like the ability to process the
link |
framework, it might not be that
link |
different. The implications of different
link |
fundamental values or fundamental assumptions
link |
in those framework
link |
frameworks may reach very different conclusions.
link |
a social perspective, the conclusions
link |
may be very different. From an intelligence
link |
just followed where my assumptions took me.
link |
Yeah, the process itself looks similar,
link |
but that's a fascinating idea
link |
how a statement will be interpreted.
link |
and read the exact same statement and the conclusions
link |
that you derive 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
link |
perspective, one set of assumptions is going to lead
link |
you here, another set of assumptions is going to lead
link |
And in fact, you know, to help people
link |
reason and say, oh, I see where
link |
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. From my perspective
link |
NLP, there's this idea
link |
that there's one way to really understand a statement.
link |
there probably isn't. There's probably
link |
an infinite number of ways to understand a statement.
link |
Well, there's lots of different interpretations
link |
and so, you know, you
link |
and I can have very different experiences
link |
with the same text obviously
link |
if we're committed to understanding each other
link |
and that's the other important point like
link |
if we're committed to understanding each other
link |
we start decomposing
link |
and breaking down our interpretation
link |
towards more and more primitive components
link |
until we get to that
link |
point where we say, oh, I see why we disagree
link |
understand how fundamental that disagreement really is.
link |
a commitment to breaking down
link |
that interpretation in terms of that
link |
framework in a logical way.
link |
Otherwise, you know, and this is why
link |
I think of AIs as really
link |
complimenting and helping human intelligence
link |
to overcome some of its biases
link |
and its predisposition
link |
by more shallow reasoning
link |
in the sense that like we get over this idea
link |
you know, I'm right
link |
because I'm Republican 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
link |
break that argument down and say, wait a second,
link |
you know, what do you really
link |
think about this, right? So, essentially
link |
holding us accountable
link |
to doing more critical thinking.
link |
We're not just sitting and thinking about that as fast
link |
and that's, I love that.
link |
I think that's really empowering use of AI
link |
for the public discourse that's completely
link |
learn how to do it on social media.
link |
one of the greatest accomplishments
link |
in the history of AI
link |
competing in a game of Jeopardy against humans
link |
a critical part of that.
link |
Let's start at the very basics. What is the game of Jeopardy?
link |
for us humans, human versus human.
link |
The game of Jeopardy. Well,
link |
actually, it's the opposite.
link |
Well, no, but it's not, right?
link |
It's really not. It's really to get a question
link |
and answer but it's what we call a factoid
link |
question. So, this notion of like
link |
it really relates to some fact that
link |
a few people would argue
link |
whether the facts are true or not. In fact,
link |
what in Jeopardy kind of counts on the idea that
link |
have factual answers
link |
to first of all determine whether or not you know
link |
the answer which is sort of an interesting twist.
link |
So, first of all, understand the question.
link |
You have to understand the question. What is it
link |
asking and that's a good point because
link |
the questions are not
link |
asked directly, right? They're all like
link |
the way the questions are asked is
link |
nonlinear. It's like
link |
it's a little bit witty. It's a little bit
link |
playful sometimes.
link |
It's a little bit tricky.
link |
Yeah, they're asking
link |
exactly in numerous witty, tricky ways
link |
they're asking is not obvious. It takes
link |
inexperienced humans a while to go,
link |
what is it even asking? Right.
link |
And it's sort of an interesting realization that
link |
you have when somebody says, oh, what's the
link |
Jeopardy! is a question answering show and they say, oh,
link |
like I know a lot and then you read it and
link |
you're still trying to process the question
link |
and the champions have answered and moved on.
link |
There are three questions ahead
link |
by the time you figured out what the question
link |
even meant. So, there's definitely
link |
an ability there to just
link |
parse out what the question even is.
link |
So, that was certainly challenging. It's
link |
interesting historically though if you look back
link |
at the Jeopardy! games much earlier
link |
you know, 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
link |
more and more interesting and subtle
link |
and nuanced and humorous and
link |
witty over time which really
link |
required the human to kind of make
link |
the right connections in figuring out what the question
link |
was even asking. So, yeah,
link |
you have to figure out the questions even asking.
link |
determine whether or not you think you know the answer
link |
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
link |
if you know the answer. I think a lot of humans
link |
will assume they'll
link |
They'll process it very superficially. In other words,
link |
what's the topic? What are some
link |
keywords and just say do I know
link |
this area or not before they actually
link |
know the answer? Then they'll buzz
link |
in and think about it.
link |
It's interesting what humans do. Now some
link |
people who know all things like
link |
Ken Jennings or something or the more recent
link |
Big Jeopardy! player
link |
they'll just assume they know all the Jeopardy!
link |
and they'll just suppose that.
link |
Watson interestingly
link |
didn't even come close to knowing all of
link |
So, for example, we had this thing called Recall
link |
of all the Jeopardy! questions
link |
how many could we even
link |
find the right answer
link |
Can we come up with if we had
link |
a big body of knowledge in the order of several
link |
ways? I mean, from a web
link |
scale was actually very small.
link |
But from a book scale, I was talking about
link |
millions of books.
link |
Equally millions of books.
link |
Cyclopedias, dictionaries, books.
link |
It's still a ton of information.
link |
I think it was only
link |
85% was the answer anywhere to be found.
link |
ready down at that level just
link |
And so it was important
link |
quick sense of do you think you know the right
link |
answer to this question? So we had to compute that
link |
confidence as quickly as we
link |
possibly could. So in effect
link |
we had to answer it.
link |
spend some time essentially answering
link |
it. And then judging
link |
the confidence that we, you know, that
link |
our answer was right. And then deciding
link |
whether or not we were confident enough to buzz
link |
in. And that would depend on what else
link |
was going on in the game because it 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
link |
to lose, then you'll buzz in with less
link |
confidence. So that was a counter for
link |
the financial standings of the different
link |
competitors. Correct.
link |
How much of the game was left, how much time
link |
was left, where you were in the standing
link |
and things like that. What, how many
link |
hundreds of milliseconds that we're talking
link |
about here? Do you have a sense of
link |
what is... We targeted
link |
what's the target? So
link |
I mean we targeted answering
link |
in under three seconds
link |
So the decision to
link |
buzz in and then the actual
link |
answering, are those two different
link |
stages? Yeah, they were two different things. In fact, we
link |
had multiple stages, whereas like we
link |
would say let's estimate our confidence
link |
which was sort of a shallow
link |
answering process.
link |
And then ultimately
link |
decide to buzz in and then we may take another
link |
second or something
link |
to kind of go in there and
link |
and large we're saying like we can't play the game.
link |
compete if we can't
link |
on average answer these questions in around
link |
three seconds or less. So you
link |
stepped in, so there's this, there's these
link |
three humans playing a game
link |
and you stepped in with the idea that
link |
IBM Watson would be one of, replace
link |
one of the humans and compete against
link |
two. Can you tell the story
link |
on this game? Sure.
link |
Seems exceptionally difficult. Yeah.
link |
it was coming up I think to the 10 year anniversary
link |
wanted to do sort of another kind of
link |
really fun challenge, public
link |
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
link |
AI at IBM for some time.
link |
I had a team doing
link |
what's called open domain
link |
factoid question answering, which is
link |
we're not going to tell you what the questions are.
link |
We're not even going to tell you what they're about.
link |
Can you go off and get accurate answers
link |
to these questions?
link |
And it was an area of
link |
AI research that I was involved in.
link |
And so it was a big, it was a very
link |
specific passion of mine. Language understanding
link |
had always been a passion of mine.
link |
One sort of narrow slice on
link |
whether or not you could do anything with language was
link |
this notion of open domain and meaning I could
link |
ask anything about anything. Factoid
link |
meaning it essentially had an answer
link |
being able to do that accurately and quickly.
link |
So that was a research area that my team had already been
link |
in. And so completely independently
link |
executives were like, what are we going to do?
link |
What's the next cool thing to do?
link |
And Ken Jennings was on his winning
link |
streak. This was like whatever
link |
was 2004, I think, was on his
link |
winning streak. And someone
link |
thought, hey, that would be really cool
link |
if the computer can play Jeopardy.
link |
And so this was like
link |
in 2004, they were shopping this thing around
link |
was telling the research
link |
Like, this is crazy.
link |
And we had some pretty senior people in the field
link |
saying, no, this is crazy. And it would come across my
link |
desk and I was like, but that's kind of what
link |
I'm really interested in doing.
link |
such this prevailing sense of this is
link |
nuts, we're not going to risk IBM's reputation on
link |
this, we're just not doing it. And this happened in
link |
2004, it happened in 2005.
link |
it was coming around again
link |
and I was coming off of a,
link |
I was doing the open domain question answering
link |
stuff, but I was coming off a couple other
link |
projects. I had a lot more time
link |
to put into this and I argued
link |
that it could be done and I argued
link |
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
link |
yourself privately
link |
that this could be done?
link |
story how you tell stories to convince others.
link |
How confident were you? What was
link |
your estimation of the problem
link |
at that time? So I thought it was
link |
possible and a lot of people
link |
thought it was impossible. I thought it was possible.
link |
The reason why I thought it was possible is
link |
because I did some brief experimentation.
link |
I knew a lot about how we were approaching
link |
open domain factoid
link |
question answering. We've been doing it for some years.
link |
I looked at the Jafferty stuff.
link |
I said this is going to be hard
link |
for a lot of the points that
link |
we mentioned earlier. Hard to interpret the question.
link |
Hard to do it quickly enough. Hard
link |
to compute an accurate confidence. None of this stuff
link |
had been done well enough before.
link |
But a lot of the technologies we're building with the kinds
link |
of technologies that should work.
link |
But more to the point
link |
what was driving me was
link |
I was an IBM research.
link |
I was a senior leader in IBM research
link |
and this is the kind of stuff we were supposed
link |
We were supposed to take things
link |
and say this is an active research
link |
It's our obligation
link |
if we have the opportunity
link |
to push it to the limits. And if it doesn't
link |
work to understand more deeply
link |
why we can't do it.
link |
I was very committed to that notion
link |
saying folks this is what we do.
link |
not to do it. This is an active
link |
research area. We've been in this for years.
link |
Why wouldn't we take this grand challenge
link |
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
link |
Here's what we tried and here's how we failed.
link |
as a scientist from that perspective
link |
and then I also argued
link |
what we did a feasibility study.
link |
Why I thought it was hard but possible
link |
for us to take some sort of examples
link |
of where it succeeded
link |
where it failed, why it failed
link |
and sort of a high level architectural 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
link |
crazy enough to say yes
link |
because for several years at that point
link |
everyone had said no.
link |
I'm not willing to risk my reputation
link |
Clearly you did not have such fears.
link |
for what I understand
link |
it was performing very poorly
link |
in the beginning. So what were the
link |
initial approaches and why did they fail?
link |
Well, there were lots
link |
of hard aspects to it.
link |
One of the reasons why prior
link |
approaches that we had worked
link |
failed was because
link |
the questions were difficult
link |
to interpret. What are you even asking for?
link |
if the question was very direct
link |
even then it could be tricky
link |
often when it would name it
link |
very clearly you would know that.
link |
And if there was just a small
link |
set of them, in other words we're going to ask
link |
about these five types.
link |
It's going to be an answer
link |
and the answer will be
link |
a city in this state
link |
or a city in this country. The answer will be
link |
a person of this type
link |
like an actor or whatever it is.
link |
But turns out that in Jeopardy
link |
there were like tens of thousands of these things
link |
and it was a very, very long
link |
Meaning it just went on and on
link |
and so even if you focused on trying
link |
to encode the types
link |
at the very top like there's
link |
five that were the most let's say five of the most frequent
link |
you still cover a very small
link |
range of the data. So you couldn't take
link |
that approach of saying
link |
I'm just going to try to collect facts
link |
or ten types or twenty types or fifty types
link |
that was like one of the first things like
link |
what do you do about that and so we came up
link |
with an approach toward that
link |
and the approach looked promising
link |
and we continued to improve
link |
our ability to handle
link |
that problem throughout the project.
link |
The other issue was that
link |
right from the outside I said we're not
link |
going to, I committed
link |
to doing this in three to five years
link |
so we did it in four
link |
But one of the things that that putting that
link |
stake in the ground
link |
was I knew how hard the language
link |
understanding problem was. I said we're not going to
link |
actually understand
link |
language to solve this problem.
link |
We are not going to
link |
interpret the question
link |
and the domain of knowledge
link |
that the question refers to and reason over
link |
to that to answer these questions. Obviously
link |
we're not going to be doing that. At the same time
link |
wasn't good enough to
link |
confidently answer with this
link |
a single correct answer.
link |
First of all it's like brilliant. It's such a great
link |
mix of innovation in practical engineering
link |
three, three, four, eight.
link |
So you're not trying to solve the general
link |
NLU problem. You're saying let's
link |
solve this in any way possible.
link |
Yeah, no I was committed to
link |
saying look we're just solving the open
link |
domain question answering problem.
link |
We're using Jeopardy as a driver
link |
for that. Hard enough. Big benchmark
link |
We could just like whatever like just figure out what works
link |
because I want to be able to go back to the academic
link |
and scientific community and say here's what
link |
we tried. Here's what worked. Here's what
link |
didn't work. I don't want to go
link |
in and say oh I only have
link |
one technology. I have a hammer and I'm only going to use
link |
this. I'm going to do whatever it takes. I'm like
link |
let's think out of the box and do whatever it takes.
link |
there's another thing I believe. I believe
link |
that the fundamental
link |
NLP technologies and machine learning
link |
technologies would be
link |
would be adequate. And this was
link |
an issue of how do we enhance
link |
them? How do we integrate them?
link |
How do we advance them?
link |
So I had one researcher and came to me
link |
who had been working on question answering with me for a very
link |
who had said we're going to need
link |
Maxwell's equations for question answering.
link |
And I said if we need
link |
some fundamental formula that
link |
breaks new ground and how we understand
link |
language, we're screwed. We're
link |
not going to get there from here.
link |
my assumption is I'm not
link |
counting on some brand new
link |
invention. What I'm counting
link |
to take everything that has done before
link |
an architecture on how to integrate
link |
it well and then see where it
link |
breaks and make the necessary
link |
advances we need to make
link |
until this thing works. Yeah. Push it
link |
hard to see where it breaks and then patch
link |
it up. I mean, that's how people change the world.
link |
I mean, that's the Elon Musk approach with
link |
rockets, SpaceX, that's the
link |
Henry Ford and so on.
link |
And I happen to be and in this case
link |
I happen to be right, but like we didn't
link |
know. Right. But you kind of have to
link |
put a stake in terms of how you're going to run the project.
link |
So yeah, and backtracking to
link |
were to do, what's the brute force
link |
solution? What would
link |
you search over? So you have a question.
link |
How would you search
link |
the possible space of answers?
link |
Look, web searches come a long way even since
link |
time, like, you know, you first of
link |
all, I mean, there are a couple other constraints
link |
around the problems. Interesting. So
link |
you couldn't go out to the web. You
link |
couldn't search the Internet. In other
link |
words, the AI experiment was
link |
we want a self contained
link |
If the device is as big as a room, fine, it's as
link |
big as a room, but we want a self
link |
contained device, contained
link |
device. 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 shoebox
link |
if you will, or at least
link |
size of a few refrigerators, whatever it might be.
link |
you couldn't just get out there. You couldn't go off
link |
network, right, to kind of go. So
link |
there was that limitation. But then
link |
we did, but the basic thing was go
link |
The problem was even when we went and did a
link |
don't remember exactly the numbers, but someone
link |
in the order of 65% of the time,
link |
the answer would be somewhere
link |
in the top 10 or 20
link |
documents. So first of
link |
all, that's not even good enough to play Jeopardy.
link |
In other words, even
link |
if you could pull the, even if you could perfectly
link |
pull the answer out of the top
link |
20 documents, top 10 documents, whatever
link |
it was, which we didn't know how to do.
link |
But even if you could do that,
link |
you'd be, and you knew it was right.
link |
We had enough confidence in it, right?
link |
You'd have to pull out the right answer. You'd have to
link |
have confidence it was the right answer.
link |
And then you'd have to do that fast enough to now go buzz
link |
in. And you'd still only
link |
get 65% of them right, which doesn't even
link |
put you in the winner circle. Winner circle
link |
you have to be up over 70.
link |
And you have to do it really, and you have to do it really
link |
quickly. But now the problem is,
link |
well, even if I had
link |
somewhere in the top 10 documents, how do I figure out
link |
where in the top 10 documents that
link |
answer is? And how do I compute
link |
a confidence of all the possible candidates?
link |
So it's not like I go in knowing
link |
the right answer and have to pick it. I don't know
link |
the right answer. I have a bunch of documents
link |
somewhere in there's the right answer.
link |
How do I, as a machine, go out and figure out
link |
which one's right? And then how do I score
link |
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
link |
about it. If you could go to the web,
link |
do you think that problem is
link |
solvable? If you just pause on it?
link |
Just thinking even beyond
link |
Do you think the problem of reading text
link |
to find where the answer is?
link |
Well, we solved that in some
link |
definition of solved, given the jeopardy challenge.
link |
How did you do it for jeopardy? So how
link |
did you take a body
link |
of work in a particular topic
link |
and extract the key pieces of information?
link |
So what, so, now, forgetting
link |
about the huge volumes that are
link |
on the web, right? So now we have to figure out
link |
we did a lot of source research. In other words,
link |
what body of knowledge
link |
is going to be small enough but
link |
broad enough to answer
link |
jeopardy? And we ultimately did find
link |
the body of knowledge that did that. I mean, it included
link |
Wikipedia and a bunch of other stuff.
link |
So, like, encyclopedia type of stuff? I don't know if you can
link |
speak to it. Encyclopedia is different times of
link |
semantic resources,
link |
like WordNet and other types of semantic resources
link |
like that, as well as, like, some web
link |
crawls. In other words, where we went out
link |
and took that content
link |
and then expanded it based on producing
link |
statistical, you know, statistically
link |
producing seeds, using those
link |
seeds for other searches
link |
and then expanding that. So
link |
using these, like, expansion techniques
link |
we went out and had found enough content
link |
and were like, okay, this is good. And even
link |
up until the end, you know, we had
link |
a threat of research that was always trying to figure
link |
out what content could we
link |
efficiently include. I mean, there's a lot of popular
link |
content, like, what is the church lady?
link |
Well, I think it was one of the, like,
link |
where do you, I guess, that's probably
link |
in encyclopedias. So, I guess,
link |
take that stuff and we would go out and we would
link |
expand. In other words, we go find
link |
other content that wasn't in the core
link |
resources and expand it. You know,
link |
the amount of content that grew it by an order of
link |
magnitude, but still, again
link |
from a web scale perspective, this is a very
link |
small amount of content. It's very select.
link |
We then took all that content,
link |
we preanalyzed the crap out of it,
link |
parsed it, you know, broke it down
link |
into all those individual words, and then we did
link |
semantic, static and semantic
link |
parses on it, you know, had computer
link |
algorithms that annotated it, and
link |
we indexed that in
link |
a very rich and very fast
link |
So, we have a relatively huge amount of, you
link |
know, let's say the equivalent of, for the sake of
link |
argument, two to five million bucks, we've
link |
now analyzed all that, blowing up its size
link |
even more, because now we have all this metadata,
link |
and then we richly indexed all of
link |
that, and by the way,
link |
in a giant in memory cache.
link |
So, Watson did not go to disk.
link |
So, the infrastructure component
link |
there, if you could just speak to it, how tough
link |
it, I mean, I know
link |
2000, maybe this is
link |
kind of a long time ago.
link |
How hard is it to use multiple
link |
machines? How hard is
link |
the infrastructure component, the hardware component?
link |
So, we used IBM hardware.
link |
We had something like, I
link |
forget exactly, but 2,000, close
link |
completely connected. So, we had a switch
link |
where, you know, every CPU was connected
link |
to every other CPU. And they were sharing memory in some kind of way.
link |
Large, shared memory,
link |
right? And all this data
link |
was preanalyzed and
link |
put into a very fast
link |
indexing structure that
link |
in memory. And then
link |
we took that question
link |
the question. So, all the content
link |
was now preanalyzed.
link |
and tried to find a piece of content, it would
link |
come back with all the metadata that we had
link |
precomputed. How do you
link |
shove that question?
link |
How do you connect the big
link |
stuff, the big knowledge base
link |
of the metadata and that's indexed to
link |
the simple little witty
link |
confusing question?
link |
lies, you know, the Watson architecture.
link |
So, we would take the question, we would
link |
analyze the question. So, which
link |
means that we would parse it
link |
and interpret it a bunch of different ways. We'd try to
link |
figure out what is it asking about. So, we
link |
would come, we had
link |
multiple strategies to kind of determine
link |
what was it asking for.
link |
That might be represented as a simple
link |
string, a character string
link |
or something we would connect back to
link |
different semantic types that were from
link |
existing resources. So, anyway,
link |
the bottom line is we would do a bunch of analysis in the question.
link |
And question analysis
link |
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
link |
open source search engines, we modified
link |
them. We had a number of different
link |
search engines we would use that had
link |
different characteristics. We went in there
link |
and engineered and modified those
link |
search engines ultimately
link |
our question analysis, produce multiple
link |
queries based on different interpretations
link |
and fire out a whole bunch of searches
link |
And they would produce, they would come back
link |
So, these are passive search algorithms, they would
link |
come back with passages. And so, now
link |
let's say you had a thousand
link |
passages. Now, for each passage
link |
you parallelize again.
link |
So, you went out and you
link |
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 five thousand whatever passages.
link |
For each passage now, you'd go and
link |
figure out whether or not there was a candidate,
link |
we would 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 |
candidate answers, called candidate answers
link |
generators, the whole bunch of those.
link |
So, for every one of these components
link |
the team was constantly doing research
link |
coming up better ways to generate
link |
search queries from the questions, better ways
link |
to analyze the question, better ways to
link |
generate candidates. And speed, so better
link |
speed. Correct. So,
link |
right, speed and accuracy for the most
link |
part were separated.
link |
We handle that sort of in separate ways, like I
link |
was, purely on accuracy and
link |
to an accuracy, are we ultimately getting more
link |
questions and producing more accurate
link |
confidences. And then a whole other team
link |
that was constantly analyzing the workflow
link |
to find the bottlenecks. And then figuring
link |
out how to both parallelize and drive
link |
the algorithm speed. But anyway, so
link |
now think of it like you have this big fan
link |
out now, right? Because you have
link |
multiple queries, now you have
link |
thousands of candidate answers.
link |
For each candidate answer, you're going to score
link |
it. So, you're going to use
link |
all the data that built up, you're going to use
link |
the question analysis,
link |
you're going to use how the query was generated,
link |
you're going to use the passage itself
link |
and you're going to use the
link |
candidate answer that was generated
link |
and you're going to score that.
link |
a group of researchers coming up with scores.
link |
There are hundreds of different
link |
scores. So, now you're getting a fan
link |
out of it again from however many
link |
candidate answers you have
link |
to all the different scores.
link |
So, if you have a 200 different scores
link |
and you have 1,000 candidates, now you have
link |
And so, now you've got to figure out
link |
answers based on the scores that
link |
came back? And I want to rank
link |
them based on the likelihood that they're a correct answer
link |
to the question. So, every
link |
score was its own research project.
link |
What do you mean by score? So, is that the
link |
annotation process of basically
link |
a human being saying that this
link |
answer has quality?
link |
Think of it, if you want to think of it, what you're doing
link |
you know, if you want to think about
link |
what a human would be doing, a human would be looking at
link |
a possible answer.
link |
They'd be reading the
link |
you know, Emily Dickinson. They'd be
link |
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
link |
likely it is that Emily Dickinson
link |
given this evidence in this passage
link |
is the right answer to that question.
link |
Got it. So, that's the annotation
link |
task. That's the annotation
link |
task. That's the scoring task.
link |
So, but scoring implies 0 to 1
link |
kind of continuous. That's right. You give it a 0 to 1 score.
link |
Since it's not a binary. No.
link |
You give it a score.
link |
You give it a 0, yeah, exactly.
link |
So, humans do give different scores so
link |
you have to somehow normalize and all that kind of stuff
link |
that deal with all that complexity. Depends on
link |
what your strategy is. We both, we
link |
could be relative to. It could be
link |
we actually looked at the raw scores
link |
as well, standardized scores because humans
link |
are not involved in this.
link |
Humans are not involved. Sorry. So, I'm
link |
misunderstanding the process here. There's
link |
passages. Where is
link |
the ground truth coming from?
link |
Grand truth is only the answers to the questions.
link |
end to end. It's end to end.
link |
driving end to end performance. It was a very
link |
interesting. Wow. A very interesting
link |
approach and ultimately
link |
scientific research approach. Always driving
link |
now. That's not to say
link |
individual component performance
link |
was related in some way
link |
to end to end performance. Of course we would
link |
because people would have to
link |
build individual components. But
link |
ultimately to get your component integrated
link |
into the system, you have to show impact
link |
on end to end performance. Question
link |
answering performance. So, there's many very
link |
smart people working on this and they're basically
link |
their ideas as a component that should be part
link |
of the system. That's right. And
link |
they would do research on their component
link |
and they would say things like
link |
I'm going to improve
link |
this as a candidate generator.
link |
I'm going to improve this as a
link |
question score or as a passive
link |
score. I'm going to improve this
link |
or as a parser. And I
link |
can improve it by 2%
link |
on its component metric.
link |
Like a better parse or better
link |
candidate or a better type estimation
link |
whatever it is. And then I would say
link |
I need to understand how
link |
the improvement on that component metric
link |
is going to 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
link |
run AI project I've ever
link |
heard. That's awesome. Okay.
link |
What breakthrough would
link |
you say? Like I'm sure there's a lot
link |
of day to day breakthroughs but was there like a breakthrough
link |
that really helped improve performance?
link |
Like wait, were people
link |
Or is it just a gradual process? Well, I think
link |
it was a gradual process but
link |
one of the things that I think
link |
gave people confidence
link |
that we can get there was that
link |
as we follow this procedure of
link |
different ideas, build different components
link |
plug them into the architecture, run the system
link |
the error analysis, start off
link |
new research projects to improve things
link |
the very important idea
link |
that the individual
link |
did not have to deeply understand
link |
everything that was going on with every other component.
link |
we leveraged machine learning in a very
link |
So while individual components could be
link |
statistically driven machine learning components
link |
some of them were heuristic, some of them were
link |
machine learning components, the system has
link |
a whole combined all the scores
link |
using machine learning.
link |
because that way you can divide
link |
and conquer. So you can say
link |
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
link |
passage search or to passage selection
link |
But when we just plug it in
link |
and we had enough training
link |
data to say now we can
link |
train and figure out how do we
link |
weigh all the scores
link |
relative to each other
link |
based on predicting
link |
the outcome which is right or wrong on
link |
jeopardy. And we had enough training data
link |
to do that. So this
link |
enabled people to work
link |
independently and to let the machine
link |
learning do the integration.
link |
Beautiful. So the machine learning
link |
is doing the fusion and then it's a human
link |
orchestrated ensemble
link |
with different approaches.
link |
Still impressive that you were able to get it
link |
done in a few years.
link |
That's not obvious to me
link |
that it's doable if I just put myself
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?
link |
It's looking back at those days.
link |
and my team's commitment
link |
to be true to the science.
link |
It's beautiful because there's so much pressure
link |
because it is a public event.
link |
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
link |
It was a success for our goal.
link |
build the most advanced
link |
open domain question answering system.
link |
to the old problems that we used to try
link |
to solve and we did
link |
dramatically better on all of them
link |
as well as we beat jeopardy.
link |
So we won the jeopardy.
link |
So it was a success.
link |
world would not understand it as a success
link |
it came down to only one game and I knew
link |
statistically speaking this can be a huge
link |
technical success and we could still lose that
link |
one game and that's a whole other theme
link |
But it was a success.
link |
It was not a success
link |
in natural language understanding
link |
but that was not the goal.
link |
I understand what you're saying
link |
in terms of the science
link |
but I would argue that
link |
the inspiration of it
link |
not a success in terms of solving
link |
natural language understanding.
link |
It was a success of being an inspiration
link |
to future challenges.
link |
To drive future efforts.
link |
What's the difference between how human being
link |
compete in jeopardy
link |
and how Watson does it.
link |
That's important in terms of intelligence.
link |
That actually came up very early
link |
on in the project also.
link |
In fact I had people who wanted to be on the project
link |
early on who approached me
link |
once I committed to do it
link |
I wanted to think about
link |
how humans do it and they were
link |
from a cognition perspective
link |
like human cognition and how that should play.
link |
take them on the project because
link |
another assumption
link |
or another state I put in the ground
link |
was 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 |
and in building an AI that understands
link |
how it needs to ultimately communicate
link |
with humans, I very much care.
link |
In fact as an AI scientist
link |
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 going to get distracted
link |
say like if I'm going to get this done
link |
I'm going to chart this path and this path says
link |
we're going to engineer a machine
link |
that's going to get this thing done
link |
search and NLP can do
link |
we have to build on that foundation
link |
if I come in and take
link |
a different approach and start wondering about
link |
how the human mind might or might not do this
link |
I'm not going to get there from here
link |
in the time frame.
link |
I think that's a great way to lead the team.
link |
there's done and there's one
link |
when you look back, analyze
link |
what's the difference actually.
link |
So I was a little bit surprised actually
link |
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 have been closer to the way humans
link |
answer questions than I might have imagined
link |
Because humans are probably in the game of Jeopardy
link |
at the level of Ken Jennings
link |
cheating their way
link |
to winning, right?
link |
Well, they're doing shallow analysis.
link |
They're doing the fastest possible.
link |
They're doing shallow analysis.
link |
very quickly analyzing the question
link |
and coming up with some
link |
key vectors or cues if you will
link |
and they're taking those cues
link |
and very quickly going through
link |
their library of stuff
link |
not deeply reasoning about what's going on
link |
what we call these scores
link |
what's kind of score
link |
in a very shallow way
link |
and then say, oh, boom, that's what it is.
link |
And so it's interesting
link |
as we reflected on that
link |
we may be doing something that's not too far off
link |
from the way humans do it
link |
didn't approach it by saying,
link |
you know, 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
link |
we're trying to do something that
link |
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
link |
with this shared understanding.
link |
So how they think about things
link |
answers, how they build explanations
link |
becomes a very important question to consider.
link |
So what's the difference
link |
constructed question answering
link |
something that requires understanding
link |
for shared communication with humans and machines?
link |
Yeah, well, this goes back
link |
to the interpretation
link |
of what we were talking about before.
link |
Jeopardy, the system is not
link |
trying to interpret the question
link |
and it's not interpreting the content
link |
that's reusing with regard to any particular
link |
framework. I mean it is
link |
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
link |
framework in some sense it has that
link |
but when you get into the richer
link |
semantic frameworks,
link |
what do people, how do they think, what motivates them,
link |
what are the events that are
link |
occurring and why they're occurring and what causes
link |
what else to happen and
link |
where are things in time and space
link |
and when you start to think about
link |
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
link |
essential challenges
link |
communication, free flowing dialogue
link |
versus question answering
link |
even with a framework, with the
link |
interpretation dialogue?
link |
free flowing dialogue
link |
fundamentally more difficult
link |
than question answering even with
link |
interpretation? So dialogue
link |
is important in a number of different ways.
link |
I mean it's a challenge. So first of all
link |
when I think about the machine
link |
that understands language
link |
and ultimately can reason
link |
in an objective way
link |
information that it perceives through language
link |
or other means and connect it back
link |
to these frameworks, reason
link |
and explain itself
link |
that system ultimately needs
link |
to be able to talk to humans, right?
link |
It needs to be able to interact with humans
link |
so in some sense it needs to dialogue.
link |
That doesn't mean that
link |
people talk about dialogue and they think
link |
talk to each other
link |
in a casual conversation
link |
and you could mimic casual conversations.
link |
We're not trying to mimic casual
link |
conversations. We're really trying
link |
to produce a machine
link |
whose goal is to help you
link |
think and help you reason
link |
about your answers and explain why.
link |
So instead of like talking to your
link |
friend down the street about having
link |
a small talk conversation with your friend
link |
down the street, this is more about
link |
like you would be communicating to the computer
link |
on Star Trek where
link |
what do you want to think about?
link |
What do you want to reason about? I'm going to tell you the information I have
link |
and I'm going to have to summarize it. I'm going to ask you questions
link |
and you're going to answer those questions.
link |
I'm going to go back and forth with you.
link |
I'm going to figure out what your mental model is.
link |
I'm going to now relate that
link |
to the information I have and present it to you
link |
in a way that you can understand
link |
and then we can ask follow up questions.
link |
So it's that type of dialogue
link |
that you want to construct.
link |
It's more structured.
link |
It's more goal oriented
link |
In other words, it can't
link |
it has to be engaging and fluid.
link |
It has to be productive
link |
and not distracting.
link |
So there has to be a model
link |
of, in other words, the machine has to have
link |
a model of how humans
link |
think through things
link |
So basically a productive, rich
link |
unlike this podcast
link |
what I'd like to think
link |
it's more similar to this podcast.
link |
I was just joking.
link |
I'll ask you about humor as well, actually.
link |
what's the hardest part of that
link |
because it seems we're quite far away
link |
from that still to be
link |
able to. So one is having a shared
link |
I think a lot of the stuff you said with frameworks
link |
is quite brilliant.
link |
creating a smooth discourse
link |
Which aspects of this whole
link |
problem that you just specified
link |
of having a productive
link |
conversation is the hardest
link |
aspect of it you can comment on because it's so shrouded
link |
So I think to do this you kind of have to
link |
be creative in the following
link |
So how to do this is purely a machine
link |
learning approach. And someone said
link |
learn how to have a
link |
good, fluent, structured
link |
knowledge acquisition conversation.
link |
and say okay I have to collect a bunch of data
link |
of people doing that. People reasoning
link |
having a good structured
link |
conversation that both acquires
link |
knowledge efficiently as well as
link |
produces answers and explanations as part of
link |
to collect the data
link |
because I don't know how much data
link |
there's a humorous comment around the lack of
link |
rational discourse but also
link |
even if it's out there
link |
say it was out there how do you
link |
successful example.
link |
So I think any problem like this
link |
where you don't have
link |
enough data to represent
link |
the phenomenon you want to learn.
link |
In other words, 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
link |
sort of thing to do. What recently came
link |
out at IBM was the debate or project
link |
interest thing. Because now you do
link |
have these structured dialogues, these debate
link |
things where they did
link |
use machine learning techniques to
link |
generate these debates.
link |
Dialogues are a little bit
link |
tougher in my opinion than
link |
generating a structured argument
link |
where you have lots of other structured
link |
arguments like this. You could potentially annotate
link |
that data and you could say this is a good response
link |
a bad response in a particular domain.
link |
I have to be responsive and I have to be
link |
with regard to what is the human saying
link |
so I'm goal oriented
link |
and saying I want to solve the problem
link |
I want to acquire the knowledge necessary. But I also
link |
have to be opportunistic and responsive
link |
to what the human is saying.
link |
So I think that it's not clear
link |
that we could just train on the body of data
link |
to do this. But we
link |
could bootstrap it. In other words we can be creative
link |
and we could say what do we think
link |
what do we think the structure of a good
link |
dialogue is that does this well
link |
and we can start to create that
link |
that more programmatically
link |
at least to get this process started
link |
create a tool that now engages humans effectively
link |
I could start both
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 |
But I have to understand what features to even learn over
link |
so I have to bootstrap the process
link |
a little bit first.
link |
And that's a creative design task
link |
that I could then use
link |
into a more automatic learning task.
link |
Some creativity in
link |
bootstrapping. What elements
link |
of a conversation do you think
link |
you would like to see?
link |
So one of the benchmarks
link |
That seems to be one of the hardest
link |
and to me the biggest contrast
link |
So one of the greatest
link |
comedy sketches of all time
link |
is the SNL celebrity
link |
with Alex Rebecca and
link |
John Connery and Bert Reynolds
link |
with John Connery commentating
link |
on Alex Rebecca's mother a lot.
link |
And I think all of them
link |
are in the negative point that's why.
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
link |
about humor in this whole interaction
link |
that's productive?
link |
Or even just whatever
link |
what humor represents to me is
link |
idea that you're saying about framework
link |
because humor only exists within a particular
link |
human framework. So what do you think
link |
about humor? What do you think about things
link |
like humor 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
link |
this way but I think that
link |
a little bit about with this
link |
with puns in Jeopardy.
link |
We literally sat down and said
link |
And it's like word play
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. You could just say
link |
what do people laugh at.
link |
And if you have enough data to represent
link |
the phenomenon, you might be able to
link |
weigh the features and figure out
link |
what humans find funny and what they don't find funny.
link |
The machine might not be able to explain
link |
why you might want to get funny
link |
unless we sit back and think about that
link |
more formally. I think, again,
link |
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 of a combination of
link |
us reflecting and being creative about
link |
how things are structured, how to formalize them
link |
and then taking advantage of large data
link |
and doing learning and figuring out how to combine
link |
these two approaches.
link |
I think there's another aspect to humor though
link |
which goes to the idea that
link |
I feel like I can relate to the person
link |
telling the story.
link |
And I think that's
link |
an interesting theme
link |
in the whole AI theme
link |
feel differently when I know it's a robot?
link |
that the robot is not conscious the way I'm
link |
conscious, when I imagine
link |
the robot does not actually have the experiences
link |
that I experience, do I find
link |
Or do, because it's not as related,
link |
that the person is relating it to it the way I relate to it.
link |
the arts and in entertainment where
link |
like, you know, sometimes you have
link |
savants who are remarkable at a thing
link |
whether it's sculpture, it's music or whatever.
link |
But the people who get the most attention
link |
are the people who can
link |
a similar emotional response
link |
who can get you to
link |
So the way they are, in other words, who can
link |
basically make the connection
link |
from the artifact, from the music
link |
or the painting of the sculpture
link |
to the emotion and get you
link |
to share that emotion with them.
link |
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 |
to the artifact. And then you feel like,
link |
oh wow, I can relate to that person.
link |
I can connect to that person.
link |
So I think humor has that
link |
you can connect to that person,
link |
person being the critical thing.
link |
able to anthropomorphize objects
link |
and AI systems pretty well.
link |
So we're almost looking
link |
to make them human.
link |
Maybe from your experience with Watson,
link |
maybe you can comment on
link |
did you consider that as part,
link |
well, obviously the problem of Jeopardy
link |
can require anthropomorphization.
link |
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 from
link |
the actual scientific task.
link |
But you're absolutely right.
link |
Humans do anthropomorphize
link |
and without necessarily
link |
a lot of work. I mean, you just put some eyes
link |
in 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
link |
cannot be mimicked.
link |
I think that connection can be mimicked
link |
that emotional response.
link |
I just wonder though
link |
what's really going on,
link |
the machine is not conscious,
link |
not having the same richness
link |
of emotional reactions and understanding
link |
that doesn't really share the understanding
link |
but essentially just moving its eyebrow
link |
or drooping its eyes or making them big
link |
or whatever it's doing. Just getting the emotional
link |
response. Will you still feel it?
link |
Interesting. I think you probably would for a while.
link |
And then when it becomes
link |
more important that there's a deeper
link |
shared understanding, it may run flat.
link |
But I don't know. I'm...
link |
I'm pretty confident that
link |
the majority of the world
link |
even if you tell them how it works...
link |
It will not matter.
link |
Especially if the machine
link |
that she is conscious.
link |
That's very possible.
link |
So you, the scientist that made the machine
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 |
You're deep into the science fiction genre now.
link |
I don't think 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
link |
for the next few decades.
link |
It's a very interesting
link |
element of intelligence.
link |
So what do you think...
link |
We talked about social constructs of intelligence
link |
in the way humans kind of
link |
interpret information.
link |
What do you think is a good test of intelligence
link |
So there's the Alan Turing
link |
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
link |
crossing a kind of
link |
that gives me pause
link |
for AR are generally high.
link |
What does high look like, by the way?
link |
So not the threshold.
link |
Test is a threshold.
link |
What do you think is the destination?
link |
What do you think is the ceiling?
link |
machines will, in many measures,
link |
will be better than us,
link |
will become more effective.
link |
In other words, better predictors
link |
about a lot of things
link |
than ultimately we can do.
link |
I think where they're going to struggle
link |
is what we've talked about before,
link |
relating to communicating
link |
with and understanding humans
link |
So I think that's a key point.
link |
You can create the super parrot.
link |
What I mean by the super parrot is
link |
given enough data, a machine can mimic
link |
your emotional response, can even
link |
generate language that will sound smart
link |
and what someone else might say
link |
under similar circumstances.
link |
I would just pause on that.
link |
That's the super parrot, right?
link |
So given similar circumstances,
link |
changes its tone of voice in similar ways,
link |
produces strings of language
link |
that would similar that a human might say,
link |
being able to produce a
link |
logical interpretation or understanding
link |
ultimately satisfy
link |
a critical interrogation
link |
or a critical understanding.
link |
I think you just described me
link |
So I think philosophically
link |
speaking, you could argue
link |
that that's all we're doing as human beings
link |
to a worse extent.
link |
It's very possible humans
link |
do behave that way too.
link |
So upon deeper probing
link |
and deeper interrogation, you may find out
link |
that there isn't a shared understanding
link |
because I think humans do both.
link |
Humans are statistical language model machines
link |
and they are capable reasoners.
link |
and you don't know which is going on.
link |
I think it's an interesting
link |
we talked earlier about like where we are
link |
in our social and political landscape.
link |
Can you distinguish
link |
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 |
With that interrogative or probing dialogue?
link |
So it's interesting because humans are
link |
really good at in their own mind
link |
justifying or explaining what they hear
link |
because they project
link |
their understanding onto yours.
link |
So you could say you could put together
link |
and someone will sit there and interpret it
link |
in a way that's extremely bias
link |
to the way they want to interpret it.
link |
They want to assume you're an idiot 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 |
AI gets better and better mimic
link |
and we create the super parrots
link |
just as we are challenged with humans.
link |
Do you really know what you're talking about?
link |
a meaningful interpretation
link |
framework that you could reason over
link |
your answers, justify
link |
your predictions and your beliefs
link |
why you think they make sense?
link |
Can you convince me what the implications are?
link |
can you reason intelligently
link |
and make me believe
link |
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 |
there should be a large group of people
link |
with a certain standard of intelligence
link |
by this particular
link |
then there will pass.
link |
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
link |
with regard to a rigorous standard
link |
of objective logic and reason
link |
you still have a problem
link |
like masses of people can be
link |
to turn their brains off.
link |
By the way, I have nothing against the one of you.
link |
a part of one of the great
link |
benchmarks, challenges
link |
What do you think about
link |
AlphaZero, OpenAI5,
link |
AlphaStar accomplishments on video games
link |
recently, which are also
link |
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 |
As one of those other things nobody thought
link |
solving Go was going to be easy
link |
hard for humans to learn, hard for humans to excel at
link |
another measure of intelligence.
link |
I mean, it's very interesting
link |
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. I think that was great.
link |
the result speaks for itself.
link |
I think it makes us think about
link |
again, what's intelligence?
link |
What aspects of intelligence are important?
link |
Can the Go machine help
link |
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
link |
very simple terms, it found the function.
link |
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 like whatever it might be?
link |
One of the interesting ideas
link |
of that system is that it plays against itself.
link |
But there's no human in the loop there.
link |
Like you're saying, it could have
link |
an alien intelligence.
link |
Imagine you're sentencing, you're judging
link |
you're sentencing people.
link |
Or you're setting policy.
link |
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 |
an interesting dilemma
link |
for the applications of
link |
AI. Do we hold AI to
link |
accountability that says,
link |
when you take responsibility
link |
decision. In other words, can you
link |
explain why you would do the thing?
link |
Will you get up and speak
link |
to other humans and convince them that this was
link |
a smart decision? Is the AI
link |
enabling you to do that?
link |
Can you get behind the logic that was
link |
Sorry to linger 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
link |
Okay, there's two possible worlds
link |
that we have in the future.
link |
One is where AI systems
link |
do medical diagnosis or
link |
things like that, or drive a car
link |
explaining to you why
link |
it fails when it does.
link |
That's one possible world
link |
we're okay with it. Or the other
link |
where we are not okay with it and
link |
we really hold back the technology
link |
from getting too good before it gets
link |
able to explain which of those worlds
link |
are more likely, do you think, and
link |
which are concerning to you or not?
link |
I think the reality is it's going to be a mix.
link |
I'm not sure I have a problem with
link |
that. I think there are tasks that are perfectly
link |
machines show a certain level
link |
of performance and that level of performance
link |
is already better than humans.
link |
So, for example, I don't know that
link |
I take driverless cars.
link |
If driverless cars learn how to be more
link |
effective drivers than humans but can't
link |
explain what they're doing, but
link |
bottom line, statistically speaking,
link |
they're 10 times safer
link |
than humans. I don't know that
link |
I think when we have these edge cases
link |
when something bad happens and we want
link |
to decide who's liable for that thing
link |
and who made that mistake and what do we do
link |
about that? And I think in those edge cases
link |
are interesting cases.
link |
And now do we go to designers of the AI
link |
and the AI says, I don't know, that's what it learned
link |
to do and it says, well, you didn't train it
link |
properly. You know, you were
link |
negligent in the training data that you gave
link |
that machine. Like, how do we drive down
link |
the real level? So, I think those are
link |
interesting questions.
link |
So, the optimization problem there, sorry,
link |
is to create an ass system that's able to
link |
explain the lawyers away.
link |
is going to be interesting. I mean, I think this is where
link |
technology and social discourse are going to get
link |
like deeply intertwined
link |
in how we start thinking about
link |
problems, decisions and problems like that.
link |
I think in other cases, it becomes more obvious
link |
why did you decide to give that person
link |
Again, policy decisions or
link |
why did you pick that treatment? Like that treatment
link |
ended up killing that guy. Like, why was that
link |
a reasonable choice to make?
link |
and people are going to demand
link |
explanations. Now, there's a reality
link |
And the reality is that it's not,
link |
I'm not sure humans are making
link |
reasonable choices when they do these
link |
things. They are using
link |
statistical hunches,
link |
systematically using
link |
statistical averages to make cause.
link |
And this is what happened. My dad, if you saw
link |
the talk I gave about that, but
link |
you know, I mean, they decided
link |
that my father was brain dead.
link |
He had went into cardiac arrest
link |
and it took a long time
link |
for the ambulance to get there and he wasn't not
link |
resuscitated right away and so forth. And they came
link |
and they told me he was brain dead and why
link |
was he brain dead? Because essentially they gave me
link |
a purely statistical argument
link |
under these conditions with these four features
link |
98% chance he's brain dead.
link |
I said, but can you just tell me
link |
not inductively, but deductively
link |
go there and tell me his brain's not functioning
link |
is the way for you to do that. And
link |
in response was, no, this is how we make this decision.
link |
I said, this is inadequate for me.
link |
I understand the statistics and
link |
I don't know how, you know,
link |
there's a 2% chance he's still alive. I just don't
link |
know the specifics. I need the specifics
link |
and I want the deductive logical argument
link |
about why you actually know he's brain dead.
link |
So I wouldn't sign that do not resuscitate.
link |
And I don't know, it was like
link |
they went through lots of procedures, a big long
link |
story, but the bottom was a fascinating
link |
story, by the way, but how I reasoned
link |
and how the doctors reasoned through this whole process.
link |
But I don't know, somewhere around
link |
24 hours later or something, he was sitting up
link |
in bed with zero brain damage.
link |
What lessons do you draw from
link |
story, that experience?
link |
That the data that
link |
the data that's being used to make statistical
link |
inferences doesn't adequately
link |
reflect the phenomenon. So in other words,
link |
you're getting shit wrong, sorry,
link |
you're getting stuff wrong
link |
because your model
link |
is not robust enough
link |
not using statistical
link |
inferences and statistical averages in certain cases
link |
when you know the model is insufficient
link |
and that you should be reasoning about the
link |
specific case more logically
link |
and more deductively
link |
and hold yourself responsible and accountable
link |
AI has a role to say the exact
link |
thing that you just said, which is
link |
perhaps this is a case
link |
you should think for yourself.
link |
You should reason deductively.
link |
So it's hard because
link |
You'd have to go back and you'd have to have enough
link |
data to essentially say, and this goes back
link |
to the case of how do we decide
link |
whether AI is good enough to do a particular
link |
And regardless of whether or not
link |
it produces an explanation.
link |
what standard do we hold
link |
broadly, for example,
link |
as my father, as a medical
link |
the medical system ultimately
link |
helped him a lot throughout his life.
link |
Without it, he probably
link |
would have died much sooner.
link |
So overall sort of
link |
in sort of a net kind of way.
link |
Actually, I don't know
link |
But it may be not in that particular case, but overall
link |
the medical system overall
link |
does more good than bad.
link |
The medical system overall was doing
link |
more good than bad. Now there's another argument
link |
that suggests that there wasn't a case,
link |
but for the sake of argument, let's say like
link |
that's let's say a net positive.
link |
And I think you have to sit there and take that
link |
into consideration. Now you
link |
look at a particular use case, like for example
link |
making this decision.
link |
Have you done enough studies
link |
how good that prediction really is?
link |
And have you done enough studies to compare
link |
it to say, well, what if we
link |
let's get the evidence, let's do
link |
the deductive thing and not use statistics here.
link |
How often would that have done better?
link |
the studies to know how good the AI actually
link |
is. And it's complicated
link |
because it depends how fast you have to make the decision.
link |
So if you have to make the decision super fast,
link |
do you have no choice?
link |
Right. If you have
link |
more time, right, but if you're ready
link |
and this is a lot of the argument that I had was a doctor,
link |
I said, what's he going to do if you do it?
link |
What's going to happen to him in that room
link |
if you do it my way?
link |
Well, he's going to 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
link |
the case with your father,
link |
but also when things like race and gender
link |
start coming into play, when
link |
when judgments are
link |
made based on things
link |
complicated in our society, at least
link |
in discourse. And it starts
link |
I'm safe to say that most
link |
of the violent crime is committed by males.
link |
So if you discriminate based
link |
with the male versus female
link |
saying that if it's a male, more likely
link |
to commit the crime. This is one of my
link |
and optimistic views
link |
the study of artificial intelligence,
link |
the process of thinking and reasoning,
link |
logically and statistically
link |
and how to combine them is so important
link |
for the discourse today because it's causing
link |
regardless of what
link |
what state AI devices
link |
dialogue to happen. This is one of the most
link |
important dialogues that
link |
in my view, the human species can have
link |
right now, which is
link |
how to think well.
link |
well, how to understand our
link |
and what to do about them.
link |
That has got to be one of the most important
link |
a species can be doing honestly.
link |
an incredibly complex society.
link |
We've created amazing
link |
abilities to amplify
link |
noise faster than we can
link |
We are challenged.
link |
We are deeply, deeply challenged.
link |
big segments of the population getting hit with
link |
enormous amounts of information.
link |
Do they know how to do critical thinking?
link |
Do they know how to objectively
link |
reason? Do they understand
link |
what they are doing, never mind
link |
what their AI is doing?
link |
This is such an important dialogue
link |
we are fundamentally
link |
thinking can be and easily becomes
link |
fundamentally bias.
link |
And there are statistics
link |
and we shouldn't blind us. We shouldn't
link |
discard statistical inference.
link |
But we should understand the nature
link |
of statistical inference.
link |
to reject statistical
link |
deciding on the individual.
link |
reject that choice.
link |
So even if the statistics said
link |
if the statistics said
link |
males are more likely to have
link |
to be violent criminals, we still take
link |
each person as an individual
link |
based on the logic
link |
and the knowledge of that
link |
We purposefully and intentionally
link |
the statistical inference.
link |
at a respect for the individual.
link |
For the individual. And that requires reasoning
link |
Looking forward, what grand challenges
link |
would you like to see in the future?
link |
the Jeopardy Challenge
link |
captivated the world.
link |
AlphaGo, AlphaZero
link |
captivated the world. DBLU, certainly beating
link |
Gary's bitterness aside
link |
captivated the world.
link |
do you have ideas for next grand challenges for
link |
future challenges of that?
link |
Look, I mean, I think there are lots of
link |
really great ideas for grand challenges.
link |
focused on one right now which is
link |
demonstrate that they understand, that they could
link |
read and understand
link |
that they can acquire these frameworks
link |
reason and communicate with humans.
link |
So it is kind of like the Turing task
link |
but it's a little bit more demanding
link |
than the Turing task. It's not enough
link |
that you might be human
link |
pair it a conversation.
link |
I think the standard
link |
is a little bit higher.
link |
the standard is higher
link |
and I think one of the challenges
link |
of devising this grand challenge
link |
what intelligence is.
link |
We're not sure how to determine
link |
whether or not two people
link |
actually understand each other
link |
and in what depth they understand it.
link |
You know, to what depth they understand
link |
the challenge becomes something along the lines
link |
a shared understanding.
link |
So if I were to probe
link |
and probe and you probe me,
link |
can machines really
link |
act like thought partners
link |
where they can satisfy me
link |
our understanding is shared enough
link |
that we can collaborate
link |
and produce answers together
link |
and that they can help me explain
link |
and justify those answers.
link |
So maybe here's an idea.
link |
We'll have AI system
link |
We can convince the voters
link |
that they should vote.
link |
So like, I guess, what does
link |
winning look like?
link |
Again, that's why I think this is such a challenge
link |
because we go back to
link |
the emotional persuasion.
link |
We go back to, you know,
link |
now we're checking off
link |
of human cognition
link |
that is in many ways
link |
weak or flawed, right?
link |
We're so easily manipulated.
link |
for often the wrong reasons,
link |
right? Not the reasons
link |
that ultimately mattered us,
link |
but the reasons that can easily persuade us.
link |
I think we can be persuaded
link |
to believe one thing or another
link |
for reasons that ultimately
link |
don't serve us well in the long term.
link |
And a good benchmark
link |
should not play with those
link |
of emotional manipulation.
link |
I don't think so. And I think that's where
link |
we set the higher standard
link |
This goes back to rationality
link |
and it goes back to objective thinking.
link |
Can you acquire information
link |
and produce reasoned arguments
link |
and to those reasons, arguments pass
link |
a certain amount of muster?
link |
acquire new knowledge?
link |
I have acquired new knowledge.
link |
Can you identify where it's
link |
consistent or contradictory
link |
with other things you've learned?
link |
And can you explain that to me and get me to understand that?
link |
So I think another way
link |
to think about it perhaps
link |
is can a machine teach you?
link |
Can it help you understand
link |
something that you didn't really understand before?
link |
it's taking you through?
link |
again, it's almost like, can it teach
link |
you? Can it help you learn?
link |
in an arbitrary space
link |
so it can open those domain space.
link |
So can you tell the machine, again, this
link |
borrows from some science fictions, but
link |
can you go off and learn about this
link |
topic that I'd like to understand
link |
better and then work with
link |
me to help me understand it?
link |
That's quite brilliant.
link |
that passes that kind of test,
link |
do you think it would need to
link |
self awareness or even consciousness?
link |
What do you think about
link |
consciousness and the importance of it?
link |
Maybe in relation to
link |
having a presence,
link |
Do you think that's important?
link |
People used to ask me if Watson was conscious
link |
and I used to say,
link |
are you conscious of what exactly?
link |
I think maybe it depends
link |
on what you're conscious of.
link |
easy for it to answer questions about
link |
it would be trivial to program it.
link |
So to answer questions about whether or not
link |
it was playing jeopardy. I mean, it could
link |
certainly answer questions that would imply
link |
that it was aware of things. Exactly.
link |
What does it mean to be aware and what does it
link |
mean to consciousness? It's sort of interesting.
link |
I mean, I think that we differ from one
link |
another based on what we're conscious
link |
We're conscious of consciousness in there.
link |
Well, there's just areas.
link |
It's not just degrees.
link |
What are you aware of?
link |
But nevertheless, there's a very subjective element
link |
to our experience.
link |
Let me even not talk about
link |
consciousness. Let me talk about
link |
to me, really interesting topic of mortality.
link |
Fear or mortality.
link |
did not have a fear of death.
link |
Wasn't conscious of death.
link |
So there's an element of finiteness
link |
to our existence that I think
link |
like we 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
link |
And that seems to add a color
link |
or motivations in a way that
link |
seems to be fundamentally important
link |
Or at least the kind of human intelligence.
link |
Well, I think for generating goals.
link |
Again, I think you could have
link |
an intelligence capability
link |
and a capability to learn,
link |
But I think without
link |
I mean, again, you get a
link |
fear, but essentially without the goal
link |
You think you can just encode that
link |
without having to really.
link |
I think you can create a robot now
link |
and you could say, you know,
link |
plug it in and say,
link |
protect your power source, you know,
link |
and give it some capabilities and we'll sit there
link |
and operate to try to protect this power source
link |
So I don't know that that's
link |
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 |
because I think as we touched on
link |
intelligence is kind of like a social construct.
link |
fact that a robot will be protecting
link |
and grounding to its intelligence
link |
us being able to respect that.
link |
I mean, ultimately, it boils down to us
link |
acknowledging 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
link |
trivial that was, you would associate
link |
that with being intelligence.
link |
I mean, I literally put in a statement of code
link |
that says, you know, you have the following actions
link |
you can take, you give it a bunch of actions
link |
like, maybe you mount the laser
link |
going on or you may
link |
or you give the ability to scream
link |
or screech or whatever.
link |
And you know, and you say, you know,
link |
you're power source threatened
link |
and you could program that in
link |
and, you know, you're going to
link |
you're going to take these actions to protect it.
link |
You know, you could teach it
link |
train it on a bunch of things.
link |
And now you can look at that and you can say,
link |
well, you know, that's intelligence
link |
because it's protecting its power source, maybe,
link |
but that's again, this human bias
link |
that says, the thing I identify
link |
my intelligence and my conscious
link |
so fundamentally with the desire
link |
or at least the behavior is associated
link |
with the desire to survive
link |
that if I see another thing doing
link |
that, I'm going to assume
link |
year will society have
link |
something that would
link |
that you would be comfortable
link |
calling an artificial general intelligence system?
link |
Well, what's your intuition?
link |
Nobody can predict the future.
link |
Certainly not the next few months
link |
or 20 years away, but
link |
what's your intuition? How far away are we?
link |
It's hard to make these predictions.
link |
I would be, you know, I would be guessing
link |
and there's so many different variables
link |
including just how much we want to invest
link |
in it and how important it, you know,
link |
and how important we think it is
link |
what kind of investment we're willing to make
link |
in it, what kind of talent
link |
we end up bringing to the table, all, you know,
link |
the incentive structure, all these things.
link |
So I think it is possible
link |
to do this sort of thing.
link |
I think trying to sort of
link |
of the variables and things like that.
link |
Is it a 10 year thing? Is it 23?
link |
It's 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
link |
to investing and incentivizing this type of
link |
work, this type of work.
link |
And it's sort of interesting.
link |
Like, I don't think it's obvious
link |
how incentivized we are.
link |
I think from a task
link |
you know, if we see business
link |
opportunities to take
link |
this technique or that technique to solve that problem,
link |
I think that's the main driver for many
link |
From a general intelligence
link |
thing, it's kind of an interesting question.
link |
Are we really motivated to do that?
link |
struggled ourselves right now to even define
link |
So it's hard to incentivize when we don't even know
link |
what it is 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
link |
of that, that there's not a clear
link |
directive. There's no clear directive to do
link |
precisely that thing.
link |
So assistance in a larger and larger
link |
So being able to assist
link |
and be 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 would be doing the vacuum
link |
cleaning. And then the same system
link |
my kids that I don't have
link |
I think that when you get into
link |
a general intelligence for
link |
tasks, and again, I want to go back
link |
to your body question because I think your body question was interesting, but
link |
to go back to, you know, learning the abilities to
link |
physical tasks, you might have
link |
we might get, I imagine
link |
in that time frame, we will get better and better
link |
at learning these kinds of tasks, whether
link |
it's mowing your lawn or driving a car
link |
or whatever it is. I think we will get better
link |
and better at that where it's learning how to make
link |
predictions over large bodies of data. I think we're
link |
going to continue to get better and better at that.
link |
machines will out, you know, outpace humans
link |
and a variety of those things.
link |
The underlying mechanisms
link |
may be the same, meaning
link |
that, you know, maybe these are deep nets,
link |
there's infrastructure to train
link |
them, reusable components
link |
to get them to different
link |
classes of tasks, and we get better
link |
and better at building these kinds of machines.
link |
You could see, argue that
link |
the general learning infrastructure in there is
link |
a form of a general type of
link |
intelligence. I think
link |
what starts getting harder
link |
effectively communicate and understand and build
link |
that shared understanding because of the
link |
layers of interpretation that are required to do
link |
that and the need for the machine
link |
to be engaged with humans at that level
link |
basis. So how do you get in there?
link |
How do you get the machine in the game?
link |
How do you get the machine in the intellectual
link |
To solve AGI, you probably
link |
have to solve that problem. You have to get
link |
the machine. So it's a little bit of a bootstrapping
link |
thing. Can we get the machine engaged
link |
in, you know, in the intellectual
link |
the intellectual dialogue
link |
with the humans? Are the humans
link |
sufficiently in intellectual dialogue with each other
link |
to generate enough
link |
data in this context?
link |
And how do you bootstrap that? Because
link |
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 |
is, and how committed
link |
so I think that's possible, but
link |
when I go back to, are we incentivized
link |
I know we're incentivized to do the former.
link |
Are we incentivized to do the latter significantly
link |
enough? Do people understand what the latter really
link |
is well enough? Part of the
link |
elemental cognition mission is to try
link |
to articulate that better and better
link |
through demonstrations and through trying to craft
link |
these grand challenges and get
link |
people to say, look, this is a class of intelligence.
link |
This is a class of AI.
link |
What is the potential of this?
link |
What are the business, what's the business potential?
link |
What's the societal potential
link |
to that? And to, you know, and to
link |
build up that incentive system
link |
Yeah, I think if people don't understand yet, I think they will.
link |
I think there's a huge business potential
link |
here. So it's exciting that you're working on it.
link |
I kind of skipped over, but
link |
physical presence of things.
link |
Watson had a body?
link |
having a body adds to
link |
the interactive element
link |
between the AI system and a human
link |
or just in general to intelligence?
link |
going back to that
link |
shared understanding bit
link |
humans are very connected to their bodies.
link |
I mean, one of the reasons,
link |
one of the challenges in getting
link |
an AI to kind of be a compatible
link |
human intelligence
link |
is that our physical bodies
link |
are generating a lot of features
link |
So in other words, where our bodies are
link |
are the tool we use to
link |
affect output, but
link |
they also generate a lot of input
link |
for our brains. So we generate
link |
emotion, we generate all these
link |
feelings, we generate all these signals
link |
that machines don't have. So it means
link |
those that have this as the input data
link |
have the feedback that says, okay, I've
link |
gotten this, I've gotten this emotion
link |
or I've gotten this idea, I now
link |
want to process it and then I can
link |
it then affects me
link |
as a physical being and then
link |
out. In other words, I could realize
link |
the implications of that, because the implications again on
link |
I then process that and
link |
the implications again, our internal features
link |
are generated. I learned from
link |
them, they have an effect on my
link |
mind body complex. So
link |
it's interesting when we think, do we want
link |
a human intelligence? Well
link |
if we want a human compatible intelligence
link |
probably the best thing to do is to embed
link |
it in a human body.
link |
Just to clarify, and both concepts are
link |
beautiful, is a humanoid
link |
that look like humans is one
link |
sort of what Elon Musk was working with
link |
embedding intelligence
link |
systems to ride along
link |
No, I mean riding along is different.
link |
I meant like if you want
link |
to create an intelligence
link |
that is human compatible
link |
it can learn and develop a shared
link |
understanding of the world around it, you have to
link |
give it a lot of the same substrate.
link |
Part of that substrate
link |
is the idea that it
link |
generates these kinds of internal features
link |
like sort of emotional stuff, it has similar
link |
senses, it has to do a lot of the same
link |
things with those same senses.
link |
if you want that, again, I don't know that you want
link |
my specific goal. I think that's a fascinating
link |
scientific goal. I think it has all kinds of other implications.
link |
That's sort of not the goal.
link |
I think of it as I create intellectual thought
link |
partners for humans, that kind
link |
I know there are other companies that are creating
link |
physical thought partners, physical partners
link |
for humans, but that's
link |
kind of not where I'm
link |
the important point is that
link |
physical experience of the world around us.
link |
On the point of thought
link |
partners, what role
link |
does an emotional connection
link |
or forgive me, love
link |
in that thought partnership?
link |
Is that something you're interested in
link |
put another way sort of having
link |
With the AI? Yeah, with the AI between human
link |
and AI. Is that something that gets
link |
the rational discourse?
link |
Is that something that's useful?
link |
I worry about biases, obviously.
link |
So in other words, if you develop
link |
an emotional relationship with the machine
link |
all of a sudden you start are more likely
link |
to believe what it's saying even if it doesn't
link |
make any sense. So I
link |
But at the same time, I think the opportunity
link |
to use machines to provide human companionship
link |
is actually not crazy.
link |
social companionship is not a crazy idea.
link |
Do you have concerns
link |
as a few people do
link |
Elon Musk, Sam Harris
link |
about long term existential threats
link |
and perhaps short term threats
link |
of AI? We talked about bias
link |
we talked about different misuses but
link |
do you have concerns about
link |
systems that are able to
link |
help us make decisions together with 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 the giving machines
link |
and what I mean by leverage
link |
control over things that can hurt us
link |
whether it's socially,
link |
psychologically, intellectually, or physically
link |
and if you give the machines too much control
link |
I think that's a concern.
link |
You forget about the AI, just when you give them
link |
too much control human bad actors
link |
and produce havoc.
link |
hackers taking over the driverless car network
link |
and creating all kinds of
link |
but you could also imagine
link |
the ease at which humans could be persuaded
link |
one way or the other
link |
and now we have algorithms that can easily
link |
take control over that
link |
all ways and move people one direction
link |
Humans do that to other humans all the time
link |
and we have marketing campaigns, we have political campaigns
link |
that take advantage of
link |
and this is done all the time
link |
machines are like giant megaphones
link |
we can amplify this in orders of magnitude
link |
and fine tune its control
link |
so we can tailor the message
link |
we can now very rapidly
link |
additionally tailor the message to the audience
link |
advantage of their
link |
biases and amplifying them
link |
and using them to pursue them in one direction
link |
or another in ways that are
link |
not fair, not logical
link |
not objective, not meaningful
link |
machines empower that
link |
so that's what I mean by leverage
link |
but wow it's powerful because
link |
machines can do it more effectively
link |
you know more quickly and we see that already
link |
going on in social media
link |
I go back to saying
link |
one of the most important public
link |
dialogues we could be having
link |
is about the nature of intelligence
link |
and logic and reason and rationality
link |
us understanding our own biases
link |
us understanding our own cognitive biases
link |
and then how machines work
link |
and how do we use them to complement
link |
basically so that in the end we have
link |
a stronger overall system
link |
that's just incredibly important
link |
most people understand that
link |
telling your kids or telling your students
link |
this goes back to the cognition
link |
here's how your brain works
link |
here's how easy it is
link |
to trick your brain
link |
there are fundamental cognizant
link |
but you should appreciate
link |
the different types of thinking
link |
and what you're prone to
link |
and what do you prefer
link |
and under what conditions
link |
does this make sense versus that makes sense
link |
and then say here's what AI can do
link |
here's how it can make this worse
link |
and here's how it can make this better
link |
and that's where the AI has a role
link |
a system that is able
link |
beyond any definition
link |
of the Turing test
link |
the benchmark really an AGI system
link |
as a thought partner
link |
topic of discussion
link |
if you get to pick one
link |
would you have with that system
link |
what would you ask
link |
and you get to find out
link |
so you threw me a little bit
link |
with finding the truth at the end but
link |
because the truth is
link |
a whole other topic
link |
but the I think the beauty of it
link |
I think what excites me is the beauty of it is
link |
if I really have that system
link |
I don't have to pick
link |
so in other words I can go to
link |
and say this is what I care about today
link |
and that's what we mean by
link |
like this general capability
link |
go out, read this stuff in the next 3 milliseconds
link |
and I want to talk to you about it
link |
I want to draw analogies
link |
I want to understand how this affects
link |
this decision or that decision
link |
what if this were true
link |
what if that were true
link |
what knowledge should I be aware of
link |
that could impact my decision
link |
here's what I'm thinking
link |
is the main implication
link |
can you prove that out
link |
can you give me the evidence that supports that
link |
can you give me evidence that supports this other thing
link |
boy would that be incredible
link |
would that be just incredible
link |
just to be part of
link |
whether it's a medical diagnosis
link |
or whether it's the various treatment options
link |
legal case or whether it's
link |
a social problem that people are discussing
link |
be part of the dialogue
link |
and us accountable
link |
to reasons and objective dialogue
link |
goosebumps talking about it right
link |
this is what I want
link |
so when you created
link |
please come back on the podcast
link |
and we can have a discussion together
link |
and make it even longer
link |
this is a record for the longest conversation
link |
and it was an honor, it was a pleasure
link |
thank you so much for that