David Ferrucci: IBM Watson, Jeopardy & Deep Conversations with AI
心理与人性AI 与机器学习音乐与艺术技术与编程生物与进化
📋 章节目录
暂无章节信息
🔑 关键词
humansintelligencehumandonmachineablehardstuffdoingdatainterestingsayingjeopardyunderstandingwordsknowledgebetterframeworksimportantgoing
💬 精彩语录
暂无语录
🎙️ 完整对话（3258 条）
Lex Fridman (00:00.000)
The following is a conversation with David Feroci.
以下是与大卫·费罗西的对话。
Lex Fridman (00:03.040)
He led the team that built Watson,
他领导了沃森的团队，
Lex Fridman (00:05.200)
the IBM question answering system
IBM问答系统
Lex Fridman (00:07.040)
that beat the top humans in the world
击败了世界上最顶尖的人类
Lex Fridman (00:09.080)
at the game of Jeopardy.
在危险游戏中。
David Ferrucci (00:11.160)
For spending a couple hours with David,
为了和大卫一起度过几个小时，
Lex Fridman (00:12.920)
I saw a genuine passion,
我看到了真诚的热情，
David Ferrucci (00:14.960)
not only for abstract understanding of intelligence,
不仅是为了对智力的抽象理解，
Lex Fridman (00:17.720)
but for engineering it to solve real world problems
但为了解决现实世界的问题
David Ferrucci (00:21.240)
under real world deadlines and resource constraints.
Lex Fridman (00:24.800)
Where science meets engineering
当科学与工程相遇
David Ferrucci (00:26.520)
is where brilliant, simple ingenuity emerges.
这就是才华横溢、简单的独创性的体现。
Lex Fridman (00:29.940)
People who work adjoining it to
在其附近工作的人
David Ferrucci (00:32.120)
have a lot of wisdom earned
获得了很多智慧
Lex Fridman (00:33.800)
through failures and eventual success.
通过失败和最终的成功。
David Ferrucci (00:36.960)
David is also the founder, CEO,
大卫也是创始人、首席执行官
Lex Fridman (00:39.080)
and chief scientist of Elemental Cognition,
元素认知首席科学家，
David Ferrucci (00:41.660)
a company working to engineer AI systems
一家致力于设计人工智能系统的公司
Lex Fridman (00:44.480)
that understand the world the way people do.
以人们的方式理解世界。
David Ferrucci (00:47.420)
This is the Artificial Intelligence Podcast.
这是人工智能播客。
Lex Fridman (00:50.260)
If you enjoy it, subscribe on YouTube,
David Ferrucci (00:52.700)
give it five stars on iTunes,
Lex Fridman (00:54.440)
support it on Patreon,
David Ferrucci (00:55.840)
or simply connect with me on Twitter
Lex Fridman (00:57.920)
at Lex Friedman, spelled F R I D M A N.
Lex Fridman (01:01.360)
And now, here's my conversation with David Ferrucci.
Lex Fridman (01:06.120)
Your undergrad was in biology
David Ferrucci (01:07.980)
with an eye toward medical school
Lex Fridman (01:11.280)
before you went on for the PhD in computer science.
Lex Fridman (01:14.320)
So let me ask you an easy question.
Lex Fridman (01:16.800)
What is the difference between biological systems
Lex Fridman (01:20.520)
and computer systems?
Lex Fridman (01:22.420)
In your, when you sit back,
David Ferrucci (01:25.240)
look at the stars, and think philosophically.
Lex Fridman (01:28.600)
I often wonder whether or not
David Ferrucci (01:30.840)
there is a substantive difference.
Lex Fridman (01:32.920)
I mean, I think the thing that got me
David Ferrucci (01:34.480)
into computer science and into artificial intelligence
Lex Fridman (01:37.220)
was exactly this presupposition
David Ferrucci (01:39.840)
that if we can get machines to think,
Lex Fridman (01:44.360)
or I should say this question,
David Ferrucci (01:45.640)
this philosophical question,
Lex Fridman (01:47.460)
if we can get machines to think,
David Ferrucci (01:50.580)
to understand, to process information the way we do,
Lex Fridman (01:54.800)
so if we can describe a procedure, describe a process,
David Ferrucci (01:57.960)
even if that process were the intelligence process itself,
Lex Fridman (02:02.480)
then what would be the difference?
Lex Fridman (02:05.280)
So from a philosophical standpoint,
Lex Fridman (02:07.680)
I'm not sure I'm convinced that there is.
David Ferrucci (02:11.640)
I mean, you can go in the direction of spirituality,
Lex Fridman (02:14.920)
you can go in the direction of the soul,
Lex Fridman (02:16.660)
but in terms of what we can experience
Lex Fridman (02:21.660)
from an intellectual and physical perspective,
David Ferrucci (02:25.980)
I'm not sure there is.
Lex Fridman (02:27.460)
Clearly, there are different implementations,
Lex Fridman (02:31.080)
but if you were to say,
Lex Fridman (02:33.220)
is a biological information processing system
David Ferrucci (02:36.160)
fundamentally more capable
Lex Fridman (02:38.420)
than one we might be able to build out of silicon
David Ferrucci (02:41.020)
or some other substrate,
Lex Fridman (02:44.900)
I don't know that there is.
Lex Fridman (02:46.500)
How distant do you think is the biological implementation?
Lex Fridman (02:50.580)
So fundamentally, they may have the same capabilities,
Lex Fridman (02:53.820)
but is it really a far mystery
Lex Fridman (02:58.260)
where a huge number of breakthroughs are needed
David Ferrucci (03:00.700)
to be able to understand it,
Lex Fridman (03:02.700)
or is it something that, for the most part,
David Ferrucci (03:06.260)
in the important aspects,
Lex Fridman (03:08.620)
echoes of the same kind of characteristics?
David Ferrucci (03:11.100)
Yeah, that's interesting.
Lex Fridman (03:12.100)
I mean, so your question presupposes
David Ferrucci (03:15.580)
that there's this goal to recreate
Lex Fridman (03:17.520)
what we perceive as biological intelligence.
David Ferrucci (03:20.880)
I'm not sure that's the,
Lex Fridman (03:24.360)
I'm not sure that's how I would state the goal.
David Ferrucci (03:26.560)
I mean, I think that studying.
Lex Fridman (03:27.760)
What is the goal?
David Ferrucci (03:29.220)
Good, so I think there are a few goals.
Lex Fridman (03:32.180)
I think that understanding the human brain
Lex Fridman (03:35.720)
and how it works is important
Lex Fridman (03:38.520)
for us to be able to diagnose and treat issues,
David Ferrucci (03:43.520)
treat issues for us to understand our own strengths
Lex Fridman (03:47.200)
and weaknesses, both intellectual,
David Ferrucci (03:51.040)
psychological, and physical.
Lex Fridman (03:52.440)
So neuroscience and understanding the brain,
David Ferrucci (03:55.000)
from that perspective, there's a clear goal there.
Lex Fridman (03:59.560)
From the perspective of saying,
David Ferrucci (04:00.920)
I wanna mimic human intelligence,
Lex Fridman (04:04.840)
that one's a little bit more interesting.
David Ferrucci (04:06.440)
Human intelligence certainly has a lot of things we envy.
Lex Fridman (04:10.480)
It's also got a lot of problems, too.
Lex Fridman (04:12.880)
So I think we're capable of sort of stepping back
Lex Fridman (04:16.660)
and saying, what do we want out of an intelligence?
Lex Fridman (04:22.160)
How do we wanna communicate with that intelligence?
Lex Fridman (04:24.400)
How do we want it to behave?
Lex Fridman (04:25.560)
How do we want it to perform?
Lex Fridman (04:27.440)
Now, of course, it's somewhat of an interesting argument
David Ferrucci (04:30.360)
because I'm sitting here as a human
Lex Fridman (04:32.040)
with a biological brain,
Lex Fridman (04:33.940)
and I'm critiquing the strengths and weaknesses
Lex Fridman (04:36.420)
of human intelligence and saying
David Ferrucci (04:38.640)
that we have the capacity to step back
Lex Fridman (04:42.200)
and say, gee, what is intelligence
Lex Fridman (04:44.120)
and what do we really want out of it?
Lex Fridman (04:46.000)
And that, in and of itself, suggests that
David Ferrucci (04:50.040)
human intelligence is something quite enviable,
Lex Fridman (04:52.080)
that it can introspect that way.
Lex Fridman (04:58.360)
And the flaws, you mentioned the flaws.
Lex Fridman (05:00.280)
Humans have flaws.
David Ferrucci (05:01.120)
Yeah, but I think that flaws that human intelligence has
Lex Fridman (05:04.720)
is extremely prejudicial and biased
David Ferrucci (05:08.420)
in the way it draws many inferences.
Lex Fridman (05:10.480)
Do you think those are, sorry to interrupt,
Lex Fridman (05:12.040)
do you think those are features or are those bugs?
Lex Fridman (05:14.360)
Do you think the prejudice, the forgetfulness, the fear,
Lex Fridman (05:21.520)
what are the flaws?
Lex Fridman (05:22.860)
List them all.
Lex Fridman (05:23.700)
What, love?
Lex Fridman (05:24.520)
Maybe that's a flaw.
David Ferrucci (05:25.560)
You think those are all things that can be gotten,
Lex Fridman (05:28.920)
get in the way of intelligence
Lex Fridman (05:30.780)
or the essential components of intelligence?
Lex Fridman (05:33.420)
Well, again, if you go back and you define intelligence
David Ferrucci (05:36.160)
as being able to sort of accurately, precisely, rigorously,
Lex Fridman (05:42.120)
reason, develop answers,
Lex Fridman (05:43.840)
and justify those answers in an objective way,
Lex Fridman (05:46.640)
yeah, then human intelligence has these flaws
David Ferrucci (05:49.720)
in that it tends to be more influenced
Lex Fridman (05:52.880)
by some of the things you said.
Lex Fridman (05:56.520)
And it's largely an inductive process,
Lex Fridman (05:59.740)
meaning it takes past data,
David Ferrucci (06:01.580)
uses that to predict the future.
Lex Fridman (06:03.560)
Very advantageous in some cases,
Lex Fridman (06:06.000)
but fundamentally biased and prejudicial in other cases
Lex Fridman (06:09.280)
because it's gonna be strongly influenced by its priors,
David Ferrucci (06:11.520)
whether they're right or wrong
Lex Fridman (06:13.880)
from some objective reasoning perspective,
David Ferrucci (06:17.420)
you're gonna favor them because those are the decisions
Lex Fridman (06:20.500)
or those are the paths that succeeded in the past.
Lex Fridman (06:24.000)
And I think that mode of intelligence makes a lot of sense
Lex Fridman (06:29.240)
for when your primary goal is to act quickly
Lex Fridman (06:33.280)
and survive and make fast decisions.
Lex Fridman (06:37.200)
And I think those create problems
David Ferrucci (06:40.300)
when you wanna think more deeply
Lex Fridman (06:42.040)
and make more objective and reasoned decisions.
David Ferrucci (06:45.120)
Of course, humans capable of doing both.
Lex Fridman (06:48.320)
They do sort of one more naturally than they do the other,
Lex Fridman (06:51.120)
but they're capable of doing both.
Lex Fridman (06:53.280)
You're saying they do the one
David Ferrucci (06:54.540)
that responds quickly more naturally.
Lex Fridman (06:56.480)
Right.
David Ferrucci (06:57.320)
Because that's the thing we kind of need
Lex Fridman (06:58.440)
to not be eaten by the predators in the world.
David Ferrucci (07:02.720)
For example, but then we've learned to reason through logic,
Lex Fridman (07:09.560)
we've developed science, we train people to do that.
David Ferrucci (07:13.960)
I think that's harder for the individual to do.
Lex Fridman (07:16.960)
I think it requires training and teaching.
David Ferrucci (07:20.960)
I think we are, the human mind certainly is capable of it,
Lex Fridman (07:24.180)
but we find it more difficult.
Lex Fridman (07:25.280)
And then there are other weaknesses, if you will,
Lex Fridman (07:27.680)
as you mentioned earlier, just memory capacity
Lex Fridman (07:30.620)
and how many chains of inference
Lex Fridman (07:33.800)
can you actually go through without like losing your way?
Lex Fridman (07:37.280)
So just focus and...
Lex Fridman (07:40.280)
So the way you think about intelligence,
Lex Fridman (07:43.240)
and we're really sort of floating
Lex Fridman (07:45.080)
in this philosophical space,
Lex Fridman (07:47.220)
but I think you're like the perfect person
Lex Fridman (07:50.080)
to talk about this,
David Ferrucci (07:52.160)
because we'll get to Jeopardy and beyond.
Lex Fridman (07:55.660)
That's like one of the most incredible accomplishments
David Ferrucci (07:58.700)
in AI, in the history of AI,
Lex Fridman (08:00.920)
but hence the philosophical discussion.
Lex Fridman (08:03.400)
So let me ask, you've kind of alluded to it,
Lex Fridman (08:06.280)
but let me ask again, what is intelligence?
David Ferrucci (08:09.400)
Underlying the discussions we'll have
Lex Fridman (08:12.460)
with Jeopardy and beyond,
Lex Fridman (08:15.520)
how do you think about intelligence?
Lex Fridman (08:17.080)
Is it a sufficiently complicated problem
Lex Fridman (08:19.800)
being able to reason your way through solving that problem?
Lex Fridman (08:22.440)
Is that kind of how you think about
Lex Fridman (08:23.800)
what it means to be intelligent?
Lex Fridman (08:25.440)
So I think of intelligence primarily two ways.
David Ferrucci (08:29.720)
One is the ability to predict.
Lex Fridman (08:33.320)
So in other words, if I have a problem,
Lex Fridman (08:35.840)
can I predict what's gonna happen next?
Lex Fridman (08:37.600)
Whether it's to predict the answer of a question
David Ferrucci (08:40.900)
or to say, look, I'm looking at all the market dynamics
Lex Fridman (08:43.880)
and I'm gonna tell you what's gonna happen next,
David Ferrucci (08:46.160)
or you're in a room and somebody walks in
Lex Fridman (08:49.400)
and you're gonna predict what they're gonna do next
David Ferrucci (08:51.320)
or what they're gonna say next.
Lex Fridman (08:53.660)
You're in a highly dynamic environment
David Ferrucci (08:55.760)
full of uncertainty, be able to predict.
Lex Fridman (08:58.680)
The more variables, the more complex.
David Ferrucci (09:01.480)
The more possibilities, the more complex.
Lex Fridman (09:04.080)
But can I take a small amount of prior data
Lex Fridman (09:07.720)
and learn the pattern and then predict
Lex Fridman (09:09.880)
what's gonna happen next accurately and consistently?
David Ferrucci (09:13.920)
That's certainly a form of intelligence.
Lex Fridman (09:16.960)
What do you need for that, by the way?
David Ferrucci (09:18.320)
You need to have an understanding
Lex Fridman (09:21.160)
of the way the world works
Lex Fridman (09:22.880)
in order to be able to unroll it into the future, right?
Lex Fridman (09:26.320)
What do you think is needed to predict?
David Ferrucci (09:28.000)
Depends what you mean by understanding.
Lex Fridman (09:29.440)
I need to be able to find that function.
David Ferrucci (09:32.240)
This is very much what deep learning does,
Lex Fridman (09:35.100)
machine learning does, is if you give me enough prior data
Lex Fridman (09:38.960)
and you tell me what the output variable is that matters,
Lex Fridman (09:41.940)
I'm gonna sit there and be able to predict it.
Lex Fridman (09:44.440)
And if I can predict it accurately
Lex Fridman (09:47.280)
so that I can get it right more often than not,
David Ferrucci (09:50.280)
I'm smart, if I can do that with less data
Lex Fridman (09:52.920)
and less training time, I'm even smarter.
David Ferrucci (09:58.000)
If I can figure out what's even worth predicting,
Lex Fridman (10:01.640)
I'm smarter, meaning I'm figuring out
Lex Fridman (10:03.920)
what path is gonna get me toward a goal.
Lex Fridman (10:06.400)
What about picking a goal?
David Ferrucci (10:07.560)
Sorry, you left again.
Lex Fridman (10:08.480)
Well, that's interesting about picking a goal,
David Ferrucci (10:10.120)
sort of an interesting thing.
Lex Fridman (10:11.080)
I think that's where you bring in
Lex Fridman (10:13.240)
what are you preprogrammed to do?
Lex Fridman (10:15.040)
We talk about humans,
Lex Fridman (10:16.000)
and well, humans are preprogrammed to survive.
Lex Fridman (10:19.400)
So it's sort of their primary driving goal.
Lex Fridman (10:23.320)
What do they have to do to do that?
Lex Fridman (10:24.700)
And that can be very complex, right?
Lex Fridman (10:27.360)
So it's not just figuring out that you need to run away
Lex Fridman (10:31.680)
from the ferocious tiger,
Lex Fridman (10:33.640)
but we survive in a social context as an example.
Lex Fridman (10:38.720)
So understanding the subtleties of social dynamics
David Ferrucci (10:42.320)
becomes something that's important for surviving,
Lex Fridman (10:45.420)
finding a mate, reproducing, right?
Lex Fridman (10:47.200)
So we're continually challenged with
Lex Fridman (10:50.320)
complex sets of variables, complex constraints,
David Ferrucci (10:53.760)
rules, if you will, or patterns.
Lex Fridman (10:56.880)
And we learn how to find the functions
Lex Fridman (10:59.320)
and predict the things.
Lex Fridman (11:00.680)
In other words, represent those patterns efficiently
Lex Fridman (11:03.580)
and be able to predict what's gonna happen.
Lex Fridman (11:04.920)
And that's a form of intelligence.
David Ferrucci (11:06.080)
That doesn't really require anything specific
Lex Fridman (11:11.400)
other than the ability to find that function
Lex Fridman (11:13.400)
and predict that right answer.
Lex Fridman (11:15.840)
That's certainly a form of intelligence.
Lex Fridman (11:18.440)
But then when we say, well, do we understand each other?
Lex Fridman (11:23.280)
In other words, would you perceive me as intelligent
Lex Fridman (11:28.640)
beyond that ability to predict?
Lex Fridman (11:31.040)
So now I can predict, but I can't really articulate
Lex Fridman (11:35.220)
how I'm going through that process,
Lex Fridman (11:37.840)
what my underlying theory is for predicting,
Lex Fridman (11:41.240)
and I can't get you to understand what I'm doing
Lex Fridman (11:43.680)
so that you can figure out how to do this yourself
David Ferrucci (11:48.040)
if you did not have, for example,
Lex Fridman (11:50.760)
the right pattern matching machinery that I did.
Lex Fridman (11:53.840)
And now we potentially have this breakdown
Lex Fridman (11:55.740)
where, in effect, I'm intelligent,
Lex Fridman (11:59.080)
but I'm sort of an alien intelligence relative to you.
Lex Fridman (12:02.640)
You're intelligent, but nobody knows about it, or I can't.
David Ferrucci (12:05.960)
Well, I can see the output.
Lex Fridman (12:08.200)
So you're saying, let's sort of separate the two things.
David Ferrucci (12:11.680)
One is you explaining why you were able
Lex Fridman (12:15.880)
to predict the future,
Lex Fridman (12:19.160)
and the second is me being able to,
Lex Fridman (12:23.360)
impressing me that you're intelligent,
David Ferrucci (12:25.520)
me being able to know
Lex Fridman (12:26.520)
that you successfully predicted the future.
Lex Fridman (12:28.680)
Do you think that's?
Lex Fridman (12:29.600)
Well, it's not impressing you that I'm intelligent.
David Ferrucci (12:31.360)
In other words, you may be convinced
Lex Fridman (12:33.640)
that I'm intelligent in some form.
Lex Fridman (12:35.960)
So how, what would convince?
Lex Fridman (12:37.160)
Because of my ability to predict.
Lex Fridman (12:38.880)
So I would look at the metrics.
Lex Fridman (12:39.720)
When you can't, I'd say, wow.
David Ferrucci (12:41.120)
You're right more times than I am.
Lex Fridman (12:44.960)
You're doing something interesting.
David Ferrucci (12:46.280)
That's a form of intelligence.
Lex Fridman (12:49.120)
But then what happens is, if I say, how are you doing that?
Lex Fridman (12:53.400)
And you can't communicate with me,
Lex Fridman (12:55.280)
and you can't describe that to me,
David Ferrucci (12:57.720)
now I may label you a savant.
Lex Fridman (13:00.700)
I may say, well, you're doing something weird,
Lex Fridman (13:03.240)
and it's just not very interesting to me,
Lex Fridman (13:06.360)
because you and I can't really communicate.
Lex Fridman (13:09.360)
And so now, so this is interesting, right?
Lex Fridman (13:12.360)
Because now this is, you're in this weird place
David Ferrucci (13:15.120)
where for you to be recognized
Lex Fridman (13:17.480)
as intelligent the way I'm intelligent,
David Ferrucci (13:21.300)
then you and I sort of have to be able to communicate.
Lex Fridman (13:24.280)
And then my, we start to understand each other,
Lex Fridman (13:28.520)
and then my respect and my appreciation,
Lex Fridman (13:33.480)
my ability to relate to you starts to change.
Lex Fridman (13:36.760)
So now you're not an alien intelligence anymore.
Lex Fridman (13:39.080)
You're a human intelligence now,
David Ferrucci (13:41.080)
because you and I can communicate.
Lex Fridman (13:43.880)
And so I think when we look at animals, for example,
David Ferrucci (13:48.120)
animals can do things we can't quite comprehend,
Lex Fridman (13:50.720)
we don't quite know how they do them,
Lex Fridman (13:51.800)
but they can't really communicate with us.
Lex Fridman (13:54.420)
They can't put what they're going through in our terms.
Lex Fridman (13:58.360)
And so we think of them as sort of,
Lex Fridman (13:59.720)
well, they're these alien intelligences,
Lex Fridman (14:01.520)
and they're not really worth necessarily what we're worth.
Lex Fridman (14:03.600)
We don't treat them the same way as a result of that.
Lex Fridman (14:06.360)
But it's hard because who knows what's going on.
Lex Fridman (14:11.640)
So just a quick elaboration on that,
David Ferrucci (14:15.640)
the explaining that you're intelligent,
Lex Fridman (14:18.760)
the explaining the reasoning that went into the prediction
David Ferrucci (14:23.480)
is not some kind of mathematical proof.
Lex Fridman (14:27.080)
If we look at humans,
David Ferrucci (14:28.840)
look at political debates and discourse on Twitter,
Lex Fridman (14:32.220)
it's mostly just telling stories.
Lex Fridman (14:35.340)
So your task is, sorry,
Lex Fridman (14:38.400)
your task is not to tell an accurate depiction
David Ferrucci (14:43.680)
of how you reason, but to tell a story, real or not,
Lex Fridman (14:48.420)
that convinces me that there was a mechanism by which you.
David Ferrucci (14:52.000)
Ultimately, that's what a proof is.
Lex Fridman (14:53.600)
I mean, even a mathematical proof is that.
David Ferrucci (14:56.240)
Because ultimately, the other mathematicians
Lex Fridman (14:58.200)
have to be convinced by your proof.
David Ferrucci (15:01.600)
Otherwise, in fact, there have been.
Lex Fridman (15:03.000)
That's the metric for success, yeah.
David Ferrucci (15:04.440)
There have been several proofs out there
Lex Fridman (15:06.080)
where mathematicians would study for a long time
David Ferrucci (15:08.040)
before they were convinced
Lex Fridman (15:08.880)
that it actually proved anything, right?
David Ferrucci (15:10.920)
You never know if it proved anything
Lex Fridman (15:12.240)
until the community of mathematicians decided that it did.
Lex Fridman (15:14.880)
So I mean, but it's a real thing, right?
Lex Fridman (15:18.680)
And that's sort of the point, right?
David Ferrucci (15:20.960)
Is that ultimately, this notion of understanding us,
Lex Fridman (15:24.640)
understanding something is ultimately a social concept.
David Ferrucci (15:28.240)
In other words, I have to convince enough people
Lex Fridman (15:30.740)
that I did this in a reasonable way.
David Ferrucci (15:33.760)
I did this in a way that other people can understand
Lex Fridman (15:36.400)
and replicate and that it makes sense to them.
Lex Fridman (15:39.840)
So human intelligence is bound together in that way.
Lex Fridman (15:44.840)
We're bound up in that sense.
David Ferrucci (15:47.460)
We sort of never really get away with it
Lex Fridman (15:49.560)
until we can sort of convince others
David Ferrucci (15:52.600)
that our thinking process makes sense.
Lex Fridman (15:55.880)
Did you think the general question of intelligence
Lex Fridman (15:59.120)
is then also a social construct?
Lex Fridman (16:01.000)
So if we ask questions of an artificial intelligence system,
Lex Fridman (16:06.720)
is this system intelligent?
Lex Fridman (16:08.640)
The answer will ultimately be a socially constructed.
David Ferrucci (16:12.640)
I think, so I think I'm making two statements.
Lex Fridman (16:16.040)
I'm saying we can try to define intelligence
David Ferrucci (16:18.040)
in this super objective way that says, here's this data.
Lex Fridman (16:23.120)
I wanna predict this type of thing, learn this function.
Lex Fridman (16:26.760)
And then if you get it right, often enough,
Lex Fridman (16:30.360)
we consider you intelligent.
Lex Fridman (16:32.080)
But that's more like a sub bond.
Lex Fridman (16:34.400)
I think it is.
David Ferrucci (16:35.720)
It doesn't mean it's not useful.
Lex Fridman (16:37.120)
It could be incredibly useful.
David Ferrucci (16:38.640)
It could be solving a problem we can't otherwise solve
Lex Fridman (16:41.460)
and can solve it more reliably than we can.
Lex Fridman (16:44.480)
But then there's this notion of,
Lex Fridman (16:46.960)
can humans take responsibility
Lex Fridman (16:50.420)
for the decision that you're making?
Lex Fridman (16:53.680)
Can we make those decisions ourselves?
Lex Fridman (16:56.120)
Can we relate to the process that you're going through?
Lex Fridman (16:58.840)
And now you as an agent,
David Ferrucci (17:01.160)
whether you're a machine or another human, frankly,
Lex Fridman (17:04.520)
are now obliged to make me understand
Lex Fridman (17:08.640)
how it is that you're arriving at that answer
Lex Fridman (17:10.860)
and allow me, me or obviously a community
David Ferrucci (17:13.840)
or a judge of people to decide
Lex Fridman (17:15.000)
whether or not that makes sense.
Lex Fridman (17:17.280)
And by the way, that happens with the humans as well.
Lex Fridman (17:20.200)
You're sitting down with your staff, for example,
Lex Fridman (17:22.080)
and you ask for suggestions about what to do next.
Lex Fridman (17:26.360)
And someone says, oh, I think you should buy.
Lex Fridman (17:28.640)
And I actually think you should buy this much
Lex Fridman (17:30.560)
or whatever or sell or whatever it is.
David Ferrucci (17:33.160)
Or I think you should launch the product today or tomorrow
Lex Fridman (17:35.720)
or launch this product versus that product,
David Ferrucci (17:37.080)
whatever the decision may be.
Lex Fridman (17:38.600)
And you ask why.
Lex Fridman (17:39.840)
And the person says,
Lex Fridman (17:40.680)
I just have a good feeling about it.
Lex Fridman (17:42.800)
And you're not very satisfied.
Lex Fridman (17:44.400)
Now, that person could be,
David Ferrucci (17:47.600)
you might say, well, you've been right before,
Lex Fridman (17:50.880)
but I'm gonna put the company on the line.
Lex Fridman (17:54.140)
Can you explain to me why I should believe this?
Lex Fridman (17:56.760)
Right.
Lex Fridman (17:58.000)
And that explanation may have nothing to do with the truth.
Lex Fridman (18:00.960)
You just, the ultimate.
David Ferrucci (18:01.800)
It's gotta convince the other person.
Lex Fridman (18:03.520)
Still be wrong, still be wrong.
David Ferrucci (18:05.280)
She's gotta be convincing.
Lex Fridman (18:06.320)
But it's ultimately gotta be convincing.
Lex Fridman (18:07.840)
And that's why I'm saying it's,
Lex Fridman (18:10.200)
we're bound together, right?
David Ferrucci (18:12.160)
Our intelligences are bound together in that sense.
Lex Fridman (18:14.160)
We have to understand each other.
Lex Fridman (18:15.360)
And if, for example, you're giving me an explanation,
Lex Fridman (18:18.920)
I mean, this is a very important point, right?
David Ferrucci (18:21.020)
You're giving me an explanation,
Lex Fridman (18:23.840)
and I'm not good,
Lex Fridman (18:29.380)
and then I'm not good at reasoning well,
Lex Fridman (18:33.520)
and being objective,
Lex Fridman (18:35.280)
and following logical paths and consistent paths,
Lex Fridman (18:39.160)
and I'm not good at measuring
Lex Fridman (18:41.400)
and sort of computing probabilities across those paths.
Lex Fridman (18:45.520)
What happens is collectively,
David Ferrucci (18:47.200)
we're not gonna do well.
Lex Fridman (18:50.120)
How hard is that problem?
David Ferrucci (18:52.240)
The second one.
Lex Fridman (18:53.180)
So I think we'll talk quite a bit about the first
David Ferrucci (18:57.960)
on a specific objective metric benchmark performing well.
Lex Fridman (19:03.840)
But being able to explain the steps,
Lex Fridman (19:07.440)
the reasoning, how hard is that problem?
Lex Fridman (19:10.580)
I think that's very hard.
David Ferrucci (19:11.800)
I mean, I think that that's,
Lex Fridman (19:16.040)
well, it's hard for humans.
David Ferrucci (19:18.160)
The thing that's hard for humans, as you know,
Lex Fridman (19:20.960)
may not necessarily be hard for computers
Lex Fridman (19:22.920)
and vice versa.
Lex Fridman (19:24.440)
So, sorry, so how hard is that problem for computers?
David Ferrucci (19:31.160)
I think it's hard for computers,
Lex Fridman (19:32.640)
and the reason why I related to,
David Ferrucci (19:34.600)
or saying that it's also hard for humans
Lex Fridman (19:36.400)
is because I think when we step back
Lex Fridman (19:38.320)
and we say we wanna design computers to do that,
Lex Fridman (19:43.520)
one of the things we have to recognize
David Ferrucci (19:46.440)
is we're not sure how to do it well.
Lex Fridman (19:50.520)
I'm not sure we have a recipe for that.
Lex Fridman (19:52.960)
And even if you wanted to learn it,
Lex Fridman (19:55.320)
it's not clear exactly what data we use
Lex Fridman (19:59.480)
and what judgments we use to learn that well.
Lex Fridman (1:00:01.100)
I was doing the open domain question answering stuff,
Lex Fridman (1:00:03.060)
but I was coming off a couple other projects.
Lex Fridman (1:00:05.940)
I had a lot more time to put into this.
Lex Fridman (1:00:08.060)
And I argued that it could be done.
Lex Fridman (1:00:10.180)
And I argue it would be crazy not to do this.
David Ferrucci (1:00:12.740)
Can I, you can be honest at this point.
Lex Fridman (1:00:15.820)
So even though you argued for it,
David Ferrucci (1:00:17.580)
what's the confidence that you had yourself privately
Lex Fridman (1:00:21.540)
that this could be done?
David Ferrucci (1:00:22.740)
Was, we just told the story,
Lex Fridman (1:00:25.620)
how you tell stories to convince others.
Lex Fridman (1:00:27.740)
How confident were you?
Lex Fridman (1:00:28.940)
What was your estimation of the problem at that time?
Lex Fridman (1:00:32.660)
So I thought it was possible.
Lex Fridman (1:00:34.300)
And a lot of people thought it was impossible.
David Ferrucci (1:00:36.300)
I thought it was possible.
Lex Fridman (1:00:37.860)
The reason why I thought it was possible
David Ferrucci (1:00:39.140)
was because I did some brief experimentation.
Lex Fridman (1:00:41.500)
I knew a lot about how we were approaching
David Ferrucci (1:00:43.460)
open domain factoid question answering.
Lex Fridman (1:00:45.940)
I've been doing it for some years.
David Ferrucci (1:00:47.620)
I looked at the Jeopardy stuff.
Lex Fridman (1:00:49.340)
I said, this is gonna be hard
David Ferrucci (1:00:50.900)
for a lot of the points that we mentioned earlier.
Lex Fridman (1:00:54.180)
Hard to interpret the question.
David Ferrucci (1:00:57.060)
Hard to do it quickly enough.
Lex Fridman (1:00:58.940)
Hard to compute an accurate confidence.
David Ferrucci (1:01:00.500)
None of this stuff had been done well enough before.
Lex Fridman (1:01:03.060)
But a lot of the technologies we're building
David Ferrucci (1:01:04.660)
were the kinds of technologies that should work.
Lex Fridman (1:01:07.500)
But more to the point, what was driving me was,
David Ferrucci (1:01:10.860)
I was in IBM research.
Lex Fridman (1:01:12.820)
I was a senior leader in IBM research.
Lex Fridman (1:01:14.900)
And this is the kind of stuff we were supposed to do.
Lex Fridman (1:01:17.140)
In other words, we were basically supposed to.
David Ferrucci (1:01:18.660)
This is the moonshot.
Lex Fridman (1:01:19.700)
This is the.
David Ferrucci (1:01:20.540)
We were supposed to take things and say,
Lex Fridman (1:01:21.900)
this is an active research area.
David Ferrucci (1:01:24.940)
It's our obligation to kind of,
Lex Fridman (1:01:27.540)
if we have the opportunity, to push it to the limits.
Lex Fridman (1:01:30.100)
And if it doesn't work,
Lex Fridman (1:01:31.460)
to understand more deeply why we can't do it.
Lex Fridman (1:01:34.740)
And so I was very committed to that notion saying,
Lex Fridman (1:01:37.900)
folks, this is what we do.
David Ferrucci (1:01:40.060)
It's crazy not to do this.
Lex Fridman (1:01:42.140)
This is an active research area.
David Ferrucci (1:01:43.740)
We've been in this for years.
Lex Fridman (1:01:44.980)
Why wouldn't we take this grand challenge
Lex Fridman (1:01:47.940)
and push it as hard as we can?
Lex Fridman (1:01:50.700)
At the very least, we'd be able to come out and say,
David Ferrucci (1:01:53.140)
here's why this problem is way hard.
Lex Fridman (1:01:57.060)
Here's what we tried and here's how we failed.
Lex Fridman (1:01:58.660)
So I was very driven as a scientist from that perspective.
Lex Fridman (1:02:03.980)
And then I also argued,
David Ferrucci (1:02:06.580)
based on what we did a feasibility study,
Lex Fridman (1:02:08.740)
why I thought it was hard but possible.
Lex Fridman (1:02:10.900)
And I showed examples of where it succeeded,
Lex Fridman (1:02:14.180)
where it failed, why it failed,
Lex Fridman (1:02:16.100)
and sort of a high level architecture approach
Lex Fridman (1:02:18.180)
for why we should do it.
Lex Fridman (1:02:19.540)
But for the most part, at that point,
Lex Fridman (1:02:22.260)
the execs really were just looking for someone crazy enough
David Ferrucci (1:02:24.660)
to say yes, because for several years at that point,
Lex Fridman (1:02:27.900)
everyone had said, no, I'm not willing to risk my reputation
Lex Fridman (1:02:32.260)
and my career on this thing.
Lex Fridman (1:02:34.820)
Clearly you did not have such fears.
David Ferrucci (1:02:36.740)
Okay. I did not.
Lex Fridman (1:02:37.980)
So you dived right in.
Lex Fridman (1:02:39.540)
And yet, for what I understand,
Lex Fridman (1:02:42.820)
it was performing very poorly in the beginning.
Lex Fridman (1:02:46.300)
So what were the initial approaches and why did they fail?
Lex Fridman (1:02:51.300)
Well, there were lots of hard aspects to it.
David Ferrucci (1:02:54.820)
I mean, one of the reasons why prior approaches
Lex Fridman (1:02:57.700)
that we had worked on in the past failed
David Ferrucci (1:03:02.380)
was because the questions were difficult to interpret.
Lex Fridman (1:03:07.780)
Like, what are you even asking for, right?
David Ferrucci (1:03:10.100)
Very often, like if the question was very direct,
Lex Fridman (1:03:12.500)
like what city, or what, even then it could be tricky,
Lex Fridman (1:03:16.620)
but what city or what person,
Lex Fridman (1:03:21.940)
is often when it would name it very clearly,
David Ferrucci (1:03:24.220)
you would know that.
Lex Fridman (1:03:25.420)
And if there were just a small set of them,
David Ferrucci (1:03:28.100)
in other words, we're gonna ask about these five types.
Lex Fridman (1:03:31.540)
Like, it's gonna be an answer,
Lex Fridman (1:03:33.580)
and the answer will be a city in this state
Lex Fridman (1:03:36.820)
or a city in this country.
Lex Fridman (1:03:37.780)
The answer will be a person of this type, right?
Lex Fridman (1:03:41.020)
Like an actor or whatever it is.
Lex Fridman (1:03:42.740)
But it turns out that in Jeopardy,
Lex Fridman (1:03:44.380)
there were like tens of thousands of these things.
Lex Fridman (1:03:47.580)
And it was a very, very long tail,
Lex Fridman (1:03:50.580)
meaning that it just went on and on.
Lex Fridman (1:03:52.500)
And so even if you focused on trying to encode the types
Lex Fridman (1:03:56.900)
at the very top, like there's five that were the most,
David Ferrucci (1:03:59.820)
let's say five of the most frequent,
Lex Fridman (1:04:01.580)
you still cover a very small percentage of the data.
Lex Fridman (1:04:04.140)
So you couldn't take that approach of saying,
Lex Fridman (1:04:07.140)
I'm just going to try to collect facts
David Ferrucci (1:04:09.780)
about these five or 10 types or 20 types
Lex Fridman (1:04:12.780)
or 50 types or whatever.
Lex Fridman (1:04:14.380)
So that was like one of the first things,
Lex Fridman (1:04:16.940)
like what do you do about that?
Lex Fridman (1:04:18.180)
And so we came up with an approach toward that.
Lex Fridman (1:04:21.500)
And the approach looked promising,
Lex Fridman (1:04:23.460)
and we continued to improve our ability
Lex Fridman (1:04:25.940)
to handle that problem throughout the project.
David Ferrucci (1:04:29.500)
The other issue was that right from the outside,
Lex Fridman (1:04:32.420)
I said, we're not going to,
David Ferrucci (1:04:34.580)
I committed to doing this in three to five years.
Lex Fridman (1:04:37.620)
So we did it in four.
Lex Fridman (1:04:39.060)
So I got lucky.
Lex Fridman (1:04:40.940)
But one of the things that that,
David Ferrucci (1:04:42.380)
putting that like stake in the ground was,
Lex Fridman (1:04:45.700)
and I knew how hard the language understanding problem was.
David Ferrucci (1:04:47.780)
I said, we're not going to actually understand language
Lex Fridman (1:04:51.620)
to solve this problem.
David Ferrucci (1:04:53.900)
We are not going to interpret the question
Lex Fridman (1:04:57.460)
and the domain of knowledge that the question refers to
Lex Fridman (1:05:00.180)
and reason over that to answer these questions.
Lex Fridman (1:05:02.420)
Obviously we're not going to be doing that.
David Ferrucci (1:05:04.140)
At the same time,
Lex Fridman (1:05:05.740)
simple search wasn't good enough to confidently answer
David Ferrucci (1:05:10.380)
with a single correct answer.
Lex Fridman (1:05:13.020)
First of all, that's like brilliant.
David Ferrucci (1:05:14.260)
That's such a great mix of innovation
Lex Fridman (1:05:16.140)
and practical engineering three, four, eight.
Lex Fridman (1:05:18.620)
So you're not trying to solve the general NLU problem.
Lex Fridman (1:05:21.780)
You're saying, let's solve this in any way possible.
David Ferrucci (1:05:25.260)
Oh, yeah.
Lex Fridman (1:05:26.100)
No, I was committed to saying, look,
David Ferrucci (1:05:28.020)
we're going to solving the open domain
Lex Fridman (1:05:29.660)
question answering problem.
David Ferrucci (1:05:31.020)
We're using Jeopardy as a driver for that.
Lex Fridman (1:05:33.540)
That's a big benchmark.
David Ferrucci (1:05:34.380)
Good enough, big benchmark, exactly.
Lex Fridman (1:05:36.500)
And now we're.
Lex Fridman (1:05:38.180)
How do we do it?
Lex Fridman (1:05:39.020)
We could just like, whatever,
David Ferrucci (1:05:39.940)
like just figure out what works
Lex Fridman (1:05:41.140)
because I want to be able to go back
David Ferrucci (1:05:42.340)
to the academic science community
Lex Fridman (1:05:44.100)
and say, here's what we tried.
David Ferrucci (1:05:45.980)
Here's what worked.
Lex Fridman (1:05:46.820)
Here's what didn't work.
David Ferrucci (1:05:47.660)
Great.
Lex Fridman (1:05:48.500)
I don't want to go in and say,
David Ferrucci (1:05:50.260)
oh, I only have one technology.
Lex Fridman (1:05:51.980)
I have a hammer.
David Ferrucci (1:05:52.820)
I'm only going to use this.
Lex Fridman (1:05:53.660)
I'm going to do whatever it takes.
David Ferrucci (1:05:54.700)
I'm like, I'm going to think out of the box
Lex Fridman (1:05:56.020)
and do whatever it takes.
David Ferrucci (1:05:57.300)
One, and I also, there was another thing I believed.
Lex Fridman (1:06:00.540)
I believed that the fundamental NLP technologies
Lex Fridman (1:06:04.580)
and machine learning technologies would be adequate.
Lex Fridman (1:06:08.780)
And this was an issue of how do we enhance them?
Lex Fridman (1:06:11.940)
How do we integrate them?
Lex Fridman (1:06:13.620)
How do we advance them?
Lex Fridman (1:06:15.300)
So I had one researcher who came to me
Lex Fridman (1:06:17.220)
who had been working on question answering
David Ferrucci (1:06:18.620)
with me for a very long time,
Lex Fridman (1:06:21.620)
who had said, we're going to need Maxwell's equations
David Ferrucci (1:06:24.260)
for question answering.
Lex Fridman (1:06:25.660)
And I said, if we need some fundamental formula
David Ferrucci (1:06:28.700)
that breaks new ground in how we understand language,
Lex Fridman (1:06:31.820)
we're screwed.
David Ferrucci (1:06:33.060)
We're not going to get there from here.
Lex Fridman (1:06:34.380)
Like I am not counting.
David Ferrucci (1:06:38.020)
My assumption is I'm not counting
Lex Fridman (1:06:39.660)
on some brand new invention.
Lex Fridman (1:06:42.380)
What I'm counting on is the ability
Lex Fridman (1:06:45.420)
to take everything it has done before
David Ferrucci (1:06:48.100)
to figure out an architecture on how to integrate it well
Lex Fridman (1:06:51.860)
and then see where it breaks
Lex Fridman (1:06:54.300)
and make the necessary advances we need to make
Lex Fridman (1:06:57.220)
until this thing works.
David Ferrucci (1:06:58.860)
Push it hard to see where it breaks
Lex Fridman (1:07:00.460)
and then patch it up.
David Ferrucci (1:07:01.660)
I mean, that's how people change the world.
Lex Fridman (1:07:03.220)
I mean, that's the Elon Musk approach to the rockets,
David Ferrucci (1:07:05.980)
SpaceX, that's the Henry Ford and so on.
Lex Fridman (1:07:08.780)
I love it.
Lex Fridman (1:07:09.620)
And I happen to be, in this case, I happen to be right,
Lex Fridman (1:07:11.940)
but like we didn't know.
Lex Fridman (1:07:14.300)
But you kind of have to put a stake in the rest
Lex Fridman (1:07:15.860)
of how you're going to run the project.
Lex Fridman (1:07:17.380)
So yeah, and backtracking to search.
Lex Fridman (1:07:20.340)
So if you were to do, what's the brute force solution?
Lex Fridman (1:07:24.660)
What would you search over?
Lex Fridman (1:07:26.100)
So you have a question,
Lex Fridman (1:07:27.940)
how would you search the possible space of answers?
Lex Fridman (1:07:31.300)
Look, web search has come a long way even since then.
Lex Fridman (1:07:34.820)
But at the time, first of all,
Lex Fridman (1:07:37.820)
I mean, there were a couple of other constraints
David Ferrucci (1:07:39.260)
around the problem, which is interesting.
Lex Fridman (1:07:40.940)
So you couldn't go out to the web.
David Ferrucci (1:07:43.100)
You couldn't search the internet.
Lex Fridman (1:07:44.980)
In other words, the AI experiment was,
David Ferrucci (1:07:47.600)
we want a self contained device.
Lex Fridman (1:07:50.460)
If the device is as big as a room, fine,
David Ferrucci (1:07:52.940)
it's as big as a room,
Lex Fridman (1:07:53.860)
but we want a self contained device.
David Ferrucci (1:07:57.980)
You're not going out to the internet.
Lex Fridman (1:07:59.260)
You don't have a lifeline to anything.
Lex Fridman (1:08:01.580)
So it had to kind of fit in a shoe box, if you will,
Lex Fridman (1:08:04.280)
or at least a size of a few refrigerators,
David Ferrucci (1:08:06.600)
whatever it might be.
Lex Fridman (1:08:08.060)
See, but also you couldn't just get out there.
David Ferrucci (1:08:10.440)
You couldn't go off network, right, to kind of go.
Lex Fridman (1:08:13.060)
So there was that limitation.
Lex Fridman (1:08:14.920)
But then we did, but the basic thing was go do web search.
Lex Fridman (1:08:19.340)
Problem was, even when we went and did a web search,
David Ferrucci (1:08:22.940)
I don't remember exactly the numbers,
Lex Fridman (1:08:24.540)
but somewhere in the order of 65% of the time,
David Ferrucci (1:08:27.580)
the answer would be somewhere, you know,
Lex Fridman (1:08:30.300)
in the top 10 or 20 documents.
Lex Fridman (1:08:32.900)
So first of all, that's not even good enough to play Jeopardy.
Lex Fridman (1:08:36.260)
You know, the words, even if you could pull the,
David Ferrucci (1:08:38.180)
even if you could perfectly pull the answer
Lex Fridman (1:08:40.240)
out of the top 20 documents, top 10 documents,
David Ferrucci (1:08:42.920)
whatever it was, which we didn't know how to do.
Lex Fridman (1:08:45.240)
But even if you could do that, you'd be,
Lex Fridman (1:08:47.940)
and you knew it was right,
Lex Fridman (1:08:49.140)
unless you had enough confidence in it, right?
Lex Fridman (1:08:50.700)
So you'd have to pull out the right answer.
Lex Fridman (1:08:52.180)
You'd have to have confidence it was the right answer.
Lex Fridman (1:08:54.820)
And then you'd have to do that fast enough to now go buzz in
Lex Fridman (1:08:58.100)
and you'd still only get 65% of them right,
David Ferrucci (1:09:00.300)
which doesn't even put you in the winner's circle.
Lex Fridman (1:09:02.660)
Winner's circle, you have to be up over 70
Lex Fridman (1:09:05.060)
and you have to do it really quick
Lex Fridman (1:09:06.060)
and you have to do it really quickly.
Lex Fridman (1:09:08.020)
But now the problem is, well,
Lex Fridman (1:09:10.100)
even if I had somewhere in the top 10 documents,
Lex Fridman (1:09:12.500)
how do I figure out where in the top 10 documents
Lex Fridman (1:09:14.980)
that answer is and how do I compute a confidence
Lex Fridman (1:09:18.040)
of all the possible candidates?
Lex Fridman (1:09:19.740)
So it's not like I go in knowing the right answer
Lex Fridman (1:09:21.820)
and I have to pick it.
Lex Fridman (1:09:22.660)
I don't know the right answer.
David Ferrucci (1:09:23.940)
I have a bunch of documents,
Lex Fridman (1:09:25.580)
somewhere in there is the right answer.
Lex Fridman (1:09:27.100)
How do I, as a machine, go out
Lex Fridman (1:09:28.700)
and figure out which one's right?
Lex Fridman (1:09:30.020)
And then how do I score it?
Lex Fridman (1:09:32.640)
So, and now how do I deal with the fact
Lex Fridman (1:09:35.300)
that I can't actually go out to the web?
Lex Fridman (1:09:37.320)
First of all, if you pause on that, just think about it.
David Ferrucci (1:09:40.020)
If you could go to the web,
Lex Fridman (1:09:42.160)
do you think that problem is solvable
Lex Fridman (1:09:44.260)
if you just pause on it?
Lex Fridman (1:09:45.540)
Just thinking even beyond jeopardy,
Lex Fridman (1:09:49.220)
do you think the problem of reading text
Lex Fridman (1:09:51.340)
defined where the answer is?
David Ferrucci (1:09:53.660)
Well, we solved that in some definition of solves
Lex Fridman (1:09:56.700)
given the jeopardy challenge.
Lex Fridman (1:09:58.020)
How did you do it for jeopardy?
Lex Fridman (1:09:59.020)
So how do you take a body of work in a particular topic
Lex Fridman (1:10:03.260)
and extract the key pieces of information?
Lex Fridman (1:10:05.940)
So now forgetting about the huge volumes
Lex Fridman (1:10:09.100)
that are on the web, right?
Lex Fridman (1:10:10.060)
So now we have to figure out,
David Ferrucci (1:10:11.260)
we did a lot of source research.
Lex Fridman (1:10:12.740)
In other words, what body of knowledge
David Ferrucci (1:10:15.720)
is gonna be small enough,
Lex Fridman (1:10:17.180)
but broad enough to answer jeopardy?
Lex Fridman (1:10:19.820)
And we ultimately did find the body of knowledge
Lex Fridman (1:10:21.920)
that did that.
David Ferrucci (1:10:22.760)
I mean, it included Wikipedia and a bunch of other stuff.
Lex Fridman (1:10:25.100)
So like encyclopedia type of stuff.
David Ferrucci (1:10:26.700)
I don't know if you can speak to it.
Lex Fridman (1:10:27.540)
Encyclopedias, dictionaries,
David Ferrucci (1:10:28.500)
different types of semantic resources,
Lex Fridman (1:10:31.140)
like WordNet and other types of semantic resources like that,
David Ferrucci (1:10:33.980)
as well as like some web crawls.
Lex Fridman (1:10:36.060)
In other words, where we went out and took that content
Lex Fridman (1:10:39.060)
and then expanded it based on producing,
Lex Fridman (1:10:41.700)
statistically producing seeds,
David Ferrucci (1:10:44.620)
using those seeds for other searches and then expanding that.
Lex Fridman (1:10:48.740)
So using these like expansion techniques,
David Ferrucci (1:10:51.500)
we went out and had found enough content
Lex Fridman (1:10:53.580)
and we're like, okay, this is good.
Lex Fridman (1:10:54.620)
And even up until the end,
Lex Fridman (1:10:56.980)
we had a thread of research.
David Ferrucci (1:10:58.380)
It was always trying to figure out
Lex Fridman (1:10:59.780)
what content could we efficiently include.
David Ferrucci (1:11:02.220)
I mean, there's a lot of popular,
Lex Fridman (1:11:03.420)
like what is the church lady?
David Ferrucci (1:11:05.420)
Well, I think was one of the, like what,
Lex Fridman (1:11:09.660)
where do you, I guess that's probably an encyclopedia, so.
Lex Fridman (1:11:12.380)
So that was an encyclopedia,
Lex Fridman (1:11:13.900)
but then we would take that stuff
Lex Fridman (1:11:16.060)
and we would go out and we would expand.
Lex Fridman (1:11:17.780)
In other words, we'd go find other content
David Ferrucci (1:11:20.140)
that wasn't in the core resources and expand it.
Lex Fridman (1:11:23.300)
The amount of content, we grew it by an order of magnitude,
Lex Fridman (1:11:26.180)
but still, again, from a web scale perspective,
Lex Fridman (1:11:28.580)
this is very small amount of content.
David Ferrucci (1:11:30.540)
It's very select.
Lex Fridman (1:11:31.380)
We then took all that content,
David Ferrucci (1:11:33.100)
we preanalyzed the crap out of it,
Lex Fridman (1:11:35.220)
meaning we parsed it,
David Ferrucci (1:11:38.500)
broke it down into all those individual words
Lex Fridman (1:11:40.700)
and then we did semantic,
David Ferrucci (1:11:42.180)
syntactic and semantic parses on it,
Lex Fridman (1:11:44.620)
had computer algorithms that annotated it
Lex Fridman (1:11:46.980)
and we indexed that in a very rich and very fast index.
Lex Fridman (1:11:53.140)
So we have a relatively huge amount of,
David Ferrucci (1:11:55.260)
let's say the equivalent of, for the sake of argument,
Lex Fridman (1:11:57.420)
two to 5 million bucks.
David Ferrucci (1:11:58.980)
We've now analyzed all that, blowing up its size even more
Lex Fridman (1:12:01.820)
because now we have all this metadata
Lex Fridman (1:12:03.620)
and then we richly indexed all of that
Lex Fridman (1:12:05.660)
and by the way, in a giant in memory cache.
Lex Fridman (1:12:08.940)
So Watson did not go to disk.
Lex Fridman (1:12:11.980)
So the infrastructure component there,
David Ferrucci (1:12:13.660)
if you could just speak to it, how tough it,
Lex Fridman (1:12:15.860)
I mean, I know 2000, maybe this is 2008, nine,
David Ferrucci (1:12:22.900)
that's kind of a long time ago.
Lex Fridman (1:12:25.900)
How hard is it to use multiple machines?
Lex Fridman (1:12:28.260)
How hard is the infrastructure component,
Lex Fridman (1:12:29.900)
the hardware component?
Lex Fridman (1:12:31.620)
So we used IBM hardware.
Lex Fridman (1:12:33.860)
We had something like, I forgot exactly,
Lex Fridman (1:12:36.100)
but close to 3000 cores completely connected.
Lex Fridman (1:12:40.740)
So you had a switch where every CPU
David Ferrucci (1:12:42.780)
was connected to every other CPU.
Lex Fridman (1:12:43.620)
And they were sharing memory in some kind of way.
Lex Fridman (1:12:46.100)
Large shared memory, right?
Lex Fridman (1:12:47.980)
And all this data was preanalyzed
Lex Fridman (1:12:50.740)
and put into a very fast indexing structure
Lex Fridman (1:12:54.860)
that was all in memory.
Lex Fridman (1:12:58.300)
And then we took that question,
Lex Fridman (1:13:02.780)
we would analyze the question.
Lex Fridman (1:13:04.380)
So all the content was now preanalyzed.
Lex Fridman (1:13:07.180)
So if I went and tried to find a piece of content,
David Ferrucci (1:13:10.820)
it would come back with all the metadata
Lex Fridman (1:13:12.540)
that we had precomputed.
Lex Fridman (1:13:14.580)
How do you shove that question?
Lex Fridman (1:13:16.940)
How do you connect the big knowledge base
David Ferrucci (1:13:20.820)
with the metadata and that's indexed
Lex Fridman (1:13:22.660)
to the simple little witty confusing question?
David Ferrucci (1:13:26.940)
Right.
Lex Fridman (1:13:27.780)
So therein lies the Watson architecture, right?
Lex Fridman (1:13:31.300)
So we would take the question,
Lex Fridman (1:13:32.940)
we would analyze the question.
Lex Fridman (1:13:34.700)
So which means that we would parse it
Lex Fridman (1:13:37.020)
and interpret it a bunch of different ways.
Lex Fridman (1:13:38.740)
We'd try to figure out what is it asking about?
Lex Fridman (1:13:40.820)
So we had multiple strategies
David Ferrucci (1:13:44.420)
to kind of determine what was it asking for.
Lex Fridman (1:13:47.100)
That might be represented as a simple string,
David Ferrucci (1:13:49.460)
a character string,
Lex Fridman (1:13:51.420)
or something we would connect back
David Ferrucci (1:13:53.140)
to different semantic types
Lex Fridman (1:13:54.820)
that were from existing resources.
Lex Fridman (1:13:56.100)
So anyway, the bottom line is
Lex Fridman (1:13:57.820)
we would do a bunch of analysis in the question.
Lex Fridman (1:14:00.420)
And question analysis had to finish and had to finish fast.
Lex Fridman (1:14:04.220)
So we do the question analysis
David Ferrucci (1:14:05.340)
because then from the question analysis,
Lex Fridman (1:14:07.900)
we would now produce searches.
Lex Fridman (1:14:09.780)
So we would, and we had built
Lex Fridman (1:14:12.700)
using open source search engines,
David Ferrucci (1:14:14.260)
we modified them,
Lex Fridman (1:14:16.100)
but we had a number of different search engines
David Ferrucci (1:14:17.940)
we would use that had different characteristics.
Lex Fridman (1:14:20.740)
We went in there and engineered
Lex Fridman (1:14:22.540)
and modified those search engines,
Lex Fridman (1:14:24.540)
ultimately to now take our question analysis,
David Ferrucci (1:14:28.500)
produce multiple queries
Lex Fridman (1:14:29.900)
based on different interpretations of the question
Lex Fridman (1:14:33.300)
and fire out a whole bunch of searches in parallel.
Lex Fridman (1:14:36.460)
And they would come back with passages.
Lex Fridman (1:14:39.820)
So these are passive search algorithms.
Lex Fridman (1:14:42.060)
They would come back with passages.
Lex Fridman (1:14:43.700)
And so now let's say you had a thousand passages.
Lex Fridman (1:14:47.140)
Now for each passage, you parallelize again.
Lex Fridman (1:14:50.700)
So you went out and you parallelize the search.
Lex Fridman (1:14:55.220)
Each search would now come back
David Ferrucci (1:14:56.460)
with a whole bunch of passages.
Lex Fridman (1:14:58.580)
Maybe you had a total of a thousand
David Ferrucci (1:15:00.460)
or 5,000 whatever passages.
Lex Fridman (1:15:02.220)
For each passage now,
David Ferrucci (1:15:03.900)
you'd go and figure out whether or not
Lex Fridman (1:15:05.220)
there was a candidate,
David Ferrucci (1:15:06.620)
we'd call it candidate answer in there.
Lex Fridman (1:15:08.540)
So you had a whole bunch of other algorithms
David Ferrucci (1:15:11.540)
that would find candidate answers,
Lex Fridman (1:15:13.220)
possible answers to the question.
Lex Fridman (1:15:15.620)
And so you had candidate answer,
Lex Fridman (1:15:17.620)
called candidate answer generators,
David Ferrucci (1:15:19.540)
a whole bunch of those.
Lex Fridman (1:15:20.780)
So for every one of these components,
David Ferrucci (1:15:23.100)
the team was constantly doing research coming up,
Lex Fridman (1:15:25.620)
better ways to generate search queries from the questions,
David Ferrucci (1:15:28.220)
better ways to analyze the question,
Lex Fridman (1:15:29.940)
better ways to generate candidates.
Lex Fridman (1:15:31.420)
And speed, so better is accuracy and speed.
Lex Fridman (1:15:35.380)
Correct, so right, speed and accuracy
David Ferrucci (1:15:38.100)
for the most part were separated.
Lex Fridman (1:15:40.500)
We handle that sort of in separate ways.
David Ferrucci (1:15:42.180)
Like I focus purely on accuracy, end to end accuracy.
Lex Fridman (1:15:45.180)
Are we ultimately getting more questions
Lex Fridman (1:15:46.900)
and producing more accurate confidences?
Lex Fridman (1:15:48.860)
And then a whole nother team
David Ferrucci (1:15:50.180)
that was constantly analyzing the workflow
Lex Fridman (1:15:52.420)
to find the bottlenecks.
Lex Fridman (1:15:53.780)
And then figuring out how to both parallelize
Lex Fridman (1:15:55.740)
and drive the algorithm speed.
Lex Fridman (1:15:58.060)
But anyway, so now think of it like,
Lex Fridman (1:15:59.980)
you have this big fan out now, right?
David Ferrucci (1:16:01.700)
Because you had multiple queries,
Lex Fridman (1:16:03.620)
now you have thousands of candidate answers.
David Ferrucci (1:16:06.940)
For each candidate answer, you're gonna score it.
Lex Fridman (1:16:09.980)
So you're gonna use all the data that built up.
David Ferrucci (1:16:12.420)
You're gonna use the question analysis,
Lex Fridman (1:16:15.460)
you're gonna use how the query was generated,
David Ferrucci (1:16:17.580)
you're gonna use the passage itself,
Lex Fridman (1:16:19.820)
and you're gonna use the candidate answer
David Ferrucci (1:16:21.620)
that was generated, and you're gonna score that.
Lex Fridman (1:16:25.460)
So now we have a group of researchers
David Ferrucci (1:16:28.020)
coming up with scores.
Lex Fridman (1:16:30.100)
There are hundreds of different scores.
Lex Fridman (1:16:32.300)
So now you're getting a fan out of it again
Lex Fridman (1:16:34.580)
from however many candidate answers you have
David Ferrucci (1:16:37.380)
to all the different scores.
Lex Fridman (1:16:39.260)
So if you have 200 different scores
Lex Fridman (1:16:41.260)
and you have a thousand candidates,
Lex Fridman (1:16:42.420)
now you have 200,000 scores.
Lex Fridman (1:16:45.180)
And so now you gotta figure out,
Lex Fridman (1:16:48.060)
how do I now rank these answers
Lex Fridman (1:16:52.340)
based on the scores that came back?
Lex Fridman (1:16:54.460)
And I wanna rank them based on the likelihood
David Ferrucci (1:16:56.300)
that they're a correct answer to the question.
Lex Fridman (1:16:58.620)
So every scorer was its own research project.
Lex Fridman (1:17:01.380)
What do you mean by scorer?
Lex Fridman (1:17:02.340)
So is that the annotation process
David Ferrucci (1:17:04.060)
of basically a human being saying that this answer
Lex Fridman (1:17:07.700)
has a quality of?
David Ferrucci (1:17:09.340)
Think of it, if you wanna think of it,
Lex Fridman (1:17:10.740)
what you're doing, you know,
David Ferrucci (1:17:12.580)
if you wanna think about what a human would be doing,
Lex Fridman (1:17:14.060)
human would be looking at a possible answer,
David Ferrucci (1:17:17.060)
they'd be reading the, you know, Emily Dickinson,
Lex Fridman (1:17:20.860)
they'd be reading the passage in which that occurred,
David Ferrucci (1:17:23.540)
they'd be looking at the question,
Lex Fridman (1:17:25.340)
and they'd be making a decision of how likely it is
David Ferrucci (1:17:28.340)
that Emily Dickinson, given this evidence in this passage,
Lex Fridman (1:17:32.260)
is the right answer to that question.
David Ferrucci (1:17:33.900)
Got it.
Lex Fridman (1:17:34.740)
So that's the annotation task.
David Ferrucci (1:17:36.180)
That's the annotation process.
Lex Fridman (1:17:37.020)
That's the scoring task.
Lex Fridman (1:17:38.780)
But scoring implies zero to one kind of continuous.
Lex Fridman (1:17:41.260)
That's right.
David Ferrucci (1:17:42.100)
You give it a zero to one score.
Lex Fridman (1:17:42.940)
So it's not a binary.
David Ferrucci (1:17:44.380)
No, you give it a score.
Lex Fridman (1:17:46.860)
Give it a zero to, yeah, exactly, zero to one score.
Lex Fridman (1:17:48.700)
But humans give different scores,
Lex Fridman (1:17:50.500)
so you have to somehow normalize and all that kind of stuff
David Ferrucci (1:17:52.940)
that deal with all that complexity.
Lex Fridman (1:17:54.260)
It depends on what your strategy is.
David Ferrucci (1:17:55.740)
We both, we...
Lex Fridman (1:17:57.060)
It could be relative, too.
David Ferrucci (1:17:58.060)
It could be...
Lex Fridman (1:17:59.420)
We actually looked at the raw scores
David Ferrucci (1:18:01.580)
as well as standardized scores,
Lex Fridman (1:18:02.780)
because humans are not involved in this.
David Ferrucci (1:18:04.980)
Humans are not involved.
Lex Fridman (1:18:05.900)
Sorry, so I'm misunderstanding the process here.
David Ferrucci (1:18:08.700)
This is passages.
Lex Fridman (1:18:10.460)
Where is the ground truth coming from?
David Ferrucci (1:18:13.300)
Ground truth is only the answers to the questions.
Lex Fridman (1:18:16.820)
So it's end to end.
David Ferrucci (1:18:17.940)
It's end to end.
Lex Fridman (1:18:19.020)
So I was always driving end to end performance.
David Ferrucci (1:18:22.420)
It's a very interesting, a very interesting
Lex Fridman (1:18:25.540)
engineering approach,
Lex Fridman (1:18:27.900)
and ultimately scientific research approach,
Lex Fridman (1:18:30.140)
always driving end to end.
David Ferrucci (1:18:31.300)
Now, that's not to say
Lex Fridman (1:18:34.780)
we wouldn't make hypotheses
David Ferrucci (1:18:38.620)
that individual component performance
Lex Fridman (1:18:42.180)
was related in some way to end to end performance.
David Ferrucci (1:18:44.460)
Of course we would,
Lex Fridman (1:18:45.300)
because people would have to build individual components.
Lex Fridman (1:18:48.940)
But ultimately, to get your component integrated
Lex Fridman (1:18:51.340)
to the system, you have to show impact
David Ferrucci (1:18:53.540)
on end to end performance, question answering performance.
Lex Fridman (1:18:55.980)
So there's many very smart people working on this,
Lex Fridman (1:18:58.420)
and they're basically trying to sell their ideas
Lex Fridman (1:19:01.540)
as a component that should be part of the system.
David Ferrucci (1:19:03.460)
That's right.
Lex Fridman (1:19:04.620)
And they would do research on their component,
Lex Fridman (1:19:07.340)
and they would say things like,
Lex Fridman (1:19:09.780)
I'm gonna improve this as a candidate generator,
David Ferrucci (1:19:13.140)
or I'm gonna improve this as a question score,
Lex Fridman (1:19:15.860)
or as a passive scorer,
David Ferrucci (1:19:17.780)
I'm gonna improve this, or as a parser,
Lex Fridman (1:19:20.700)
and I can improve it by 2% on its component metric,
David Ferrucci (1:19:25.500)
like a better parse, or a better candidate,
Lex Fridman (1:19:27.940)
or a better type estimation, whatever it is.
Lex Fridman (1:19:30.260)
And then I would say,
Lex Fridman (1:19:31.700)
I need to understand how the improvement
David Ferrucci (1:19:33.900)
on that component metric
Lex Fridman (1:19:35.340)
is gonna affect the end to end performance.
David Ferrucci (1:19:37.740)
If you can't estimate that,
Lex Fridman (1:19:39.460)
and can't do experiments to demonstrate that,
David Ferrucci (1:19:41.860)
it doesn't get in.
Lex Fridman (1:19:43.420)
That's like the best run AI project I've ever heard.
David Ferrucci (1:19:47.540)
That's awesome.
Lex Fridman (1:19:48.380)
Okay, what breakthrough would you say,
David Ferrucci (1:19:51.780)
like, I'm sure there's a lot of day to day breakthroughs,
Lex Fridman (1:19:54.260)
but was there like a breakthrough
Lex Fridman (1:19:55.620)
that really helped improve performance?
Lex Fridman (1:19:57.860)
Like where people began to believe,
Lex Fridman (1:20:01.140)
or is it just a gradual process?
Lex Fridman (1:20:02.500)
Well, I think it was a gradual process,
Lex Fridman (1:20:04.500)
but one of the things that I think gave people confidence
Lex Fridman (1:20:08.980)
that we can get there was that,
David Ferrucci (1:20:11.620)
as we follow this procedure of different ideas,
Lex Fridman (1:20:16.620)
build different components,
David Ferrucci (1:20:19.140)
plug them into the architecture,
Lex Fridman (1:20:20.460)
run the system, see how we do,
David Ferrucci (1:20:23.180)
do the error analysis,
Lex Fridman (1:20:24.700)
start off new research projects to improve things.
Lex Fridman (1:20:28.140)
And the very important idea
Lex Fridman (1:20:31.260)
that the individual component work
David Ferrucci (1:20:37.420)
did not have to deeply understand everything
Lex Fridman (1:20:40.020)
that was going on with every other component.
Lex Fridman (1:20:42.220)
And this is where we leverage machine learning
Lex Fridman (1:20:45.060)
in a very important way.
Lex Fridman (1:20:47.380)
So while individual components
Lex Fridman (1:20:48.780)
could be statistically driven machine learning components,
David Ferrucci (1:20:51.260)
some of them were heuristic,
Lex Fridman (1:20:52.740)
some of them were machine learning components,
David Ferrucci (1:20:54.580)
the system has a whole combined all the scores
Lex Fridman (1:20:58.100)
using machine learning.
David Ferrucci (1:21:00.540)
This was critical because that way
Lex Fridman (1:21:02.700)
you can divide and conquer.
Lex Fridman (1:21:04.340)
So you can say, okay, you work on your candidate generator,
Lex Fridman (1:21:07.500)
or you work on this approach to answer scoring,
David Ferrucci (1:21:09.740)
you work on this approach to type scoring,
Lex Fridman (1:21:11.740)
you work on this approach to passage search
David Ferrucci (1:21:14.500)
or to pass a selection and so forth.
Lex Fridman (1:21:17.340)
But when we just plug it in,
Lex Fridman (1:21:19.580)
and we had enough training data to say,
Lex Fridman (1:21:22.020)
now we can train and figure out
Lex Fridman (1:21:24.540)
how do we weigh all the scores relative to each other
Lex Fridman (1:21:29.300)
based on the predicting the outcome,
David Ferrucci (1:21:31.900)
which is right or wrong on Jeopardy.
Lex Fridman (1:21:33.820)
And we had enough training data to do that.
Lex Fridman (1:21:36.780)
So this enabled people to work independently
Lex Fridman (1:21:40.500)
and to let the machine learning do the integration.
David Ferrucci (1:21:43.340)
Beautiful, so yeah, the machine learning
Lex Fridman (1:21:45.100)
is doing the fusion,
Lex Fridman (1:21:46.340)
and then it's a human orchestrated ensemble
Lex Fridman (1:21:48.980)
of different approaches.
David Ferrucci (1:21:50.380)
That's great.
Lex Fridman (1:21:53.420)
Still impressive that you were able
David Ferrucci (1:21:54.940)
to get it done in a few years.
Lex Fridman (1:21:57.620)
That's not obvious to me that it's doable,
David Ferrucci (1:22:00.420)
if I just put myself in that mindset.
Lex Fridman (1:22:03.340)
But when you look back at the Jeopardy challenge,
David Ferrucci (1:22:07.820)
again, when you're looking up at the stars,
Lex Fridman (1:22:10.220)
what are you most proud of, looking back at those days?
David Ferrucci (1:22:17.420)
I'm most proud of my,
Lex Fridman (1:22:27.900)
my commitment and my team's commitment
David Ferrucci (1:22:32.260)
to be true to the science,
Lex Fridman (1:22:35.060)
to not be afraid to fail.
David Ferrucci (1:22:38.020)
That's beautiful because there's so much pressure,
Lex Fridman (1:22:41.540)
because it is a public event, it is a public show,
David Ferrucci (1:22:44.380)
that you were dedicated to the idea.
Lex Fridman (1:22:46.980)
That's right.
Lex Fridman (1:22:50.460)
Do you think it was a success?
Lex Fridman (1:22:53.140)
In the eyes of the world, it was a success.
David Ferrucci (1:22:56.620)
By your, I'm sure, exceptionally high standards,
Lex Fridman (1:23:00.860)
is there something you regret you would do differently?
David Ferrucci (1:23:03.700)
It was a success.
Lex Fridman (1:23:05.900)
It was a success for our goal.
David Ferrucci (1:23:08.180)
Our goal was to build the most advanced
Lex Fridman (1:23:11.340)
open domain question answering system.
David Ferrucci (1:23:14.700)
We went back to the old problems
Lex Fridman (1:23:16.420)
that we used to try to solve,
Lex Fridman (1:23:17.900)
and we did dramatically better on all of them,
Lex Fridman (1:23:21.220)
as well as we beat Jeopardy.
Lex Fridman (1:23:24.140)
So we won at Jeopardy.
Lex Fridman (1:23:25.780)
So it was a success.
David Ferrucci (1:23:28.460)
It was, I worry that the community
Lex Fridman (1:23:32.540)
or the world would not understand it as a success
David Ferrucci (1:23:36.100)
because it came down to only one game.
Lex Fridman (1:23:38.660)
And I knew statistically speaking,
David Ferrucci (1:23:40.380)
this can be a huge technical success,
Lex Fridman (1:23:42.260)
and we could still lose that one game.
Lex Fridman (1:23:43.820)
And that's a whole nother theme of this, of the journey.
Lex Fridman (1:23:47.220)
But it was a success.
David Ferrucci (1:23:50.260)
It was not a success in natural language understanding,
Lex Fridman (1:23:53.620)
but that was not the goal.
David Ferrucci (1:23:56.620)
Yeah, that was, but I would argue,
Lex Fridman (1:24:00.700)
I understand what you're saying
David Ferrucci (1:24:02.020)
in terms of the science,
Lex Fridman (1:24:04.140)
but I would argue that the inspiration of it, right?
David Ferrucci (1:24:07.540)
The, not a success in terms of solving
Lex Fridman (1:24:11.180)
natural language understanding.
David Ferrucci (1:24:12.820)
There was a success of being an inspiration
Lex Fridman (1:24:16.300)
to future challenges.
David Ferrucci (1:24:17.900)
Absolutely.
Lex Fridman (1:24:18.820)
That drive future efforts.
David Ferrucci (1:24:21.140)
What's the difference between how human being
Lex Fridman (1:24:23.740)
compete in Jeopardy and how Watson does it?
David Ferrucci (1:24:26.860)
That's important in terms of intelligence.
Lex Fridman (1:24:28.740)
Yeah, so that actually came up very early on
David Ferrucci (1:24:31.380)
in the project also.
Lex Fridman (1:24:32.620)
In fact, I had people who wanted to be on the project
David Ferrucci (1:24:35.180)
who were early on, who sort of approached me
Lex Fridman (1:24:39.100)
once I committed to do it,
David Ferrucci (1:24:42.380)
had wanted to think about how humans do it.
Lex Fridman (1:24:44.300)
And they were, from a cognition perspective,
David Ferrucci (1:24:47.060)
like human cognition and how that should play.
Lex Fridman (1:24:49.900)
And I would not take them on the project
David Ferrucci (1:24:52.180)
because another assumption or another stake
Lex Fridman (1:24:55.780)
I put in the ground was,
David Ferrucci (1:24:57.620)
I don't really care how humans do this.
Lex Fridman (1:25:00.180)
At least in the context of this project.
David Ferrucci (1:25:01.540)
I need to build in the context of this project.
Lex Fridman (1:25:03.900)
In NLU and in building an AI that understands
Lex Fridman (1:25:06.980)
how it needs to ultimately communicate with humans,
Lex Fridman (1:25:09.660)
I very much care.
Lex Fridman (1:25:11.260)
So it wasn't that I didn't care in general.
Lex Fridman (1:25:16.540)
In fact, as an AI scientist, I care a lot about that,
Lex Fridman (1:25:20.780)
but I'm also a practical engineer
Lex Fridman (1:25:22.620)
and I committed to getting this thing done
Lex Fridman (1:25:25.540)
and I wasn't gonna get distracted.
Lex Fridman (1:25:27.500)
I had to kind of say, like, if I'm gonna get this done,
David Ferrucci (1:25:30.740)
I'm gonna chart this path.
Lex Fridman (1:25:31.740)
And this path says, we're gonna engineer a machine
David Ferrucci (1:25:35.980)
that's gonna get this thing done.
Lex Fridman (1:25:37.540)
And we know what search and NLP can do.
David Ferrucci (1:25:41.500)
We have to build on that foundation.
Lex Fridman (1:25:44.140)
If I come in and take a different approach
Lex Fridman (1:25:46.260)
and start wondering about how the human mind
Lex Fridman (1:25:48.060)
might or might not do this,
David Ferrucci (1:25:49.700)
I'm not gonna get there from here in the timeframe.
Lex Fridman (1:25:54.380)
I think that's a great way to lead the team.
Lex Fridman (1:25:56.620)
But now that it's done and there's one,
Lex Fridman (1:25:59.180)
when you look back, analyze what's the difference actually.
Lex Fridman (1:26:03.540)
So I was a little bit surprised actually
Lex Fridman (1:26:05.460)
to discover over time, as this would come up
David Ferrucci (1:26:09.020)
from time to time and we'd reflect on it,
Lex Fridman (1:26:13.300)
and talking to Ken Jennings a little bit
Lex Fridman (1:26:14.980)
and hearing Ken Jennings talk about
Lex Fridman (1:26:16.780)
how he answered questions,
David Ferrucci (1:26:18.860)
that it might've been closer to the way humans
Lex Fridman (1:26:21.260)
answer questions than I might've imagined previously.
David Ferrucci (1:26:24.700)
Because humans are probably in the game of Jeopardy!
Lex Fridman (1:26:27.860)
at the level of Ken Jennings,
Lex Fridman (1:26:29.500)
are probably also cheating their way to winning, right?
Lex Fridman (1:26:35.180)
Not cheating, but shallow.
David Ferrucci (1:26:36.020)
Well, they're doing shallow analysis.
Lex Fridman (1:26:37.180)
They're doing the fastest possible.
David Ferrucci (1:26:39.340)
They're doing shallow analysis.
Lex Fridman (1:26:40.860)
So they are very quickly analyzing the question
Lex Fridman (1:26:44.820)
and coming up with some key vectors or cues, if you will.
Lex Fridman (1:26:49.980)
And they're taking those cues
Lex Fridman (1:26:51.060)
and they're very quickly going through
Lex Fridman (1:26:52.540)
like their library of stuff,
David Ferrucci (1:26:54.900)
not deeply reasoning about what's going on.
Lex Fridman (1:26:57.700)
And then sort of like a lots of different,
David Ferrucci (1:27:00.580)
like what we would call these scores,
Lex Fridman (1:27:03.180)
would kind of score that in a very shallow way
Lex Fridman (1:27:06.060)
and then say, oh, boom, you know, that's what it is.
Lex Fridman (1:27:08.900)
And so it's interesting as we reflected on that.
Lex Fridman (1:27:12.420)
So we may be doing something that's not too far off
Lex Fridman (1:27:16.060)
from the way humans do it,
Lex Fridman (1:27:17.220)
but we certainly didn't approach it by saying,
Lex Fridman (1:27:21.420)
how would a human do this?
David Ferrucci (1:27:22.660)
Now in elemental cognition,
Lex Fridman (1:27:24.620)
like the project I'm leading now,
David Ferrucci (1:27:27.300)
we ask those questions all the time
Lex Fridman (1:27:28.740)
because ultimately we're trying to do something
David Ferrucci (1:27:31.660)
that is to make the intelligence of the machine
Lex Fridman (1:27:35.060)
and the intelligence of the human very compatible.
David Ferrucci (1:27:37.740)
Well, compatible in the sense
Lex Fridman (1:27:38.660)
they can communicate with one another
Lex Fridman (1:27:40.940)
and they can reason with this shared understanding.
Lex Fridman (1:27:44.540)
So how they think about things and how they build answers,
Lex Fridman (1:27:48.020)
how they build explanations
Lex Fridman (1:27:49.740)
becomes a very important question to consider.
Lex Fridman (1:27:52.100)
So what's the difference between this open domain,
Lex Fridman (1:27:56.900)
but cold constructed question answering of Jeopardy
Lex Fridman (1:28:02.340)
and more something that requires understanding
Lex Fridman (1:28:07.380)
for shared communication with humans and machines?
David Ferrucci (1:28:10.220)
Yeah, well, this goes back to the interpretation
Lex Fridman (1:28:13.300)
of what we were talking about before.
David Ferrucci (1:28:15.540)
Jeopardy, the system's not trying to interpret the question
Lex Fridman (1:28:19.140)
and it's not interpreting the content it's reusing
David Ferrucci (1:28:22.060)
with regard to any particular framework.
Lex Fridman (1:28:23.860)
I mean, it is parsing it and parsing the content
Lex Fridman (1:28:26.900)
and using grammatical cues and stuff like that.
Lex Fridman (1:28:29.460)
So if you think of grammar as a human framework
David Ferrucci (1:28:31.700)
in some sense, it has that,
Lex Fridman (1:28:33.420)
but when you get into the richer semantic frameworks,
Lex Fridman (1:28:36.900)
what do people, how do they think, what motivates them,
Lex Fridman (1:28:40.060)
what are the events that are occurring
Lex Fridman (1:28:41.620)
and why are they occurring
Lex Fridman (1:28:42.580)
and what causes what else to happen
Lex Fridman (1:28:44.420)
and where are things in time and space?
Lex Fridman (1:28:47.460)
And like when you start thinking about how humans formulate
Lex Fridman (1:28:51.260)
and structure the knowledge that they acquire in their head
Lex Fridman (1:28:54.020)
and wasn't doing any of that.
Lex Fridman (1:28:57.060)
What do you think are the essential challenges
Lex Fridman (1:29:01.500)
of like free flowing communication, free flowing dialogue
David Ferrucci (1:29:05.860)
versus question answering even with the framework
Lex Fridman (1:29:09.060)
of the interpretation dialogue?
David Ferrucci (1:29:11.260)
Yep.
Lex Fridman (1:29:12.340)
Do you see free flowing dialogue
David Ferrucci (1:29:14.980)
as a fundamentally more difficult than question answering
Lex Fridman (1:29:20.420)
even with shared interpretation?
Lex Fridman (1:29:23.580)
So dialogue is important in a number of different ways.
Lex Fridman (1:29:26.660)
I mean, it's a challenge.
Lex Fridman (1:29:27.500)
So first of all, when I think about the machine that,
Lex Fridman (1:29:30.540)
when I think about a machine that understands language
Lex Fridman (1:29:33.300)
and ultimately can reason in an objective way
Lex Fridman (1:29:36.780)
that can take the information that it perceives
David Ferrucci (1:29:40.580)
through language or other means
Lex Fridman (1:29:42.260)
and connect it back to these frameworks,
David Ferrucci (1:29:44.540)
reason and explain itself,
Lex Fridman (1:29:48.020)
that system ultimately needs to be able to talk to humans
David Ferrucci (1:29:50.700)
or it needs to be able to interact with humans.
Lex Fridman (1:29:52.940)
So in some sense it needs to dialogue.
David Ferrucci (1:29:55.180)
That doesn't mean that it,
Lex Fridman (1:29:58.660)
sometimes people talk about dialogue and they think,
David Ferrucci (1:30:01.820)
you know, how do humans talk to like,
Lex Fridman (1:30:04.300)
talk to each other in a casual conversation
Lex Fridman (1:30:07.700)
and you can mimic casual conversations.
Lex Fridman (1:30:11.780)
We're not trying to mimic casual conversations.
David Ferrucci (1:30:14.340)
We're really trying to produce a machine
Lex Fridman (1:30:17.580)
whose goal is to help you think
Lex Fridman (1:30:20.260)
and help you reason about your answers and explain why.
Lex Fridman (1:30:23.620)
So instead of like talking to your friend down the street
David Ferrucci (1:30:26.580)
about having a small talk conversation
Lex Fridman (1:30:28.900)
with your friend down the street,
David Ferrucci (1:30:30.500)
this is more about like you would be communicating
Lex Fridman (1:30:32.380)
to the computer on Star Trek
Lex Fridman (1:30:34.060)
where like, what do you wanna think about?
Lex Fridman (1:30:36.780)
Like, what do you wanna reason about?
David Ferrucci (1:30:37.620)
I'm gonna tell you the information I have.
Lex Fridman (1:30:38.780)
I'm gonna have to summarize it.
David Ferrucci (1:30:39.860)
I'm gonna ask you questions.
Lex Fridman (1:30:41.060)
You're gonna answer those questions.
David Ferrucci (1:30:42.700)
I'm gonna go back and forth with you.
Lex Fridman (1:30:44.300)
I'm gonna figure out what your mental model is.
David Ferrucci (1:30:46.620)
I'm gonna now relate that to the information I have
Lex Fridman (1:30:49.940)
and present it to you in a way that you can understand it
Lex Fridman (1:30:53.060)
and then we could ask followup questions.
Lex Fridman (1:30:54.940)
So it's that type of dialogue that you wanna construct.
David Ferrucci (1:30:58.340)
It's more structured, it's more goal oriented,
Lex Fridman (1:31:02.380)
but it needs to be fluid.
David Ferrucci (1:31:04.900)
In other words, it has to be engaging and fluid.
Lex Fridman (1:31:09.300)
It has to be productive and not distracting.
Lex Fridman (1:31:13.100)
So there has to be a model of,
Lex Fridman (1:31:15.700)
in other words, the machine has to have a model
David Ferrucci (1:31:17.580)
of how humans think through things and discuss them.
Lex Fridman (1:31:22.660)
So basically a productive, rich conversation
David Ferrucci (1:31:28.700)
unlike this podcast.
Lex Fridman (1:31:32.700)
I'd like to think it's more similar to this podcast.
David Ferrucci (1:31:34.940)
I wasn't joking.
Lex Fridman (1:31:37.020)
I'll ask you about humor as well, actually.
Lex Fridman (1:31:39.740)
But what's the hardest part of that?
Lex Fridman (1:31:43.300)
Because it seems we're quite far away
David Ferrucci (1:31:46.620)
as a community from that still to be able to,
Lex Fridman (1:31:49.820)
so one is having a shared understanding.
David Ferrucci (1:31:53.020)
That's, I think, a lot of the stuff you said
Lex Fridman (1:31:54.920)
with frameworks is quite brilliant.
Lex Fridman (1:31:57.160)
But just creating a smooth discourse.
Lex Fridman (1:32:02.740)
It feels clunky right now.
David Ferrucci (1:32:05.300)
Which aspects of this whole problem
Lex Fridman (1:32:07.620)
that you just specified of having
Lex Fridman (1:32:10.540)
a productive conversation is the hardest?
Lex Fridman (1:32:12.900)
And that we're, or maybe any aspect of it
David Ferrucci (1:32:17.700)
you can comment on because it's so shrouded in mystery.
Lex Fridman (1:32:20.780)
So I think to do this you kind of have to be creative
David Ferrucci (1:32:24.280)
in the following sense.
Lex Fridman (1:32:26.820)
If I were to do this as purely a machine learning approach
Lex Fridman (1:32:29.820)
and someone said learn how to have a good,
Lex Fridman (1:32:32.940)
fluent, structured knowledge acquisition conversation,
David Ferrucci (1:32:38.420)
I'd go out and say, okay, I have to collect
Lex Fridman (1:32:39.980)
a bunch of data of people doing that.
David Ferrucci (1:32:42.360)
People reasoning well, having a good, structured
Lex Fridman (1:32:47.100)
conversation that both acquires knowledge efficiently
David Ferrucci (1:32:50.320)
as well as produces answers and explanations
Lex Fridman (1:32:52.420)
as part of the process.
Lex Fridman (1:32:54.600)
And you struggle.
Lex Fridman (1:32:57.340)
I don't know.
David Ferrucci (1:32:58.180)
To collect the data.
Lex Fridman (1:32:59.000)
To collect the data because I don't know
Lex Fridman (1:33:00.700)
how much data is like that.
Lex Fridman (1:33:02.640)
Okay, there's one, there's a humorous commentary
David Ferrucci (1:33:06.140)
on the lack of rational discourse.
Lex Fridman (1:33:08.500)
But also even if it's out there, say it was out there,
Lex Fridman (1:33:12.700)
how do you actually annotate, like how do you collect
Lex Fridman (1:33:16.380)
an accessible example?
David Ferrucci (1:33:17.220)
Right, so I think any problem like this
Lex Fridman (1:33:19.200)
where you don't have enough data to represent
David Ferrucci (1:33:23.140)
the phenomenon you want to learn,
Lex Fridman (1:33:24.740)
in other words you want, if you have enough data
David Ferrucci (1:33:26.740)
you could potentially learn the pattern.
Lex Fridman (1:33:28.540)
In an example like this it's hard to do.
David Ferrucci (1:33:30.340)
This is sort of a human sort of thing to do.
Lex Fridman (1:33:34.420)
What recently came out at IBM was the debater projects
Lex Fridman (1:33:37.020)
and it's interesting, right, because now you do have
Lex Fridman (1:33:39.460)
these structured dialogues, these debate things
David Ferrucci (1:33:42.580)
where they did use machine learning techniques
Lex Fridman (1:33:44.700)
to generate these debates.
David Ferrucci (1:33:49.220)
Dialogues are a little bit tougher in my opinion
Lex Fridman (1:33:52.460)
than generating a structured argument
David Ferrucci (1:33:56.100)
where you have lots of other structured arguments
Lex Fridman (1:33:57.580)
like this, you could potentially annotate that data
Lex Fridman (1:33:59.540)
and you could say this is a good response,
Lex Fridman (1:34:00.820)
this is a bad response in a particular domain.
David Ferrucci (1:34:03.060)
Here I have to be responsive and I have to be opportunistic
Lex Fridman (1:34:08.900)
with regard to what is the human saying.
Lex Fridman (1:34:11.820)
So I'm goal oriented in saying I want to solve the problem,
Lex Fridman (1:34:14.900)
I want to acquire the knowledge necessary,
Lex Fridman (1:34:16.580)
but I also have to be opportunistic and responsive
Lex Fridman (1:34:19.140)
to what the human is saying.
Lex Fridman (1:34:21.040)
So I think that it's not clear that we could just train
Lex Fridman (1:34:24.060)
on the body of data to do this, but we could bootstrap it.
David Ferrucci (1:34:28.020)
In other words, we can be creative and we could say,
Lex Fridman (1:34:30.540)
what do we think the structure of a good dialogue is
Lex Fridman (1:34:34.020)
that does this well?
Lex Fridman (1:34:35.820)
And we can start to create that.
David Ferrucci (1:34:37.860)
If we can create that more programmatically,
Lex Fridman (1:34:42.100)
at least to get this process started
Lex Fridman (1:34:44.700)
and I can create a tool that now engages humans effectively,
Lex Fridman (1:34:47.980)
I could start generating data,
David Ferrucci (1:34:51.340)
I could start the human learning process
Lex Fridman (1:34:53.020)
and I can update my machine,
Lex Fridman (1:34:55.060)
but I could also start the automatic learning process
Lex Fridman (1:34:57.700)
as well, but I have to understand
Lex Fridman (1:34:59.860)
what features to even learn over.
Lex Fridman (1:35:01.860)
So I have to bootstrap the process a little bit first.
Lex Fridman (1:35:04.740)
And that's a creative design task
Lex Fridman (1:35:07.740)
that I could then use as input
David Ferrucci (1:35:11.060)
into a more automatic learning task.
Lex Fridman (1:35:13.420)
So some creativity in bootstrapping.
Lex Fridman (1:35:16.740)
What elements of a conversation
Lex Fridman (1:35:18.020)
do you think you would like to see?
Lex Fridman (1:35:21.140)
So one of the benchmarks for me is humor, right?
Lex Fridman (1:35:25.620)
That seems to be one of the hardest.
Lex Fridman (1:35:27.580)
And to me, the biggest contrast is sort of Watson.
Lex Fridman (1:35:31.340)
So one of the greatest sketches,
David Ferrucci (1:35:33.380)
comedy sketches of all time, right,
Lex Fridman (1:35:35.260)
is the SNL celebrity Jeopardy
David Ferrucci (1:35:38.580)
with Alex Trebek and Sean Connery
Lex Fridman (1:35:42.060)
and Burt Reynolds and so on,
David Ferrucci (1:35:44.060)
with Sean Connery commentating on Alex Trebek's
Lex Fridman (1:35:47.900)
while they're alive.
Lex Fridman (1:35:49.380)
And I think all of them are in the negative pointwise.
Lex Fridman (1:35:52.860)
So they're clearly all losing
David Ferrucci (1:35:55.100)
in terms of the game of Jeopardy,
Lex Fridman (1:35:56.340)
but they're winning in terms of comedy.
Lex Fridman (1:35:58.340)
So what do you think about humor in this whole interaction
Lex Fridman (1:36:03.780)
in the dialogue that's productive?
David Ferrucci (1:36:06.500)
Or even just what humor represents to me
Lex Fridman (1:36:09.780)
is the same idea that you're saying about framework,
David Ferrucci (1:36:15.420)
because humor only exists
Lex Fridman (1:36:16.420)
within a particular human framework.
Lex Fridman (1:36:18.340)
So what do you think about humor?
Lex Fridman (1:36:19.580)
What do you think about things like humor
David Ferrucci (1:36:21.540)
that connect to the kind of creativity
Lex Fridman (1:36:23.340)
you mentioned that's needed?
David Ferrucci (1:36:25.100)
I think there's a couple of things going on there.
Lex Fridman (1:36:26.380)
So I sort of feel like,
Lex Fridman (1:36:29.500)
and I might be too optimistic this way,
Lex Fridman (1:36:31.780)
but I think that there are,
David Ferrucci (1:36:34.700)
we did a little bit about with puns in Jeopardy.
Lex Fridman (1:36:39.020)
We literally sat down and said,
Lex Fridman (1:36:41.660)
how do puns work?
Lex Fridman (1:36:43.180)
And it's like wordplay,
Lex Fridman (1:36:44.820)
and you could formalize these things.
Lex Fridman (1:36:46.140)
So I think there's a lot aspects of humor
David Ferrucci (1:36:48.260)
that you could formalize.
Lex Fridman (1:36:50.220)
You could also learn humor.
Lex Fridman (1:36:51.620)
You could just say, what do people laugh at?
Lex Fridman (1:36:53.460)
And if you have enough, again,
David Ferrucci (1:36:54.860)
if you have enough data to represent the phenomenon,
Lex Fridman (1:36:56.860)
you might be able to weigh the features
Lex Fridman (1:36:59.460)
and figure out what humans find funny
Lex Fridman (1:37:01.300)
and what they don't find funny.
David Ferrucci (1:37:02.700)
The machine might not be able to explain
Lex Fridman (1:37:05.140)
why the human is funny unless we sit back
Lex Fridman (1:37:08.060)
and think about that more formally.
Lex Fridman (1:37:10.180)
I think, again, I think you do a combination of both.
Lex Fridman (1:37:12.420)
And I'm always a big proponent of that.
Lex Fridman (1:37:13.900)
I think robust architectures and approaches
David Ferrucci (1:37:16.700)
are always a little bit combination of us reflecting
Lex Fridman (1:37:19.620)
and being creative about how things are structured,
Lex Fridman (1:37:22.500)
how to formalize them,
Lex Fridman (1:37:23.780)
and then taking advantage of large data and doing learning
Lex Fridman (1:37:26.220)
and figuring out how to combine these two approaches.
Lex Fridman (1:37:29.100)
I think there's another aspect to humor though,
David Ferrucci (1:37:31.420)
which goes to the idea that I feel like I can relate
Lex Fridman (1:37:34.340)
to the person telling the story.
Lex Fridman (1:37:38.820)
And I think that's an interesting theme
Lex Fridman (1:37:42.140)
in the whole AI theme,
Lex Fridman (1:37:43.380)
which is, do I feel differently when I know it's a robot?
Lex Fridman (1:37:48.460)
And when I imagine that the robot is not conscious
David Ferrucci (1:37:52.860)
the way I'm conscious,
Lex Fridman (1:37:54.180)
when I imagine the robot does not actually
David Ferrucci (1:37:56.300)
have the experiences that I experience,
Lex Fridman (1:37:58.700)
do I find it funny?
David Ferrucci (1:38:00.980)
Or do, because it's not as related,
Lex Fridman (1:38:03.060)
I don't imagine that the person's relating it to it
David Ferrucci (1:38:06.540)
the way I relate to it.
Lex Fridman (1:38:07.860)
I think this also, you see this in the arts
Lex Fridman (1:38:11.340)
and in entertainment where,
Lex Fridman (1:38:14.260)
sometimes you have savants who are remarkable at a thing,
David Ferrucci (1:38:17.380)
whether it's sculpture or it's music or whatever,
Lex Fridman (1:38:19.820)
but the people who get the most attention
David Ferrucci (1:38:21.300)
are the people who can evoke a similar emotional response,
Lex Fridman (1:38:26.660)
who can get you to emote, right?
David Ferrucci (1:38:30.740)
About the way they are.
Lex Fridman (1:38:31.940)
In other words, who can basically make the connection
David Ferrucci (1:38:34.460)
from the artifact, from the music or the painting
Lex Fridman (1:38:37.020)
of the sculpture to the emotion
Lex Fridman (1:38:39.780)
and get you to share that emotion with them.
Lex Fridman (1:38:42.380)
And then, and that's when it becomes compelling.
Lex Fridman (1:38:44.700)
So they're communicating at a whole different level.
Lex Fridman (1:38:46.980)
They're just not communicating the artifact.
David Ferrucci (1:38:49.340)
They're communicating their emotional response
Lex Fridman (1:38:50.980)
to the artifact.
Lex Fridman (1:38:51.980)
And then you feel like, oh wow,
Lex Fridman (1:38:53.380)
I can relate to that person, I can connect to that,
David Ferrucci (1:38:55.540)
I can connect to that person.
Lex Fridman (1:38:57.140)
So I think humor has that aspect as well.
Lex Fridman (1:39:00.660)
So the idea that you can connect to that person,
Lex Fridman (1:39:04.820)
person being the critical thing,
Lex Fridman (1:39:07.260)
but we're also able to anthropomorphize objects pretty,
Lex Fridman (1:39:12.620)
robots and AI systems pretty well.
Lex Fridman (1:39:15.180)
So we're almost looking to make them human.
Lex Fridman (1:39:18.740)
So maybe from your experience with Watson,
David Ferrucci (1:39:20.820)
maybe you can comment on, did you consider that as part,
Lex Fridman (1:39:24.940)
well, obviously the problem of jeopardy
David Ferrucci (1:39:27.020)
doesn't require anthropomorphization, but nevertheless.
Lex Fridman (1:39:30.500)
Well, there was some interest in doing that.
Lex Fridman (1:39:32.300)
And that's another thing I didn't want to do
Lex Fridman (1:39:35.020)
because I didn't want to distract
David Ferrucci (1:39:36.220)
from the actual scientific task.
Lex Fridman (1:39:38.740)
But you're absolutely right.
David Ferrucci (1:39:39.620)
I mean, humans do anthropomorphize
Lex Fridman (1:39:43.380)
and without necessarily a lot of work.
David Ferrucci (1:39:45.900)
I mean, you just put some eyes
Lex Fridman (1:39:47.100)
and a couple of eyebrow movements
Lex Fridman (1:39:49.220)
and you're getting humans to react emotionally.
Lex Fridman (1:39:51.820)
And I think you can do that.
Lex Fridman (1:39:53.540)
So I didn't mean to suggest that,
Lex Fridman (1:39:56.780)
that that connection cannot be mimicked.
David Ferrucci (1:40:00.620)
I think that connection can be mimicked
Lex Fridman (1:40:02.260)
and can produce that emotional response.
David Ferrucci (1:40:07.300)
I just wonder though, if you're told what's really going on,
Lex Fridman (1:40:13.020)
if you know that the machine is not conscious,
David Ferrucci (1:40:17.180)
not having the same richness of emotional reactions
Lex Fridman (1:40:20.740)
and understanding that it doesn't really
David Ferrucci (1:40:21.980)
share the understanding,
Lex Fridman (1:40:23.380)
but it's essentially just moving its eyebrow
David Ferrucci (1:40:25.100)
or drooping its eyes or making them bigger,
Lex Fridman (1:40:27.180)
whatever it's doing, just getting the emotional response,
Lex Fridman (1:40:30.180)
will you still feel it?
Lex Fridman (1:40:31.580)
Interesting.
David Ferrucci (1:40:32.420)
I think you probably would for a while.
Lex Fridman (1:40:34.380)
And then when it becomes more important
David Ferrucci (1:40:35.860)
that there's a deeper share of understanding,
Lex Fridman (1:40:38.700)
it may run flat, but I don't know.
David Ferrucci (1:40:40.700)
I'm pretty confident that majority of the world,
Lex Fridman (1:40:45.300)
even if you tell them how it works,
David Ferrucci (1:40:47.460)
well, it will not matter,
Lex Fridman (1:40:49.100)
especially if the machine herself says that she is conscious.
David Ferrucci (1:40:55.420)
That's very possible.
Lex Fridman (1:40:56.260)
So you, the scientist that made the machine is saying
David Ferrucci (1:41:00.700)
that this is how the algorithm works.
Lex Fridman (1:41:02.860)
Everybody will just assume you're lying
Lex Fridman (1:41:04.460)
and that there's a conscious being there.
Lex Fridman (1:41:06.140)
So you're deep into the science fiction genre now,
Lex Fridman (1:41:09.220)
but yeah.
Lex Fridman (1:41:10.060)
I don't think it's, it's actually psychology.
David Ferrucci (1:41:12.020)
I think it's not science fiction.
Lex Fridman (1:41:13.780)
I think it's reality.
David Ferrucci (1:41:14.900)
I think it's a really powerful one
Lex Fridman (1:41:16.780)
that we'll have to be exploring in the next few decades.
David Ferrucci (1:41:19.980)
I agree.
Lex Fridman (1:41:20.820)
It's a very interesting element of intelligence.
Lex Fridman (1:41:23.540)
So what do you think,
Lex Fridman (1:41:25.220)
we've talked about social constructs of intelligences
Lex Fridman (1:41:28.500)
and frameworks and the way humans
Lex Fridman (1:41:31.140)
kind of interpret information.
Lex Fridman (1:41:33.940)
What do you think is a good test of intelligence
Lex Fridman (1:41:35.700)
in your view?
Lex Fridman (1:41:36.540)
So there's the Alan Turing with the Turing test.
Lex Fridman (1:41:41.300)
Watson accomplished something very impressive with Jeopardy.
Lex Fridman (1:41:44.940)
What do you think is a test
Lex Fridman (1:41:47.820)
that would impress the heck out of you
Lex Fridman (1:41:49.740)
that you saw that a computer could do?
Lex Fridman (1:41:52.980)
They would say, this is crossing a kind of threshold
David Ferrucci (1:41:57.260)
that gives me pause in a good way.
Lex Fridman (1:42:02.620)
My expectations for AI are generally high.
Lex Fridman (1:42:06.100)
What does high look like by the way?
Lex Fridman (1:42:07.420)
So not the threshold, test is a threshold.
Lex Fridman (1:42:10.380)
What do you think is the destination?
Lex Fridman (1:42:12.460)
What do you think is the ceiling?
David Ferrucci (1:42:15.780)
I think machines will in many measures
Lex Fridman (1:42:18.460)
will be better than us, will become more effective.
David Ferrucci (1:42:21.660)
In other words, better predictors about a lot of things
Lex Fridman (1:42:25.140)
than ultimately we can do.
David Ferrucci (1:42:28.540)
I think where they're gonna struggle
Lex Fridman (1:42:30.780)
is what we talked about before,
David Ferrucci (1:42:32.260)
which is relating to communicating with
Lex Fridman (1:42:36.540)
and understanding humans in deeper ways.
Lex Fridman (1:42:40.580)
And so I think that's a key point,
Lex Fridman (1:42:42.420)
like we can create the super parrot.
Lex Fridman (1:42:44.820)
What I mean by the super parrot is given enough data,
Lex Fridman (1:42:47.660)
a machine can mimic your emotional response,
David Ferrucci (1:42:50.140)
can even generate language that will sound smart
Lex Fridman (1:42:52.780)
and what someone else might say under similar circumstances.
David Ferrucci (1:42:57.860)
Like I would just pause on that,
Lex Fridman (1:42:58.940)
like that's the super parrot, right?
Lex Fridman (1:43:01.180)
So given similar circumstances,
Lex Fridman (1:43:03.660)
moves its faces in similar ways,
David Ferrucci (1:43:06.940)
changes its tone of voice in similar ways,
Lex Fridman (1:43:09.420)
produces strings of language that would similar
David Ferrucci (1:43:12.460)
that a human might say,
Lex Fridman (1:43:14.300)
not necessarily being able to produce
David Ferrucci (1:43:16.740)
a logical interpretation or understanding
Lex Fridman (1:43:20.620)
that would ultimately satisfy a critical interrogation
David Ferrucci (1:43:25.260)
or a critical understanding.
Lex Fridman (1:43:27.700)
I think you just described me in a nutshell.
Lex Fridman (1:43:30.540)
So I think philosophically speaking,
Lex Fridman (1:43:34.380)
you could argue that that's all we're doing
David Ferrucci (1:43:36.580)
as human beings to work super parrots.
Lex Fridman (1:43:37.860)
So I was gonna say, it's very possible,
David Ferrucci (1:43:40.300)
you know, humans do behave that way too.
Lex Fridman (1:43:42.620)
And so upon deeper probing and deeper interrogation,
David Ferrucci (1:43:45.860)
you may find out that there isn't a shared understanding
Lex Fridman (1:43:48.940)
because I think humans do both.
David Ferrucci (1:43:50.340)
Like humans are statistical language model machines
Lex Fridman (1:43:54.580)
and they are capable reasoners.
David Ferrucci (1:43:57.660)
You know, they're both.
Lex Fridman (1:43:59.900)
And you don't know which is going on, right?
David Ferrucci (1:44:02.900)
So, and I think it's an interesting problem.
Lex Fridman (1:44:09.140)
We talked earlier about like where we are
David Ferrucci (1:44:11.380)
in our social and political landscape.
Lex Fridman (1:44:14.700)
Can you distinguish someone who can string words together
Lex Fridman (1:44:19.540)
and sound like they know what they're talking about
Lex Fridman (1:44:21.820)
from someone who actually does?
Lex Fridman (1:44:24.020)
Can you do that without dialogue,
Lex Fridman (1:44:25.620)
without interrogative or probing dialogue?
Lex Fridman (1:44:27.780)
So it's interesting because humans are really good
Lex Fridman (1:44:31.100)
in their own mind, justifying or explaining what they hear
David Ferrucci (1:44:34.660)
because they project their understanding onto yours.
Lex Fridman (1:44:38.860)
So you could say, you could put together a string of words
Lex Fridman (1:44:41.540)
and someone will sit there and interpret it
Lex Fridman (1:44:44.020)
in a way that's extremely biased
David Ferrucci (1:44:46.060)
to the way they wanna interpret it.
Lex Fridman (1:44:47.140)
They wanna assume that you're an idiot
Lex Fridman (1:44:48.460)
and they'll interpret it one way.
Lex Fridman (1:44:50.060)
They will assume you're a genius
Lex Fridman (1:44:51.380)
and they'll interpret it another way that suits their needs.
Lex Fridman (1:44:54.380)
So this is tricky business.
Lex Fridman (1:44:56.460)
So I think to answer your question,
Lex Fridman (1:44:59.060)
as AI gets better and better, better and better mimic,
David Ferrucci (1:45:02.220)
you recreate the super parrots,
Lex Fridman (1:45:03.900)
we're challenged just as we are with,
David Ferrucci (1:45:06.580)
we're challenged with humans.
Lex Fridman (1:45:08.220)
Do you really know what you're talking about?
Lex Fridman (1:45:10.700)
Do you have a meaningful interpretation,
Lex Fridman (1:45:14.500)
a powerful framework that you could reason over
Lex Fridman (1:45:17.940)
and justify your answers, justify your predictions
Lex Fridman (1:45:23.420)
and your beliefs, why you think they make sense.
Lex Fridman (1:45:25.620)
Can you convince me what the implications are?
Lex Fridman (1:45:28.620)
So can you reason intelligently and make me believe
Lex Fridman (1:45:34.260)
that the implications of your prediction and so forth?
Lex Fridman (1:45:40.260)
So what happens is it becomes reflective.
David Ferrucci (1:45:44.060)
My standard for judging your intelligence
Lex Fridman (1:45:46.420)
depends a lot on mine.
Lex Fridman (1:45:49.940)
But you're saying there should be a large group of people
Lex Fridman (1:45:54.380)
with a certain standard of intelligence
David Ferrucci (1:45:56.900)
that would be convinced by this particular AI system.
Lex Fridman (1:46:02.580)
Then they'll pass.
David Ferrucci (1:46:03.540)
There should be, but I think depending on the content,
Lex Fridman (1:46:07.660)
one of the problems we have there
David Ferrucci (1:46:09.500)
is that if that large community of people
Lex Fridman (1:46:12.580)
are not judging it with regard to a rigorous standard
David Ferrucci (1:46:16.620)
of objective logic and reason, you still have a problem.
Lex Fridman (1:46:19.500)
Like masses of people can be persuaded.
David Ferrucci (1:46:23.780)
The millennials, yeah.
Lex Fridman (1:46:25.020)
To turn their brains off.
David Ferrucci (1:46:29.020)
Right, okay.
Lex Fridman (1:46:31.980)
Sorry.
David Ferrucci (1:46:32.820)
By the way, I have nothing against the millennials.
Lex Fridman (1:46:33.780)
No, I don't, I'm just, just.
Lex Fridman (1:46:36.060)
So you're a part of one of the great benchmarks,
Lex Fridman (1:46:40.980)
challenges of AI history.
Lex Fridman (1:46:43.280)
What do you think about AlphaZero, OpenAI5,
Lex Fridman (1:46:47.220)
AlphaStar accomplishments on video games recently,
David Ferrucci (1:46:50.740)
which are also, I think, at least in the case of Go,
Lex Fridman (1:46:55.300)
with AlphaGo and AlphaZero playing Go,
David Ferrucci (1:46:57.180)
was a monumental accomplishment as well.
Lex Fridman (1:46:59.700)
What are your thoughts about that challenge?
David Ferrucci (1:47:01.740)
I think it was a giant landmark for AI.
Lex Fridman (1:47:03.460)
I think it was phenomenal.
David Ferrucci (1:47:04.460)
I mean, it was one of those other things
Lex Fridman (1:47:06.020)
nobody thought like solving Go was gonna be easy,
David Ferrucci (1:47:08.540)
particularly because it's hard for,
Lex Fridman (1:47:10.460)
particularly hard for humans.
David Ferrucci (1:47:12.700)
Hard for humans to learn, hard for humans to excel at.
Lex Fridman (1:47:15.540)
And so it was another measure, a measure of intelligence.
David Ferrucci (1:47:21.380)
It's very cool.
Lex Fridman (1:47:22.500)
I mean, it's very interesting what they did.
Lex Fridman (1:47:25.260)
And I loved how they solved the data problem,
Lex Fridman (1:47:27.940)
which again, they bootstrapped it
Lex Fridman (1:47:29.180)
and got the machine to play itself,
Lex Fridman (1:47:30.420)
to generate enough data to learn from.
David Ferrucci (1:47:32.720)
I think that was brilliant.
Lex Fridman (1:47:33.860)
I think that was great.
Lex Fridman (1:47:35.660)
And of course, the result speaks for itself.
Lex Fridman (1:47:38.900)
I think it makes us think about,
Lex Fridman (1:47:40.900)
again, it is, okay, what's intelligence?
Lex Fridman (1:47:42.940)
What aspects of intelligence are important?
Lex Fridman (1:47:45.520)
Can the Go machine help me make me a better Go player?
Lex Fridman (1:47:49.340)
Is it an alien intelligence?
David Ferrucci (1:47:51.660)
Am I even capable of,
Lex Fridman (1:47:53.860)
like again, if we put in very simple terms,
David Ferrucci (1:47:56.060)
it found the function, it found the Go function.
Lex Fridman (1:47:59.180)
Can I even comprehend the Go function?
Lex Fridman (1:48:00.820)
Can I talk about the Go function?
Lex Fridman (1:48:02.260)
Can I conceptualize the Go function,
Lex Fridman (1:48:03.880)
like whatever it might be?
Lex Fridman (1:48:05.500)
So one of the interesting ideas of that system
Lex Fridman (1:48:08.040)
is that it plays against itself, right?
Lex Fridman (1:48:10.060)
But there's no human in the loop there.
Lex Fridman (1:48:12.660)
So like you're saying, it could have by itself
Lex Fridman (1:48:16.460)
created an alien intelligence.
Lex Fridman (1:48:18.900)
How?
Lex Fridman (1:48:19.740)
Toward a Go, imagine you're sentencing,
David Ferrucci (1:48:21.820)
you're a judge and you're sentencing people,
Lex Fridman (1:48:24.700)
or you're setting policy,
David Ferrucci (1:48:26.420)
or you're making medical decisions,
Lex Fridman (1:48:31.160)
and you can't explain,
David Ferrucci (1:48:33.340)
you can't get anybody to understand
Lex Fridman (1:48:34.880)
what you're doing or why.
Lex Fridman (1:48:37.300)
So it's an interesting dilemma
Lex Fridman (1:48:40.700)
for the applications of AI.
David Ferrucci (1:48:43.620)
Do we hold AI to this accountability
Lex Fridman (1:48:47.420)
that says humans have to be willing
Lex Fridman (1:48:51.460)
to take responsibility for the decision?
Lex Fridman (1:48:56.380)
In other words, can you explain why you would do the thing?
David Ferrucci (1:48:58.780)
Will you get up and speak to other humans
Lex Fridman (1:49:02.040)
and convince them that this was a smart decision?
Lex Fridman (1:49:04.660)
Is the AI enabling you to do that?
Lex Fridman (1:49:07.180)
Can you get behind the logic that was made there?
Lex Fridman (1:49:10.220)
Do you think, sorry to land on this point,
Lex Fridman (1:49:13.420)
because it's a fascinating one.
David Ferrucci (1:49:15.420)
It's a great goal for AI.
Lex Fridman (1:49:17.540)
Do you think it's achievable in many cases?
David Ferrucci (1:49:21.460)
Or, okay, there's two possible worlds
Lex Fridman (1:49:23.860)
that we have in the future.
David Ferrucci (1:49:25.820)
One is where AI systems do like medical diagnosis
Lex Fridman (1:49:28.940)
or things like that, or drive a car
David Ferrucci (1:49:32.420)
without ever explaining to you why it fails when it does.
Lex Fridman (1:49:36.580)
That's one possible world and we're okay with it.
David Ferrucci (1:49:40.340)
Or the other where we are not okay with it
Lex Fridman (1:49:42.980)
and we really hold back the technology
David Ferrucci (1:49:45.380)
from getting too good before it's able to explain.
Lex Fridman (1:49:48.780)
Which of those worlds are more likely, do you think,
Lex Fridman (1:49:50.800)
and which are concerning to you or not?
Lex Fridman (1:49:53.500)
I think the reality is it's gonna be a mix.
David Ferrucci (1:49:56.140)
I'm not sure I have a problem with that.
Lex Fridman (1:49:57.460)
I mean, I think there are tasks that are perfectly fine
David Ferrucci (1:49:59.940)
with machines show a certain level of performance
Lex Fridman (1:50:03.980)
and that level of performance is already better than humans.
Lex Fridman (1:50:07.740)
So for example, I don't know that I take driverless cars.
Lex Fridman (1:50:11.260)
If driverless cars learn how to be more effective drivers
David Ferrucci (1:50:14.300)
than humans but can't explain what they're doing,
Lex Fridman (1:50:16.900)
but bottom line, statistically speaking,
David Ferrucci (1:50:19.020)
they're 10 times safer than humans,
Lex Fridman (1:50:22.380)
I don't know that I care.
David Ferrucci (1:50:24.960)
I think when we have these edge cases
Lex Fridman (1:50:27.540)
when something bad happens and we wanna decide
David Ferrucci (1:50:29.700)
who's liable for that thing and who made that mistake
Lex Fridman (1:50:32.540)
and what do we do about that?
Lex Fridman (1:50:33.500)
And I think those edge cases are interesting cases.
Lex Fridman (1:50:36.740)
And now do we go to designers of the AI
Lex Fridman (1:50:38.940)
and the AI says, I don't know if that's what it learned
Lex Fridman (1:50:40.700)
to do and it says, well, you didn't train it properly.
David Ferrucci (1:50:43.620)
You were negligent in the training data
Lex Fridman (1:50:46.740)
that you gave that machine.
Lex Fridman (1:50:47.800)
Like, how do we drive down the reliability?
Lex Fridman (1:50:49.380)
So I think those are interesting questions.
Lex Fridman (1:50:53.180)
So the optimization problem there, sorry,
Lex Fridman (1:50:55.300)
is to create an AI system that's able
David Ferrucci (1:50:56.900)
to explain the lawyers away.
Lex Fridman (1:51:00.100)
There you go.
David Ferrucci (1:51:01.620)
I think it's gonna be interesting.
Lex Fridman (1:51:04.040)
I mean, I think this is where technology
Lex Fridman (1:51:05.820)
and social discourse are gonna get like deeply intertwined
Lex Fridman (1:51:09.460)
and how we start thinking about problems, decisions
Lex Fridman (1:51:12.380)
and problems like that.
Lex Fridman (1:51:13.500)
I think in other cases it becomes more obvious
David Ferrucci (1:51:15.860)
where it's like, why did you decide
Lex Fridman (1:51:20.260)
to give that person a longer sentence or deny them parole?
Lex Fridman (1:51:27.180)
Again, policy decisions or why did you pick that treatment?
Lex Fridman (1:51:30.580)
Like that treatment ended up killing that guy.
Lex Fridman (1:51:32.300)
Like, why was that a reasonable choice to make?
Lex Fridman (1:51:36.940)
And people are gonna demand explanations.
David Ferrucci (1:51:40.100)
Now there's a reality though here.
Lex Fridman (1:51:43.460)
And the reality is that it's not,
David Ferrucci (1:51:45.960)
I'm not sure humans are making reasonable choices
Lex Fridman (1:51:48.620)
when they do these things.
David Ferrucci (1:51:49.940)
They are using statistical hunches, biases,
Lex Fridman (1:51:54.780)
or even systematically using statistical averages
David Ferrucci (1:51:58.520)
to make calls.
Lex Fridman (1:51:59.360)
This is what happened to my dad
Lex Fridman (1:52:00.700)
and if you saw the talk I gave about that.
Lex Fridman (1:52:01.940)
But they decided that my father was brain dead.
David Ferrucci (1:52:07.300)
He had went into cardiac arrest
Lex Fridman (1:52:09.340)
and it took a long time for the ambulance to get there
Lex Fridman (1:52:12.420)
and he was not resuscitated right away and so forth.
Lex Fridman (1:52:14.580)
And they came and they told me he was brain dead
Lex Fridman (1:52:16.900)
and why was he brain dead?
Lex Fridman (1:52:17.860)
Because essentially they gave me
David Ferrucci (1:52:19.060)
a purely statistical argument under these conditions
Lex Fridman (1:52:22.020)
with these four features, 98% chance he's brain dead.
David Ferrucci (1:52:25.340)
I said, but can you just tell me not inductively,
Lex Fridman (1:52:28.940)
but deductively go there and tell me
David Ferrucci (1:52:30.460)
his brain's not functioning is the way for you to do that.
Lex Fridman (1:52:32.820)
And the protocol in response was,
David Ferrucci (1:52:35.960)
no, this is how we make this decision.
Lex Fridman (1:52:37.980)
I said, this is inadequate for me.
David Ferrucci (1:52:39.720)
I understand the statistics and I don't know how,
Lex Fridman (1:52:43.060)
there's a 2% chance he's still alive.
David Ferrucci (1:52:44.740)
I just don't know the specifics.
Lex Fridman (1:52:46.500)
I need the specifics of this case
Lex Fridman (1:52:49.380)
and I want the deductive logical argument
Lex Fridman (1:52:51.420)
about why you actually know he's brain dead.
Lex Fridman (1:52:53.580)
So I wouldn't sign the do not resuscitate.
Lex Fridman (1:52:55.980)
And I don't know, it was like they went through
David Ferrucci (1:52:57.820)
lots of procedures, it was a big long story,
Lex Fridman (1:53:00.020)
but the bottom was a fascinating story by the way,
Lex Fridman (1:53:02.060)
but how I reasoned and how the doctors reasoned
Lex Fridman (1:53:04.340)
through this whole process.
Lex Fridman (1:53:05.980)
But I don't know, somewhere around 24 hours later
Lex Fridman (1:53:07.900)
or something, he was sitting up in bed
David Ferrucci (1:53:09.460)
with zero brain damage.
Lex Fridman (1:53:13.940)
I mean, what lessons do you draw from that story,
Lex Fridman (1:53:18.020)
that experience?
Lex Fridman (1:53:19.500)
That the data that's being used
David Ferrucci (1:53:22.700)
to make statistical inferences
Lex Fridman (1:53:24.100)
doesn't adequately reflect the phenomenon.
Lex Fridman (1:53:26.440)
So in other words, you're getting shit wrong,
Lex Fridman (1:53:28.660)
I'm sorry, but you're getting stuff wrong
David Ferrucci (1:53:31.720)
because your model is not robust enough
Lex Fridman (1:53:35.260)
and you might be better off not using statistical inference
Lex Fridman (1:53:41.320)
and statistical averages in certain cases
Lex Fridman (1:53:43.060)
when you know the model's insufficient
Lex Fridman (1:53:45.220)
and that you should be reasoning about the specific case
Lex Fridman (1:53:48.440)
more logically and more deductibly
Lex Fridman (1:53:51.060)
and hold yourself responsible
Lex Fridman (1:53:52.420)
and hold yourself accountable to doing that.
Lex Fridman (1:53:55.360)
And perhaps AI has a role to say the exact thing
Lex Fridman (1:53:59.380)
what you just said, which is perhaps this is a case
David Ferrucci (1:54:02.980)
you should think for yourself,
Lex Fridman (1:54:05.420)
you should reason deductively.
David Ferrucci (1:54:08.020)
Well, so it's hard because it's hard to know that.
Lex Fridman (1:54:14.780)
You'd have to go back and you'd have to have enough data
David Ferrucci (1:54:17.220)
to essentially say, and this goes back to how do we,
Lex Fridman (1:54:20.180)
this goes back to the case of how do we decide
David Ferrucci (1:54:22.000)
whether the AI is good enough to do a particular task
Lex Fridman (1:54:25.380)
and regardless of whether or not
David Ferrucci (1:54:27.280)
it produces an explanation.
Lex Fridman (1:54:30.980)
And what standard do we hold for that?
Lex Fridman (1:54:34.940)
So if you look more broadly, for example,
Lex Fridman (1:54:42.860)
as my father, as a medical case,
David Ferrucci (1:54:48.180)
the medical system ultimately helped him a lot
Lex Fridman (1:54:50.140)
throughout his life, without it,
David Ferrucci (1:54:52.500)
he probably would have died much sooner.
Lex Fridman (1:54:55.640)
So overall, it sort of worked for him
David Ferrucci (1:54:58.900)
in sort of a net, net kind of way.
Lex Fridman (1:55:02.280)
Actually, I don't know that that's fair.
Lex Fridman (1:55:04.820)
But maybe not in that particular case, but overall,
Lex Fridman (1:55:07.660)
like the medical system overall does more good than bad.
David Ferrucci (1:55:10.580)
Yeah, the medical system overall
Lex Fridman (1:55:12.420)
was doing more good than bad.
David Ferrucci (1:55:14.300)
Now, there's another argument that suggests
Lex Fridman (1:55:16.560)
that wasn't the case, but for the sake of argument,
David Ferrucci (1:55:18.620)
let's say like that's, let's say a net positive.
Lex Fridman (1:55:21.060)
And I think you have to sit there and there
Lex Fridman (1:55:22.380)
and take that into consideration.
Lex Fridman (1:55:24.820)
Now you look at a particular use case,
David Ferrucci (1:55:26.660)
like for example, making this decision,
Lex Fridman (1:55:29.960)
have you done enough studies to know
Lex Fridman (1:55:33.400)
how good that prediction really is?
Lex Fridman (1:55:37.140)
And have you done enough studies to compare it,
David Ferrucci (1:55:40.060)
to say, well, what if we dug in in a more direct,
Lex Fridman (1:55:45.420)
let's get the evidence, let's do the deductive thing
Lex Fridman (1:55:47.980)
and not use statistics here,
Lex Fridman (1:55:49.420)
how often would that have done better?
Lex Fridman (1:55:52.460)
So you have to do the studies
Lex Fridman (1:55:53.700)
to know how good the AI actually is.
Lex Fridman (1:55:56.180)
And it's complicated because it depends how fast
Lex Fridman (1:55:58.500)
you have to make the decision.
Lex Fridman (1:55:59.540)
So if you have to make the decision super fast,
Lex Fridman (1:56:02.320)
you have no choice.
Lex Fridman (1:56:04.620)
If you have more time, right?
Lex Fridman (1:56:06.740)
But if you're ready to pull the plug,
Lex Fridman (1:56:09.040)
and this is a lot of the argument that I had with a doctor,
Lex Fridman (1:56:11.480)
I said, what's he gonna do if you do it,
Lex Fridman (1:56:13.220)
what's gonna happen to him in that room if you do it my way?
Lex Fridman (1:56:16.340)
You know, well, he's gonna die anyway.
Lex Fridman (1:56:18.740)
So let's do it my way then.
Lex Fridman (1:56:20.940)
I mean, it raises questions for our society
David Ferrucci (1:56:22.860)
to struggle with, as the case with your father,
Lex Fridman (1:56:26.500)
but also when things like race and gender
David Ferrucci (1:56:28.660)
start coming into play when certain,
Lex Fridman (1:56:31.740)
when judgments are made based on things
David Ferrucci (1:56:35.620)
that are complicated in our society,
Lex Fridman (1:56:39.060)
at least in the discourse.
Lex Fridman (1:56:40.120)
And it starts, you know, I think I'm safe to say
Lex Fridman (1:56:43.980)
that most of the violent crimes committed
David Ferrucci (1:56:46.300)
by males, so if you discriminate based,
Lex Fridman (1:56:51.300)
you know, it's a male versus female saying that
David Ferrucci (1:56:53.900)
if it's a male, more likely to commit the crime.
Lex Fridman (1:56:56.380)
This is one of my very positive and optimistic views
David Ferrucci (1:57:01.100)
of why the study of artificial intelligence,
Lex Fridman (1:57:05.540)
the process of thinking and reasoning logically
Lex Fridman (1:57:08.540)
and statistically, and how to combine them
Lex Fridman (1:57:10.500)
is so important for the discourse today,
David Ferrucci (1:57:12.180)
because it's causing a, regardless of what state AI devices
Lex Fridman (1:57:17.620)
are or not, it's causing this dialogue to happen.
David Ferrucci (1:57:22.220)
This is one of the most important dialogues
Lex Fridman (1:57:24.820)
that in my view, the human species can have right now,
David Ferrucci (1:57:28.180)
which is how to think well, how to reason well,
Lex Fridman (1:57:33.820)
how to understand our own cognitive biases
Lex Fridman (1:57:39.220)
and what to do about them.
Lex Fridman (1:57:40.980)
That has got to be one of the most important things
David Ferrucci (1:57:43.620)
we as a species can be doing, honestly.
Lex Fridman (1:57:47.240)
We are, we've created an incredibly complex society.
David Ferrucci (1:57:51.180)
We've created amazing abilities to amplify noise faster
Lex Fridman (1:57:56.300)
than we can amplify signal.
David Ferrucci (1:57:59.320)
We are challenged.
Lex Fridman (1:58:01.220)
We are deeply, deeply challenged.
David Ferrucci (1:58:03.620)
We have, you know, big segments of the population
Lex Fridman (1:58:06.260)
getting hit with enormous amounts of information.
Lex Fridman (1:58:08.940)
Do they know how to do critical thinking?
Lex Fridman (1:58:10.940)
Do they know how to objectively reason?
David Ferrucci (1:58:14.200)
Do they understand what they are doing,
Lex Fridman (1:58:16.980)
nevermind what their AI is doing?
David Ferrucci (1:58:19.700)
This is such an important dialogue to be having.
Lex Fridman (1:58:23.180)
And, you know, we are fundamentally,
David Ferrucci (1:58:26.860)
our thinking can be and easily becomes fundamentally bias.
Lex Fridman (1:58:31.420)
And there are statistics and we shouldn't blind our,
David Ferrucci (1:58:34.460)
we shouldn't discard statistical inference,
Lex Fridman (1:58:37.300)
but we should understand the nature
David Ferrucci (1:58:39.020)
of statistical inference.
Lex Fridman (1:58:40.900)
As a society, as you know,
David Ferrucci (1:58:44.020)
we decide to reject statistical inference
Lex Fridman (1:58:48.220)
to favor understanding and deciding on the individual.
David Ferrucci (1:58:55.600)
Yes.
Lex Fridman (1:58:57.180)
We consciously make that choice.
Lex Fridman (1:59:00.660)
So even if the statistics said,
Lex Fridman (1:59:04.140)
even if the statistics said males are more likely to have,
David Ferrucci (1:59:08.300)
you know, to be violent criminals,
Lex Fridman (1:59:09.660)
we still take each person as an individual
Lex Fridman (1:59:12.580)
and we treat them based on the logic
Lex Fridman (1:59:16.820)
and the knowledge of that situation.
David Ferrucci (1:59:20.260)
We purposefully and intentionally
Lex Fridman (1:59:24.100)
reject the statistical inference.
David Ferrucci (1:59:28.320)
We do that out of respect for the individual.
Lex Fridman (1:59:31.260)
For the individual.
David Ferrucci (1:59:32.100)
Yeah, and that requires reasoning and thinking.
Lex Fridman (1:59:34.060)
Correct.
David Ferrucci (1:59:35.180)
Looking forward, what grand challenges
Lex Fridman (1:59:37.420)
would you like to see in the future?
David Ferrucci (1:59:38.940)
Because the Jeopardy challenge, you know,
Lex Fridman (1:59:43.380)
captivated the world.
David Ferrucci (1:59:45.140)
AlphaGo, AlphaZero captivated the world.
Lex Fridman (1:59:48.060)
Deep Blue certainly beating Kasparov.
David Ferrucci (1:59:51.580)
Gary's bitterness aside captivated the world.
Lex Fridman (1:59:55.700)
What do you think, do you have ideas
Lex Fridman (1:59:57.880)
for next grand challenges for future challenges of that?
Lex Fridman (20:03.720)
And so what I mean by that is
David Ferrucci (20:05.480)
if you look at the entire enterprise of science,
Lex Fridman (20:09.500)
science is supposed to be at about
Lex Fridman (20:11.640)
objective reason and reason, right?
Lex Fridman (20:13.680)
So we think about, gee, who's the most intelligent person
Lex Fridman (20:17.680)
or group of people in the world?
Lex Fridman (20:20.500)
Do we think about the savants who can close their eyes
Lex Fridman (20:24.080)
and give you a number?
Lex Fridman (20:25.540)
We think about the think tanks,
David Ferrucci (20:27.680)
or the scientists or the philosophers
Lex Fridman (20:29.500)
who kind of work through the details
Lex Fridman (20:32.680)
and write the papers and come up with the thoughtful,
Lex Fridman (20:35.960)
logical proofs and use the scientific method.
David Ferrucci (20:39.480)
I think it's the latter.
Lex Fridman (20:42.760)
And my point is that how do you train someone to do that?
Lex Fridman (20:45.800)
And that's what I mean by it's hard.
Lex Fridman (20:46.920)
How do you, what's the process of training people
Lex Fridman (20:49.400)
to do that well?
Lex Fridman (20:50.800)
That's a hard process.
David Ferrucci (20:52.400)
We work, as a society, we work pretty hard
Lex Fridman (20:56.020)
to get other people to understand our thinking
Lex Fridman (20:59.240)
and to convince them of things.
Lex Fridman (21:02.220)
Now we could persuade them,
David Ferrucci (21:04.040)
obviously you talked about this,
Lex Fridman (21:05.300)
like human flaws or weaknesses,
David Ferrucci (21:07.520)
we can persuade them through emotional means.
Lex Fridman (21:12.160)
But to get them to understand and connect to
Lex Fridman (21:16.140)
and follow a logical argument is difficult.
Lex Fridman (21:19.960)
We try it, we do it, we do it as scientists,
David Ferrucci (21:22.440)
we try to do it as journalists,
Lex Fridman (21:24.200)
we try to do it as even artists in many forms,
David Ferrucci (21:27.280)
as writers, as teachers.
Lex Fridman (21:29.780)
We go through a fairly significant training process
David Ferrucci (21:33.860)
to do that.
Lex Fridman (21:34.700)
And then we could ask, well, why is that so hard?
Lex Fridman (21:39.040)
But it's hard.
Lex Fridman (21:39.920)
And for humans, it takes a lot of work.
Lex Fridman (21:44.060)
And when we step back and say,
Lex Fridman (21:45.960)
well, how do we get a machine to do that?
David Ferrucci (21:49.160)
It's a vexing question.
Lex Fridman (21:51.840)
How would you begin to try to solve that?
Lex Fridman (21:55.240)
And maybe just a quick pause,
Lex Fridman (21:57.400)
because there's an optimistic notion
David Ferrucci (21:59.840)
in the things you're describing,
Lex Fridman (22:01.040)
which is being able to explain something through reason.
Lex Fridman (22:05.980)
But if you look at algorithms that recommend things
Lex Fridman (22:08.660)
that we'll look at next, whether it's Facebook, Google,
David Ferrucci (22:11.800)
advertisement based companies, their goal is to convince you
Lex Fridman (22:18.000)
to buy things based on anything.
Lex Fridman (22:23.480)
So that could be reason,
Lex Fridman (22:25.440)
because the best of advertisement is showing you things
David Ferrucci (22:28.100)
that you really do need and explain why you need it.
Lex Fridman (22:31.980)
But it could also be through emotional manipulation.
David Ferrucci (22:37.000)
The algorithm that describes why a certain decision
Lex Fridman (22:41.760)
was made, how hard is it to do it
Lex Fridman (22:45.760)
through emotional manipulation?
Lex Fridman (22:48.160)
And why is that a good or a bad thing?
Lex Fridman (22:52.760)
So you've kind of focused on reason, logic,
Lex Fridman (22:56.840)
really showing in a clear way why something is good.
Lex Fridman (23:02.520)
One, is that even a thing that us humans do?
Lex Fridman (23:05.920)
And two, how do you think of the difference
Lex Fridman (23:09.920)
in the reasoning aspect and the emotional manipulation?
Lex Fridman (23:15.160)
So you call it emotional manipulation,
Lex Fridman (23:17.360)
but more objectively is essentially saying,
Lex Fridman (23:20.220)
there are certain features of things
David Ferrucci (23:22.660)
that seem to attract your attention.
Lex Fridman (23:24.440)
I mean, it kind of give you more of that stuff.
David Ferrucci (23:26.800)
Manipulation is a bad word.
Lex Fridman (23:28.280)
Yeah, I mean, I'm not saying it's good right or wrong.
David Ferrucci (23:31.140)
It works to get your attention
Lex Fridman (23:32.960)
and it works to get you to buy stuff.
Lex Fridman (23:34.440)
And when you think about algorithms that look
Lex Fridman (23:36.960)
at the patterns of features
David Ferrucci (23:40.000)
that you seem to be spending your money on
Lex Fridman (23:41.920)
and say, I'm gonna give you something
David Ferrucci (23:43.280)
with a similar pattern.
Lex Fridman (23:44.820)
So I'm gonna learn that function
David Ferrucci (23:46.080)
because the objective is to get you to click on it
Lex Fridman (23:48.200)
or get you to buy it or whatever it is.
David Ferrucci (23:51.120)
I don't know, I mean, it is what it is.
Lex Fridman (23:53.400)
I mean, that's what the algorithm does.
David Ferrucci (23:55.840)
You can argue whether it's good or bad.
Lex Fridman (23:57.440)
It depends what your goal is.
David Ferrucci (24:00.440)
I guess this seems to be very useful
Lex Fridman (24:02.480)
for convincing, for telling a story.
David Ferrucci (24:05.280)
For convincing humans, it's good
Lex Fridman (24:07.680)
because again, this goes back to what is the human behavior
Lex Fridman (24:11.800)
like, how does the human brain respond to things?
Lex Fridman (24:17.000)
I think there's a more optimistic view of that too,
David Ferrucci (24:19.360)
which is that if you're searching
Lex Fridman (24:22.000)
for certain kinds of things,
David Ferrucci (24:23.120)
you've already reasoned that you need them.
Lex Fridman (24:26.160)
And these algorithms are saying, look, that's up to you
David Ferrucci (24:30.080)
to reason whether you need something or not.
Lex Fridman (24:32.180)
That's your job.
David Ferrucci (24:33.400)
You may have an unhealthy addiction to this stuff
Lex Fridman (24:36.920)
or you may have a reasoned and thoughtful explanation
David Ferrucci (24:42.880)
for why it's important to you.
Lex Fridman (24:44.520)
And the algorithms are saying, hey, that's like, whatever.
David Ferrucci (24:47.040)
Like, that's your problem.
Lex Fridman (24:48.040)
All I know is you're buying stuff like that.
David Ferrucci (24:50.580)
You're interested in stuff like that.
Lex Fridman (24:51.920)
Could be a bad reason, could be a good reason.
David Ferrucci (24:53.920)
That's up to you.
Lex Fridman (24:55.080)
I'm gonna show you more of that stuff.
Lex Fridman (24:57.520)
And I think that it's not good or bad.
Lex Fridman (25:01.640)
It's not reasoned or not reasoned.
David Ferrucci (25:03.520)
The algorithm is doing what it does,
Lex Fridman (25:04.880)
which is saying, you seem to be interested in this.
David Ferrucci (25:06.920)
I'm gonna show you more of that stuff.
Lex Fridman (25:09.320)
And I think we're seeing this not just in buying stuff,
Lex Fridman (25:11.200)
but even in social media.
Lex Fridman (25:12.160)
You're reading this kind of stuff.
David Ferrucci (25:13.960)
I'm not judging on whether it's good or bad.
Lex Fridman (25:15.740)
I'm not reasoning at all.
David Ferrucci (25:16.920)
I'm just saying, I'm gonna show you other stuff
Lex Fridman (25:19.200)
with similar features.
Lex Fridman (25:20.800)
And like, and that's it.
Lex Fridman (25:22.360)
And I wash my hands from it and I say,
David Ferrucci (25:23.840)
that's all that's going on.
Lex Fridman (25:25.940)
You know, there is, people are so harsh on AI systems.
Lex Fridman (25:31.900)
So one, the bar of performance is extremely high.
Lex Fridman (25:34.940)
And yet we also ask them to, in the case of social media,
David Ferrucci (25:39.560)
to help find the better angels of our nature
Lex Fridman (25:42.940)
and help make a better society.
Lex Fridman (25:45.980)
What do you think about the role of AI there?
Lex Fridman (25:47.860)
So that, I agree with you.
Lex Fridman (25:48.860)
That's the interesting dichotomy, right?
Lex Fridman (25:51.580)
Because on one hand, we're sitting there
Lex Fridman (25:54.160)
and we're sort of doing the easy part,
Lex Fridman (25:55.880)
which is finding the patterns.
David Ferrucci (25:57.980)
We're not building, the system's not building a theory
Lex Fridman (26:01.820)
that is consumable and understandable to other humans
David Ferrucci (26:04.220)
that can be explained and justified.
Lex Fridman (26:06.380)
And so on one hand to say, oh, you know, AI is doing this.
Lex Fridman (26:11.500)
Why isn't doing this other thing?
Lex Fridman (26:13.700)
Well, this other thing's a lot harder.
Lex Fridman (26:16.300)
And it's interesting to think about why it's harder.
Lex Fridman (26:20.180)
And because you're interpreting the data
David Ferrucci (26:23.980)
in the context of prior models.
Lex Fridman (26:26.300)
In other words, understandings
David Ferrucci (26:28.140)
of what's important in the world, what's not important.
Lex Fridman (26:30.260)
What are all the other abstract features
Lex Fridman (26:32.060)
that drive our decision making?
Lex Fridman (26:35.420)
What's sensible, what's not sensible,
David Ferrucci (26:36.980)
what's good, what's bad, what's moral,
Lex Fridman (26:38.620)
what's valuable, what isn't?
Lex Fridman (26:40.060)
Where is that stuff?
Lex Fridman (26:41.160)
No one's applying the interpretation.
Lex Fridman (26:43.300)
So when I see you clicking on a bunch of stuff
Lex Fridman (26:46.660)
and I look at these simple features, the raw features,
David Ferrucci (26:49.780)
the features that are there in the data,
Lex Fridman (26:51.140)
like what words are being used or how long the material is
David Ferrucci (26:57.700)
or other very superficial features,
Lex Fridman (27:00.620)
what colors are being used in the material.
David Ferrucci (27:02.500)
Like, I don't know why you're clicking
Lex Fridman (27:03.880)
on this stuff you're clicking.
David Ferrucci (27:04.980)
Or if it's products, what the price is
Lex Fridman (27:07.620)
or what the categories and stuff like that.
Lex Fridman (27:09.540)
And I just feed you more of the same stuff.
Lex Fridman (27:11.540)
That's very different than kind of getting in there
Lex Fridman (27:13.740)
and saying, what does this mean?
Lex Fridman (27:16.020)
The stuff you're reading, like why are you reading it?
Lex Fridman (27:21.380)
What assumptions are you bringing to the table?
Lex Fridman (27:23.900)
Are those assumptions sensible?
Lex Fridman (27:26.380)
Does the material make any sense?
Lex Fridman (27:28.980)
Does it lead you to thoughtful, good conclusions?
David Ferrucci (27:34.080)
Again, there's interpretation and judgment involved
Lex Fridman (27:37.420)
in that process that isn't really happening in the AI today.
David Ferrucci (27:42.420)
That's harder because you have to start getting
Lex Fridman (27:47.200)
at the meaning of the stuff, of the content.
David Ferrucci (27:52.040)
You have to get at how humans interpret the content
Lex Fridman (27:55.760)
relative to their value system
Lex Fridman (27:58.720)
and deeper thought processes.
Lex Fridman (28:00.600)
So that's what meaning means is not just some kind
David Ferrucci (28:04.520)
of deep, timeless, semantic thing
Lex Fridman (28:09.300)
that the statement represents,
Lex Fridman (28:10.960)
but also how a large number of people
Lex Fridman (28:13.400)
are likely to interpret.
Lex Fridman (28:15.220)
So that's again, even meaning is a social construct.
Lex Fridman (28:19.120)
So you have to try to predict how most people
David Ferrucci (28:22.800)
would understand this kind of statement.
Lex Fridman (28:24.520)
Yeah, meaning is often relative,
Lex Fridman (28:27.300)
but meaning implies that the connections go beneath
Lex Fridman (28:30.400)
the surface of the artifacts.
David Ferrucci (28:31.840)
If I show you a painting, it's a bunch of colors on a canvas,
Lex Fridman (28:35.480)
what does it mean to you?
Lex Fridman (28:37.140)
And it may mean different things to different people
Lex Fridman (28:39.400)
because of their different experiences.
David Ferrucci (28:42.240)
It may mean something even different
Lex Fridman (28:44.720)
to the artist who painted it.
David Ferrucci (28:47.440)
As we try to get more rigorous with our communication,
Lex Fridman (28:50.720)
we try to really nail down that meaning.
Lex Fridman (28:53.280)
So we go from abstract art to precise mathematics,
Lex Fridman (28:58.840)
precise engineering drawings and things like that.
David Ferrucci (29:01.520)
We're really trying to say, I wanna narrow
Lex Fridman (29:04.480)
that space of possible interpretations
David Ferrucci (29:08.300)
because the precision of the communication
Lex Fridman (29:10.720)
ends up becoming more and more important.
Lex Fridman (29:13.400)
And so that means that I have to specify,
Lex Fridman (29:17.920)
and I think that's why this becomes really hard,
David Ferrucci (29:21.400)
because if I'm just showing you an artifact
Lex Fridman (29:24.200)
and you're looking at it superficially,
David Ferrucci (29:26.000)
whether it's a bunch of words on a page,
Lex Fridman (29:28.220)
or whether it's brushstrokes on a canvas
David Ferrucci (29:31.940)
or pixels on a photograph,
Lex Fridman (29:33.600)
you can sit there and you can interpret
David Ferrucci (29:35.080)
lots of different ways at many, many different levels.
Lex Fridman (29:38.880)
But when I wanna align our understanding of that,
David Ferrucci (29:45.640)
I have to specify a lot more stuff
Lex Fridman (29:48.360)
that's actually not directly in the artifact.
David Ferrucci (29:52.320)
Now I have to say, well, how are you interpreting
Lex Fridman (29:55.960)
this image and that image?
Lex Fridman (29:57.240)
And what about the colors and what do they mean to you?
Lex Fridman (29:59.360)
What perspective are you bringing to the table?
David Ferrucci (2:00:00.900)
You know, look, I mean, I think there are lots
Lex Fridman (2:00:03.280)
of really great ideas for grand challenges.
David Ferrucci (2:00:05.800)
I'm particularly focused on one right now,
Lex Fridman (2:00:08.500)
which is, you know, can you demonstrate
David Ferrucci (2:00:11.660)
that they understand, that they could read and understand,
Lex Fridman (2:00:14.980)
that they can acquire these frameworks
Lex Fridman (2:00:18.020)
and communicate, you know,
Lex Fridman (2:00:19.420)
reason and communicate with humans.
Lex Fridman (2:00:21.160)
So it is kind of like the Turing test,
Lex Fridman (2:00:23.380)
but it's a little bit more demanding than the Turing test.
David Ferrucci (2:00:26.540)
It's not enough to convince me that you might be human
Lex Fridman (2:00:31.260)
because you could, you know, you can parrot a conversation.
David Ferrucci (2:00:34.920)
I think, you know, the standard is a little bit higher,
Lex Fridman (2:00:38.820)
is for example, can you, you know, the standard is higher.
Lex Fridman (2:00:43.380)
And I think one of the challenges
Lex Fridman (2:00:45.540)
of devising this grand challenge is that we're not sure
Lex Fridman (2:00:51.960)
what intelligence is, we're not sure how to determine
Lex Fridman (2:00:56.220)
whether or not two people actually understand each other
Lex Fridman (2:00:59.140)
and in what depth they understand it, you know,
Lex Fridman (2:01:02.260)
to what depth they understand each other.
Lex Fridman (2:01:04.340)
So the challenge becomes something along the lines of,
Lex Fridman (2:01:08.380)
can you satisfy me that we have a shared understanding?
Lex Fridman (2:01:14.800)
So if I were to probe and probe and you probe me,
Lex Fridman (2:01:18.380)
can machines really act like thought partners
David Ferrucci (2:01:23.420)
where they can satisfy me that we have a shared,
Lex Fridman (2:01:27.340)
our understanding is shared enough
David Ferrucci (2:01:29.420)
that we can collaborate and produce answers together
Lex Fridman (2:01:33.300)
and that, you know, they can help me explain
Lex Fridman (2:01:35.460)
and justify those answers.
Lex Fridman (2:01:36.740)
So maybe here's an idea.
Lex Fridman (2:01:38.100)
So we'll have AI system run for president and convince.
Lex Fridman (2:01:44.500)
That's too easy.
David Ferrucci (2:01:46.100)
I'm sorry, go ahead.
Lex Fridman (2:01:46.940)
Well, no, you have to convince the voters
David Ferrucci (2:01:49.380)
that they should vote.
Lex Fridman (2:01:51.580)
So like, I guess what does winning look like?
David Ferrucci (2:01:53.780)
Again, that's why I think this is such a challenge
Lex Fridman (2:01:55.860)
because we go back to the emotional persuasion.
David Ferrucci (2:01:59.980)
We go back to, you know, now we're checking off an aspect
Lex Fridman (2:02:06.040)
of human cognition that is in many ways weak or flawed,
David Ferrucci (2:02:11.460)
right, we're so easily manipulated.
Lex Fridman (2:02:13.940)
Our minds are drawn for often the wrong reasons, right?
David Ferrucci (2:02:19.620)
Not the reasons that ultimately matter to us,
Lex Fridman (2:02:21.840)
but the reasons that can easily persuade us.
David Ferrucci (2:02:23.980)
I think we can be persuaded to believe one thing or another
Lex Fridman (2:02:28.420)
for reasons that ultimately don't serve us well
David Ferrucci (2:02:31.340)
in the longterm.
Lex Fridman (2:02:33.160)
And a good benchmark should not play with those elements
David Ferrucci (2:02:38.460)
of emotional manipulation.
Lex Fridman (2:02:40.780)
I don't think so.
Lex Fridman (2:02:41.620)
And I think that's where we have to set the higher standard
Lex Fridman (2:02:44.340)
for ourselves of what, you know, what does it mean?
David Ferrucci (2:02:47.100)
This goes back to rationality
Lex Fridman (2:02:48.900)
and it goes back to objective thinking.
Lex Fridman (2:02:50.580)
And can you produce, can you acquire information
Lex Fridman (2:02:53.300)
and produce reasoned arguments
Lex Fridman (2:02:54.800)
and to those reasoned arguments
Lex Fridman (2:02:56.300)
pass a certain amount of muster and is it,
Lex Fridman (2:03:00.140)
and can you acquire new knowledge?
Lex Fridman (2:03:02.220)
You know, can you, for example, can you reason,
David Ferrucci (2:03:06.260)
I have acquired new knowledge,
Lex Fridman (2:03:07.460)
can you identify where it's consistent or contradictory
Lex Fridman (2:03:11.220)
with other things you've learned?
Lex Fridman (2:03:12.860)
And can you explain that to me
Lex Fridman (2:03:14.020)
and get me to understand that?
Lex Fridman (2:03:15.580)
So I think another way to think about it perhaps
David Ferrucci (2:03:18.540)
is can a machine teach you, can it help you understand
Lex Fridman (2:03:31.900)
something that you didn't really understand before
David Ferrucci (2:03:35.260)
where it's taking you, so you're not,
Lex Fridman (2:03:39.140)
again, it's almost like can it teach you,
David Ferrucci (2:03:41.360)
can it help you learn and in an arbitrary space
Lex Fridman (2:03:46.360)
so it can open those domain space?
Lex Fridman (2:03:49.000)
So can you tell the machine, and again,
Lex Fridman (2:03:50.440)
this borrows from some science fiction,
Lex Fridman (2:03:52.840)
but can you go off and learn about this topic
Lex Fridman (2:03:55.820)
that I'd like to understand better
Lex Fridman (2:03:58.300)
and then work with me to help me understand it?
Lex Fridman (2:04:02.000)
That's quite brilliant.
David Ferrucci (2:04:03.600)
What, the machine that passes that kind of test,
Lex Fridman (2:04:06.940)
do you think it would need to have self awareness
Lex Fridman (2:04:11.520)
or even consciousness?
Lex Fridman (2:04:13.120)
What do you think about consciousness
Lex Fridman (2:04:16.140)
and the importance of it maybe in relation to having a body,
Lex Fridman (2:04:21.080)
having a presence, an entity?
Lex Fridman (2:04:24.720)
Do you think that's important?
Lex Fridman (2:04:26.720)
You know, people used to ask me if Watson was conscious
Lex Fridman (2:04:28.740)
and I used to say, he's conscious of what exactly?
Lex Fridman (2:04:32.240)
I mean, I think, you know, maybe it depends
Lex Fridman (2:04:34.580)
what it is that you're conscious of.
Lex Fridman (2:04:36.020)
I mean, like, so, you know, did it, if you, you know,
David Ferrucci (2:04:39.380)
it's certainly easy for it to answer questions
Lex Fridman (2:04:42.400)
about, it would be trivial to program it
David Ferrucci (2:04:44.760)
to answer questions about whether or not
Lex Fridman (2:04:46.600)
it was playing Jeopardy.
David Ferrucci (2:04:47.540)
I mean, it could certainly answer questions
Lex Fridman (2:04:48.980)
that would imply that it was aware of things.
David Ferrucci (2:04:51.240)
Exactly, what does it mean to be aware
Lex Fridman (2:04:52.620)
and what does it mean to be conscious of?
David Ferrucci (2:04:53.640)
It's sort of interesting.
Lex Fridman (2:04:54.480)
I mean, I think that we differ from one another
David Ferrucci (2:04:57.960)
based on what we're conscious of.
Lex Fridman (2:05:01.080)
But wait, wait a minute, yes, for sure.
David Ferrucci (2:05:02.660)
There's degrees of consciousness in there, so.
Lex Fridman (2:05:05.320)
Well, and there's just areas.
Lex Fridman (2:05:06.920)
Like, it's not just degrees, what are you aware of?
Lex Fridman (2:05:10.120)
Like, what are you not aware of?
Lex Fridman (2:05:11.120)
But nevertheless, there's a very subjective element
Lex Fridman (2:05:13.440)
to our experience.
David Ferrucci (2:05:16.000)
Let me even not talk about consciousness.
Lex Fridman (2:05:18.340)
Let me talk about another, to me,
David Ferrucci (2:05:21.560)
really interesting topic of mortality, fear of mortality.
Lex Fridman (2:05:25.560)
Watson, as far as I could tell,
David Ferrucci (2:05:29.280)
did not have a fear of death.
Lex Fridman (2:05:32.160)
Certainly not.
David Ferrucci (2:05:33.000)
Most, most humans do.
Lex Fridman (2:05:36.960)
Wasn't conscious of death.
David Ferrucci (2:05:39.040)
It wasn't, yeah.
Lex Fridman (2:05:40.000)
So there's an element of finiteness to our existence
David Ferrucci (2:05:44.160)
that I think, like you mentioned, survival,
Lex Fridman (2:05:47.240)
that adds to the whole thing.
David Ferrucci (2:05:49.040)
I mean, consciousness is tied up with that,
Lex Fridman (2:05:50.860)
that we are a thing.
David Ferrucci (2:05:52.880)
It's a subjective thing that ends.
Lex Fridman (2:05:56.200)
And that seems to add a color and flavor
David Ferrucci (2:05:59.000)
to our motivations in a way
Lex Fridman (2:06:00.440)
that seems to be fundamentally important for intelligence,
David Ferrucci (2:06:05.960)
or at least the kind of human intelligence.
Lex Fridman (2:06:07.920)
Well, I think for generating goals, again,
David Ferrucci (2:06:10.200)
I think you could have,
Lex Fridman (2:06:12.280)
you could have an intelligence capability
Lex Fridman (2:06:14.560)
and a capability to learn, a capability to predict.
Lex Fridman (2:06:18.560)
But I think without,
David Ferrucci (2:06:22.480)
I mean, again, you get fear,
Lex Fridman (2:06:23.960)
but essentially without the goal to survive.
Lex Fridman (2:06:27.040)
So you think you can just encode that
Lex Fridman (2:06:29.120)
without having to really?
David Ferrucci (2:06:30.600)
I think you could encode.
Lex Fridman (2:06:31.440)
I mean, you could create a robot now,
Lex Fridman (2:06:32.880)
and you could say, you know, plug it in,
Lex Fridman (2:06:36.060)
and say, protect your power source, you know,
Lex Fridman (2:06:38.520)
and give it some capabilities,
Lex Fridman (2:06:39.720)
and it'll sit there and operate
David Ferrucci (2:06:40.900)
to try to protect its power source and survive.
Lex Fridman (2:06:42.760)
I mean, so I don't know
David Ferrucci (2:06:44.240)
that that's philosophically a hard thing to demonstrate.
Lex Fridman (2:06:46.680)
It sounds like a fairly easy thing to demonstrate
David Ferrucci (2:06:48.960)
that you can give it that goal.
Lex Fridman (2:06:50.040)
Will it come up with that goal by itself?
David Ferrucci (2:06:52.360)
I think you have to program that goal in.
Lex Fridman (2:06:54.520)
But there's something,
David Ferrucci (2:06:56.660)
because I think, as we touched on,
Lex Fridman (2:06:58.580)
intelligence is kind of like a social construct.
David Ferrucci (2:07:01.480)
The fact that a robot will be protecting its power source
Lex Fridman (2:07:07.080)
would add depth and grounding to its intelligence
David Ferrucci (2:07:12.960)
in terms of us being able to respect it.
Lex Fridman (2:07:15.800)
I mean, ultimately, it boils down to us acknowledging
David Ferrucci (2:07:18.880)
that it's intelligent.
Lex Fridman (2:07:20.660)
And the fact that it can die,
David Ferrucci (2:07:23.480)
I think, is an important part of that.
Lex Fridman (2:07:26.120)
The interesting thing to reflect on
David Ferrucci (2:07:27.820)
is how trivial that would be.
Lex Fridman (2:07:29.520)
And I don't think, if you knew how trivial that was,
David Ferrucci (2:07:32.080)
you would associate that with being intelligence.
Lex Fridman (2:07:35.360)
I mean, I literally put in a statement of code
David Ferrucci (2:07:37.440)
that says you have the following actions you can take.
Lex Fridman (2:07:40.400)
You give it a bunch of actions,
David Ferrucci (2:07:41.600)
like maybe you mount a laser gun on it,
Lex Fridman (2:07:44.000)
or you give it the ability to scream or screech or whatever.
Lex Fridman (2:07:48.920)
And you say, if you see your power source threatened,
Lex Fridman (2:07:52.680)
then you could program that in,
Lex Fridman (2:07:53.880)
and you're gonna take these actions to protect it.
Lex Fridman (2:07:58.040)
You know, you could train it on a bunch of things.
David Ferrucci (2:08:02.200)
So, and now you're gonna look at that and you say,
Lex Fridman (2:08:04.080)
well, you know, that's intelligence,
Lex Fridman (2:08:05.280)
which is protecting its power source?
Lex Fridman (2:08:06.840)
Maybe, but that's, again, this human bias that says,
David Ferrucci (2:08:10.220)
the thing I identify, my intelligence and my conscious,
Lex Fridman (2:08:14.600)
so fundamentally with the desire,
David Ferrucci (2:08:16.720)
or at least the behaviors associated
Lex Fridman (2:08:18.680)
with the desire to survive,
David Ferrucci (2:08:21.340)
that if I see another thing doing that,
Lex Fridman (2:08:24.720)
I'm going to assume it's intelligent.
Lex Fridman (2:08:27.280)
What timeline, year,
Lex Fridman (2:08:29.760)
will society have something that would,
David Ferrucci (2:08:34.640)
that you would be comfortable calling
Lex Fridman (2:08:36.000)
an artificial general intelligence system?
Lex Fridman (2:08:39.560)
Well, what's your intuition?
Lex Fridman (2:08:41.080)
Nobody can predict the future,
David Ferrucci (2:08:42.480)
certainly not the next few months or 20 years away,
Lex Fridman (2:08:46.480)
but what's your intuition?
Lex Fridman (2:08:47.600)
How far away are we?
Lex Fridman (2:08:50.080)
I don't know.
David Ferrucci (2:08:50.920)
It's hard to make these predictions.
Lex Fridman (2:08:52.120)
I mean, I would be guessing,
Lex Fridman (2:08:54.760)
and there's so many different variables,
Lex Fridman (2:08:57.000)
including just how much we want to invest in it
Lex Fridman (2:08:59.080)
and how important we think it is,
Lex Fridman (2:09:03.480)
what kind of investment we're willing to make in it,
Lex Fridman (2:09:06.160)
what kind of talent we end up bringing to the table,
Lex Fridman (2:09:08.440)
the incentive structure, all these things.
Lex Fridman (2:09:10.160)
So I think it is possible to do this sort of thing.
Lex Fridman (2:09:15.220)
I think it's, I think trying to sort of
David Ferrucci (2:09:20.360)
ignore many of the variables and things like that,
Lex Fridman (2:09:23.040)
is it a 10 year thing, is it a 23 year?
David Ferrucci (2:09:25.440)
Probably closer to a 20 year thing, I guess.
Lex Fridman (2:09:27.880)
But not several hundred years.
David Ferrucci (2:09:29.720)
No, I don't think it's several hundred years.
Lex Fridman (2:09:32.080)
I don't think it's several hundred years.
Lex Fridman (2:09:33.660)
But again, so much depends on how committed we are
Lex Fridman (2:09:38.860)
to investing and incentivizing this type of work.
Lex Fridman (2:09:43.120)
And it's sort of interesting.
Lex Fridman (2:09:45.200)
Like, I don't think it's obvious how incentivized we are.
David Ferrucci (2:09:50.280)
I think from a task perspective,
Lex Fridman (2:09:53.160)
if we see business opportunities to take this technique
David Ferrucci (2:09:57.880)
or that technique to solve that problem,
Lex Fridman (2:09:59.120)
I think that's the main driver for many of these things.
David Ferrucci (2:10:03.240)
From a general intelligence,
Lex Fridman (2:10:05.600)
it's kind of an interesting question.
Lex Fridman (2:10:06.920)
Are we really motivated to do that?
Lex Fridman (2:10:09.360)
And like, we just struggled ourselves right now
David Ferrucci (2:10:12.520)
to even define what it is.
Lex Fridman (2:10:14.760)
So it's hard to incentivize
David Ferrucci (2:10:16.160)
when we don't even know what it is
Lex Fridman (2:10:17.260)
we're incentivized to create.
Lex Fridman (2:10:18.800)
And if you said mimic a human intelligence,
Lex Fridman (2:10:23.280)
I just think there are so many challenges
David Ferrucci (2:10:25.520)
with the significance and meaning of that.
Lex Fridman (2:10:27.720)
That there's not a clear directive.
David Ferrucci (2:10:29.640)
There's no clear directive to do precisely that thing.
Lex Fridman (2:10:32.280)
So assistance in a larger and larger number of tasks.
Lex Fridman (2:10:36.480)
So being able to,
Lex Fridman (2:10:38.080)
a system that's particularly able to operate my microwave
Lex Fridman (2:10:41.080)
and making a grilled cheese sandwich.
Lex Fridman (2:10:42.600)
I don't even know how to make one of those.
Lex Fridman (2:10:44.960)
And then the same system will be doing the vacuum cleaning.
Lex Fridman (2:10:48.020)
And then the same system would be teaching
David Ferrucci (2:10:53.540)
my kids that I don't have math.
Lex Fridman (2:10:56.280)
I think that when you get into a general intelligence
David Ferrucci (2:11:00.720)
for learning physical tasks,
Lex Fridman (2:11:04.240)
and again, I wanna go back to your body question
David Ferrucci (2:11:06.000)
because I think your body question was interesting,
Lex Fridman (2:11:07.260)
but you wanna go back to learning the abilities
David Ferrucci (2:11:11.080)
to physical tasks.
Lex Fridman (2:11:11.920)
You might have, we might get,
David Ferrucci (2:11:14.420)
I imagine in that timeframe,
Lex Fridman (2:11:16.020)
we will get better and better at learning these kinds
David Ferrucci (2:11:18.440)
of tasks, whether it's mowing your lawn
Lex Fridman (2:11:20.320)
or driving a car or whatever it is.
David Ferrucci (2:11:22.720)
I think we will get better and better at that
Lex Fridman (2:11:24.420)
where it's learning how to make predictions
David Ferrucci (2:11:25.840)
over large bodies of data.
Lex Fridman (2:11:27.000)
I think we're gonna continue to get better
Lex Fridman (2:11:28.280)
and better at that.
Lex Fridman (2:11:30.520)
And machines will outpace humans
David Ferrucci (2:11:33.520)
in a variety of those things.
Lex Fridman (2:11:35.560)
The underlying mechanisms for doing that may be the same,
David Ferrucci (2:11:40.760)
meaning that maybe these are deep nets,
Lex Fridman (2:11:43.680)
there's infrastructure to train them,
David Ferrucci (2:11:46.280)
reusable components to get them to do different classes
Lex Fridman (2:11:49.840)
of tasks, and we get better and better
David Ferrucci (2:11:51.920)
at building these kinds of machines.
Lex Fridman (2:11:53.980)
You could argue that the general learning infrastructure
David Ferrucci (2:11:56.600)
in there is a form of a general type of intelligence.
Lex Fridman (2:12:01.040)
I think what starts getting harder is this notion of,
David Ferrucci (2:12:06.400)
can we effectively communicate and understand
Lex Fridman (2:12:09.120)
and build that shared understanding?
David Ferrucci (2:12:10.840)
Because of the layers of interpretation that are required
Lex Fridman (2:12:13.200)
to do that, and the need for the machine to be engaged
David Ferrucci (2:12:16.320)
with humans at that level in a continuous basis.
Lex Fridman (2:12:20.320)
So how do you get the machine in the game?
Lex Fridman (2:12:23.480)
How do you get the machine in the intellectual game?
Lex Fridman (2:12:26.600)
Yeah, and to solve AGI,
David Ferrucci (2:12:29.120)
you probably have to solve that problem.
Lex Fridman (2:12:31.000)
You have to get the machine,
Lex Fridman (2:12:31.920)
so it's a little bit of a bootstrapping thing.
Lex Fridman (2:12:33.800)
Can we get the machine engaged in the intellectual game,
Lex Fridman (2:12:39.160)
but in the intellectual dialogue with the humans?
Lex Fridman (2:12:42.360)
Are the humans sufficiently in intellectual dialogue
Lex Fridman (2:12:44.840)
with each other to generate enough data in this context?
Lex Fridman (2:12:49.640)
And how do you bootstrap that?
David Ferrucci (2:12:51.020)
Because every one of those conversations,
Lex Fridman (2:12:54.080)
every one of those conversations,
David Ferrucci (2:12:55.760)
those intelligent interactions,
Lex Fridman (2:12:58.040)
require so much prior knowledge
David Ferrucci (2:12:59.680)
that it's a challenge to bootstrap it.
Lex Fridman (2:13:01.640)
So the question is, and how committed?
Lex Fridman (2:13:05.840)
So I think that's possible, but when I go back to,
Lex Fridman (2:13:08.800)
are we incentivized to do that?
David Ferrucci (2:13:10.880)
I know we're incentivized to do the former.
Lex Fridman (2:13:13.160)
Are we incentivized to do the latter significantly enough?
Lex Fridman (2:13:15.900)
Do people understand what the latter really is well enough?
Lex Fridman (2:13:18.460)
Part of the elemental cognition mission
David Ferrucci (2:13:20.880)
is to try to articulate that better and better
Lex Fridman (2:13:23.520)
through demonstrations
Lex Fridman (2:13:24.560)
and through trying to craft these grand challenges
Lex Fridman (2:13:26.960)
and get people to say, look,
David Ferrucci (2:13:28.120)
this is a class of intelligence.
Lex Fridman (2:13:30.440)
This is a class of AI.
Lex Fridman (2:13:31.840)
Do we want this?
Lex Fridman (2:13:33.420)
What is the potential of this?
Lex Fridman (2:13:35.800)
What's the business potential?
Lex Fridman (2:13:37.840)
What's the societal potential to that?
Lex Fridman (2:13:40.120)
And to build up that incentive system around that.
Lex Fridman (2:13:45.080)
Yeah, I think if people don't understand yet,
David Ferrucci (2:13:46.820)
I think they will.
Lex Fridman (2:13:47.660)
I think there's a huge business potential here.
Lex Fridman (2:13:49.620)
So it's exciting that you're working on it.
Lex Fridman (2:13:54.000)
We kind of skipped over,
Lex Fridman (2:13:54.960)
but I'm a huge fan of physical presence of things.
Lex Fridman (2:13:59.560)
Do you think Watson had a body?
Lex Fridman (2:14:03.320)
Do you think having a body adds to the interactive element
Lex Fridman (2:14:08.320)
between the AI system and a human,
Lex Fridman (2:14:11.640)
or just in general to intelligence?
Lex Fridman (2:14:14.600)
So I think going back to that shared understanding bit,
David Ferrucci (2:14:19.780)
humans are very connected to their bodies.
Lex Fridman (2:14:21.640)
I mean, one of the challenges in getting an AI
David Ferrucci (2:14:26.320)
to kind of be a compatible human intelligence
Lex Fridman (2:14:29.120)
is that our physical bodies are generating a lot of features
David Ferrucci (2:14:33.660)
that make up the input.
Lex Fridman (2:14:37.720)
So in other words, our bodies are the tool
David Ferrucci (2:14:40.800)
we use to affect output,
Lex Fridman (2:14:42.720)
but they also generate a lot of input for our brains.
Lex Fridman (2:14:46.360)
So we generate emotion, we generate all these feelings,
Lex Fridman (2:14:49.800)
we generate all these signals that machines don't have.
Lex Fridman (2:14:52.720)
So machines don't have this as the input data
Lex Fridman (2:14:56.800)
and they don't have the feedback that says,
David Ferrucci (2:14:58.720)
I've gotten this emotion or I've gotten this idea,
Lex Fridman (2:15:02.940)
I now want to process it,
Lex Fridman (2:15:04.320)
and then it then affects me as a physical being,
Lex Fridman (2:15:08.960)
and I can play that out.
David Ferrucci (2:15:12.200)
In other words, I could realize the implications of that,
Lex Fridman (2:15:14.120)
implications again, on my mind body complex,
David Ferrucci (2:15:17.520)
I then process that, and the implications again,
Lex Fridman (2:15:19.960)
our internal features are generated, I learn from them,
David Ferrucci (2:15:23.620)
they have an effect on my mind body complex.
Lex Fridman (2:15:26.760)
So it's interesting when we think,
Lex Fridman (2:15:28.900)
do we want a human intelligence?
Lex Fridman (2:15:30.440)
Well, if we want a human compatible intelligence,
David Ferrucci (2:15:33.200)
probably the best thing to do
Lex Fridman (2:15:34.320)
is to embed it in a human body.
David Ferrucci (2:15:36.840)
Just to clarify, and both concepts are beautiful,
Lex Fridman (2:15:39.960)
is humanoid robots, so robots that look like humans is one,
David Ferrucci (2:15:45.440)
or did you mean actually sort of what Elon Musk
Lex Fridman (2:15:50.720)
was working with Neuralink,
David Ferrucci (2:15:52.940)
really embedding intelligence systems
Lex Fridman (2:15:55.840)
to ride along human bodies?
David Ferrucci (2:15:59.800)
No, I mean riding along is different.
Lex Fridman (2:16:01.840)
I meant like if you want to create an intelligence
David Ferrucci (2:16:05.840)
that is human compatible,
Lex Fridman (2:16:08.720)
meaning that it can learn and develop
David Ferrucci (2:16:10.840)
a shared understanding of the world around it,
Lex Fridman (2:16:13.040)
you have to give it a lot of the same substrate.
David Ferrucci (2:16:15.120)
Part of that substrate is the idea
Lex Fridman (2:16:18.200)
that it generates these kinds of internal features,
David Ferrucci (2:16:21.120)
like sort of emotional stuff, it has similar senses,
Lex Fridman (2:16:24.020)
it has to do a lot of the same things
Lex Fridman (2:16:25.640)
with those same senses, right?
Lex Fridman (2:16:28.200)
So I think if you want that,
David Ferrucci (2:16:29.760)
again, I don't know that you want that.
Lex Fridman (2:16:32.520)
That's not my specific goal,
David Ferrucci (2:16:34.280)
I think that's a fascinating scientific goal,
Lex Fridman (2:16:35.860)
I think it has all kinds of other implications.
David Ferrucci (2:16:37.860)
That's sort of not the goal.
Lex Fridman (2:16:39.480)
I want to create, I think of it
David Ferrucci (2:16:41.600)
as I create intellectual thought partners for humans,
Lex Fridman (2:16:44.120)
so that kind of intelligence.
David Ferrucci (2:16:47.640)
I know there are other companies
Lex Fridman (2:16:48.560)
that are creating physical thought partners,
David Ferrucci (2:16:50.160)
physical partners for humans,
Lex Fridman (2:16:52.460)
but that's kind of not where I'm at.
Lex Fridman (2:16:56.420)
But the important point is that a big part
Lex Fridman (2:17:00.760)
of what we process is that physical experience
David Ferrucci (2:17:06.240)
of the world around us.
Lex Fridman (2:17:08.080)
On the point of thought partners,
Lex Fridman (2:17:10.520)
what role does an emotional connection,
Lex Fridman (2:17:13.920)
or forgive me, love, have to play
Lex Fridman (2:17:17.820)
in that thought partnership?
Lex Fridman (2:17:19.840)
Is that something you're interested in,
David Ferrucci (2:17:22.000)
put another way, sort of having a deep connection,
Lex Fridman (2:17:26.700)
beyond intellectual?
Lex Fridman (2:17:29.300)
With the AI?
Lex Fridman (2:17:30.200)
Yeah, with the AI, between human and AI.
David Ferrucci (2:17:32.740)
Is that something that gets in the way
Lex Fridman (2:17:34.440)
of the rational discourse?
Lex Fridman (2:17:37.560)
Is that something that's useful?
Lex Fridman (2:17:39.240)
I worry about biases, obviously.
Lex Fridman (2:17:41.920)
So in other words, if you develop an emotional relationship
Lex Fridman (2:17:44.280)
with a machine, all of a sudden you start,
David Ferrucci (2:17:46.640)
are more likely to believe what it's saying,
Lex Fridman (2:17:48.320)
even if it doesn't make any sense.
Lex Fridman (2:17:50.240)
So I worry about that.
Lex Fridman (2:17:53.640)
But at the same time,
David Ferrucci (2:17:54.600)
I think the opportunity to use machines
Lex Fridman (2:17:56.560)
to provide human companionship is actually not crazy.
Lex Fridman (2:17:59.440)
And intellectual and social companionship
Lex Fridman (2:18:04.060)
is not a crazy idea.
Lex Fridman (2:18:06.320)
Do you have concerns, as a few people do,
Lex Fridman (2:18:09.960)
Elon Musk, Sam Harris,
David Ferrucci (2:18:11.760)
about long term existential threats of AI,
Lex Fridman (2:18:15.460)
and perhaps short term threats of AI?
David Ferrucci (2:18:18.760)
We talked about bias,
Lex Fridman (2:18:19.800)
we talked about different misuses,
Lex Fridman (2:18:21.120)
but do you have concerns about thought partners,
Lex Fridman (2:18:25.680)
systems that are able to help us make decisions
David Ferrucci (2:18:28.600)
together as humans,
Lex Fridman (2:18:29.780)
somehow having a significant negative impact
Lex Fridman (2:18:31.960)
on society in the long term?
Lex Fridman (2:18:33.760)
I think there are things to worry about.
David Ferrucci (2:18:35.340)
I think giving machines too much leverage is a problem.
Lex Fridman (2:18:41.500)
And what I mean by leverage is,
David Ferrucci (2:18:44.320)
is too much control over things that can hurt us,
Lex Fridman (2:18:47.040)
whether it's socially, psychologically, intellectually,
David Ferrucci (2:18:50.240)
or physically.
Lex Fridman (2:18:51.640)
And if you give the machines too much control,
David Ferrucci (2:18:53.480)
I think that's a concern.
Lex Fridman (2:18:54.760)
You forget about the AI,
David Ferrucci (2:18:56.320)
just once you give them too much control,
Lex Fridman (2:18:58.640)
human bad actors can hack them and produce havoc.
Lex Fridman (2:19:04.760)
So that's a problem.
Lex Fridman (2:19:07.000)
And you'd imagine hackers taking over
David Ferrucci (2:19:10.040)
the driverless car network
Lex Fridman (2:19:11.080)
and creating all kinds of havoc.
Lex Fridman (2:19:15.220)
But you could also imagine given the ease
Lex Fridman (2:19:19.760)
at which humans could be persuaded one way or the other,
Lex Fridman (2:19:22.800)
and now we have algorithms that can easily take control
Lex Fridman (2:19:25.840)
over that and amplify noise
Lex Fridman (2:19:29.640)
and move people one direction or another.
Lex Fridman (2:19:32.000)
I mean, humans do that to other humans all the time.
Lex Fridman (2:19:34.140)
And we have marketing campaigns,
Lex Fridman (2:19:35.420)
we have political campaigns that take advantage
David Ferrucci (2:19:38.220)
of our emotions or our fears.
Lex Fridman (2:19:41.960)
And this is done all the time.
Lex Fridman (2:19:44.160)
But with machines, machines are like giant megaphones.
Lex Fridman (2:19:47.760)
We can amplify this in orders of magnitude
Lex Fridman (2:19:50.680)
and fine tune its control so we can tailor the message.
Lex Fridman (2:19:54.840)
We can now very rapidly and efficiently tailor the message
David Ferrucci (2:19:58.640)
to the audience, taking advantage of their biases
Lex Fridman (2:20:04.200)
and amplifying them and using them to persuade them
David Ferrucci (2:20:06.640)
in one direction or another in ways that are not fair,
Lex Fridman (2:20:10.740)
not logical, not objective, not meaningful.
Lex Fridman (2:20:13.440)
And humans, machines empower that.
Lex Fridman (2:20:17.000)
So that's what I mean by leverage.
David Ferrucci (2:20:18.920)
Like it's not new, but wow, it's powerful
Lex Fridman (2:20:22.840)
because machines can do it more effectively,
David Ferrucci (2:20:24.400)
more quickly and we see that already going on
Lex Fridman (2:20:27.720)
in social media and other places.
David Ferrucci (2:20:31.720)
That's scary.
Lex Fridman (2:20:33.100)
And that's why I go back to saying one of the most important
David Ferrucci (2:20:38.100)
That's why I go back to saying one of the most important
Lex Fridman (2:20:42.860)
public dialogues we could be having
David Ferrucci (2:20:45.420)
is about the nature of intelligence
Lex Fridman (2:20:47.980)
and the nature of inference and logic
Lex Fridman (2:20:52.140)
and reason and rationality and us understanding
Lex Fridman (2:20:56.100)
our own biases, us understanding our own cognitive biases
Lex Fridman (2:20:59.820)
and how they work and then how machines work
Lex Fridman (2:21:03.140)
and how do we use them to compliment basically
Lex Fridman (2:21:06.020)
so that in the end we have a stronger overall system.
Lex Fridman (2:21:09.660)
That's just incredibly important.
David Ferrucci (2:21:13.020)
I don't think most people understand that.
Lex Fridman (2:21:15.780)
So like telling your kids or telling your students,
David Ferrucci (2:21:20.620)
this goes back to the cognition.
Lex Fridman (2:21:22.540)
Here's how your brain works.
Lex Fridman (2:21:24.300)
Here's how easy it is to trick your brain, right?
Lex Fridman (2:21:28.060)
There are fundamental cognitive,
David Ferrucci (2:21:29.460)
you should appreciate the different types of thinking
Lex Fridman (2:21:34.060)
and how they work and what you're prone to
Lex Fridman (2:21:36.820)
and what do you prefer?
Lex Fridman (2:21:40.580)
And under what conditions does this make sense
Lex Fridman (2:21:42.340)
versus does that make sense?
Lex Fridman (2:21:43.620)
And then say, here's what AI can do.
David Ferrucci (2:21:46.340)
Here's how it can make this worse
Lex Fridman (2:21:48.620)
and here's how it can make this better.
Lex Fridman (2:21:51.020)
And then that's where the AI has a role
Lex Fridman (2:21:52.740)
is to reveal that trade off.
Lex Fridman (2:21:56.620)
So if you imagine a system that is able to
Lex Fridman (2:22:00.700)
beyond any definition of the Turing test to the benchmark,
David Ferrucci (2:22:06.960)
really an AGI system as a thought partner
Lex Fridman (2:22:10.240)
that you one day will create,
Lex Fridman (2:22:14.320)
what question, what topic of discussion,
Lex Fridman (2:22:19.520)
if you get to pick one, would you have with that system?
Lex Fridman (2:22:23.960)
What would you ask and you get to find out
Lex Fridman (2:22:28.240)
the truth together?
Lex Fridman (2:22:33.440)
So you threw me a little bit with finding the truth
Lex Fridman (2:22:36.200)
at the end, but because the truth is a whole nother topic.
Lex Fridman (2:22:41.040)
But I think the beauty of it,
Lex Fridman (2:22:43.600)
I think what excites me is the beauty of it is
David Ferrucci (2:22:46.060)
if I really have that system, I don't have to pick.
Lex Fridman (2:22:48.700)
So in other words, I can go to and say,
David Ferrucci (2:22:51.600)
this is what I care about today.
Lex Fridman (2:22:54.060)
And that's what we mean by like this general capability,
David Ferrucci (2:22:57.180)
go out, read this stuff in the next three milliseconds.
Lex Fridman (2:23:00.560)
And I wanna talk to you about it.
David Ferrucci (2:23:02.800)
I wanna draw analogies, I wanna understand
Lex Fridman (2:23:05.000)
how this affects this decision or that decision.
Lex Fridman (2:23:08.080)
What if this were true?
Lex Fridman (2:23:09.200)
What if that were true?
Lex Fridman (2:23:10.720)
What knowledge should I be aware of
Lex Fridman (2:23:13.200)
that could impact my decision?
David Ferrucci (2:23:16.000)
Here's what I'm thinking is the main implication.
Lex Fridman (2:23:18.960)
Can you prove that out?
Lex Fridman (2:23:21.120)
Can you give me the evidence that supports that?
Lex Fridman (2:23:23.280)
Can you give me evidence that supports this other thing?
Lex Fridman (2:23:25.600)
Boy, would that be incredible?
Lex Fridman (2:23:27.400)
Would that be just incredible?
David Ferrucci (2:23:28.560)
Just a long discourse.
Lex Fridman (2:23:30.400)
Just to be part of whether it's a medical diagnosis
David Ferrucci (2:23:33.340)
or whether it's the various treatment options
Lex Fridman (2:23:35.880)
or whether it's a legal case
David Ferrucci (2:23:38.400)
or whether it's a social problem
Lex Fridman (2:23:40.040)
that people are discussing,
David Ferrucci (2:23:41.040)
like be part of the dialogue,
Lex Fridman (2:23:43.740)
one that holds itself and us accountable
David Ferrucci (2:23:49.520)
to reasons and objective dialogue.
Lex Fridman (2:23:52.760)
I get goosebumps talking about it, right?
David Ferrucci (2:23:54.520)
It's like, this is what I want.
Lex Fridman (2:23:57.440)
So when you create it, please come back on the podcast
Lex Fridman (2:24:01.000)
and we can have a discussion together
Lex Fridman (2:24:03.480)
and make it even longer.
David Ferrucci (2:24:04.800)
This is a record for the longest conversation
Lex Fridman (2:24:07.320)
in the world.
David Ferrucci (2:24:08.160)
It was an honor, it was a pleasure, David.
Lex Fridman (2:24:09.400)
Thank you so much for talking to me.
David Ferrucci (2:24:10.240)
Thanks so much, a lot of fun.
Lex Fridman (30:02.360)
What are your prior experiences with those artifacts?
Lex Fridman (30:05.440)
What are your fundamental assumptions and values?
Lex Fridman (30:08.800)
What is your ability to kind of reason,
David Ferrucci (30:10.840)
to chain together logical implication
Lex Fridman (30:13.640)
as you're sitting there and saying,
David Ferrucci (30:14.480)
well, if this is the case, then I would conclude this.
Lex Fridman (30:16.520)
And if that's the case, then I would conclude that.
Lex Fridman (30:19.120)
So your reasoning processes and how they work,
Lex Fridman (30:22.520)
your prior models and what they are,
David Ferrucci (30:25.360)
your values and your assumptions,
Lex Fridman (30:27.200)
all those things now come together into the interpretation.
David Ferrucci (30:30.640)
Getting in sync of that is hard.
Lex Fridman (30:34.860)
And yet humans are able to intuit some of that
David Ferrucci (30:37.620)
without any pre.
Lex Fridman (30:39.580)
Because they have the shared experience.
Lex Fridman (30:41.580)
And we're not talking about shared,
Lex Fridman (30:42.940)
two people having shared experience.
David Ferrucci (30:44.420)
I mean, as a society.
Lex Fridman (30:45.540)
That's correct.
David Ferrucci (30:46.540)
We have the shared experience and we have similar brains.
Lex Fridman (30:51.180)
So we tend to, in other words,
David Ferrucci (30:54.060)
part of our shared experiences are shared local experience.
Lex Fridman (30:56.460)
Like we may live in the same culture,
David Ferrucci (30:57.860)
we may live in the same society
Lex Fridman (30:59.060)
and therefore we have similar educations.
David Ferrucci (31:02.020)
We have some of what we like to call prior models
Lex Fridman (31:04.100)
about the word prior experiences.
Lex Fridman (31:05.860)
And we use that as a,
Lex Fridman (31:07.380)
think of it as a wide collection of interrelated variables
Lex Fridman (31:10.940)
and they're all bound to similar things.
Lex Fridman (31:12.780)
And so we take that as our background
Lex Fridman (31:15.060)
and we start interpreting things similarly.
Lex Fridman (31:17.540)
But as humans, we have a lot of shared experience.
David Ferrucci (31:21.860)
We do have similar brains, similar goals,
Lex Fridman (31:24.980)
similar emotions under similar circumstances.
David Ferrucci (31:28.060)
Because we're both humans.
Lex Fridman (31:29.020)
So now one of the early questions you asked,
Lex Fridman (31:31.420)
how is biological and computer information systems
Lex Fridman (31:37.020)
fundamentally different?
David Ferrucci (31:37.980)
Well, one is humans come with a lot of pre programmed stuff.
Lex Fridman (31:43.840)
A ton of program stuff.
Lex Fridman (31:45.940)
And they're able to communicate
Lex Fridman (31:47.220)
because they share that stuff.
Lex Fridman (31:50.340)
Do you think that shared knowledge,
Lex Fridman (31:54.100)
if we can maybe escape the hard work question,
Lex Fridman (31:57.580)
how much is encoded in the hardware?
Lex Fridman (31:59.460)
Just the shared knowledge in the software,
David Ferrucci (32:01.260)
the history, the many centuries of wars and so on
Lex Fridman (32:05.340)
that came to today, that shared knowledge.
Lex Fridman (32:09.660)
How hard is it to encode?
Lex Fridman (32:14.340)
Do you have a hope?
Lex Fridman (32:15.860)
Can you speak to how hard is it to encode that knowledge
Lex Fridman (32:19.340)
systematically in a way that could be used by a computer?
Lex Fridman (32:22.780)
So I think it is possible to learn to,
Lex Fridman (32:25.100)
for a machine to program a machine,
David Ferrucci (32:27.900)
to acquire that knowledge with a similar foundation.
Lex Fridman (32:31.460)
In other words, a similar interpretive foundation
David Ferrucci (32:36.120)
for processing that knowledge.
Lex Fridman (32:38.060)
What do you mean by that?
Lex Fridman (32:39.100)
So in other words, we view the world in a particular way.
Lex Fridman (32:44.540)
So in other words, we have a, if you will,
David Ferrucci (32:48.820)
as humans, we have a framework
Lex Fridman (32:50.120)
for interpreting the world around us.
Lex Fridman (32:52.260)
So we have multiple frameworks for interpreting
Lex Fridman (32:55.220)
the world around us.
Lex Fridman (32:56.060)
But if you're interpreting, for example,
Lex Fridman (32:59.780)
socio political interactions,
David Ferrucci (33:01.340)
you're thinking about where there's people,
Lex Fridman (33:03.140)
there's collections and groups of people,
David Ferrucci (33:05.540)
they have goals, goals largely built around survival
Lex Fridman (33:08.460)
and quality of life.
David Ferrucci (33:10.860)
There are fundamental economics around scarcity of resources.
Lex Fridman (33:16.640)
And when humans come and start interpreting
David Ferrucci (33:19.660)
a situation like that, because you brought up
Lex Fridman (33:21.860)
like historical events,
David Ferrucci (33:23.600)
they start interpreting situations like that.
Lex Fridman (33:25.500)
They apply a lot of this fundamental framework
David Ferrucci (33:29.500)
for interpreting that.
Lex Fridman (33:30.740)
Well, who are the people?
Lex Fridman (33:32.260)
What were their goals?
Lex Fridman (33:33.300)
What resources did they have?
Lex Fridman (33:35.020)
How much power influence did they have over the other?
Lex Fridman (33:37.020)
Like this fundamental substrate, if you will,
David Ferrucci (33:40.540)
for interpreting and reasoning about that.
Lex Fridman (33:43.820)
So I think it is possible to imbue a computer
David Ferrucci (33:46.920)
with that stuff that humans like take for granted
Lex Fridman (33:50.660)
when they go and sit down and try to interpret things.
Lex Fridman (33:54.020)
And then with that foundation, they acquire,
Lex Fridman (33:58.860)
they start acquiring the details,
David Ferrucci (34:00.300)
the specifics in a given situation,
Lex Fridman (34:02.820)
are then able to interpret it with regard to that framework.
Lex Fridman (34:05.700)
And then given that interpretation, they can do what?
Lex Fridman (34:08.700)
They can predict.
Lex Fridman (34:10.300)
But not only can they predict,
Lex Fridman (34:12.220)
they can predict now with an explanation
David Ferrucci (34:15.940)
that can be given in those terms,
Lex Fridman (34:17.940)
in the terms of that underlying framework
David Ferrucci (34:20.200)
that most humans share.
Lex Fridman (34:22.300)
Now you could find humans that come and interpret events
David Ferrucci (34:24.620)
very differently than other humans
Lex Fridman (34:26.300)
because they're like using a different framework.
David Ferrucci (34:30.620)
The movie Matrix comes to mind
Lex Fridman (34:32.500)
where they decided humans were really just batteries,
Lex Fridman (34:36.420)
and that's how they interpreted the value of humans
Lex Fridman (34:39.940)
as a source of electrical energy.
David Ferrucci (34:41.640)
So, but I think that for the most part,
Lex Fridman (34:45.460)
we have a way of interpreting the events
David Ferrucci (34:50.780)
or the social events around us
Lex Fridman (34:52.260)
because we have this shared framework.
David Ferrucci (34:54.140)
It comes from, again, the fact that we're similar beings
Lex Fridman (34:58.700)
that have similar goals, similar emotions,
Lex Fridman (35:01.100)
and we can make sense out of these.
Lex Fridman (35:02.900)
These frameworks make sense to us.
Lex Fridman (35:05.020)
So how much knowledge is there, do you think?
Lex Fridman (35:08.060)
So you said it's possible.
David Ferrucci (35:09.580)
Well, there's a tremendous amount of detailed knowledge
Lex Fridman (35:12.140)
in the world.
David Ferrucci (35:12.980)
You could imagine effectively infinite number
Lex Fridman (35:17.580)
of unique situations and unique configurations
David Ferrucci (35:20.840)
of these things.
Lex Fridman (35:22.100)
But the knowledge that you need,
Lex Fridman (35:25.100)
what I refer to as like the frameworks,
Lex Fridman (35:27.600)
for you need for interpreting them, I don't think.
David Ferrucci (35:29.580)
I think those are finite.
Lex Fridman (35:31.500)
You think the frameworks are more important
Lex Fridman (35:35.020)
than the bulk of the knowledge?
Lex Fridman (35:36.780)
So it's like framing.
David Ferrucci (35:37.780)
Yeah, because what the frameworks do
Lex Fridman (35:39.220)
is they give you now the ability to interpret and reason,
Lex Fridman (35:41.580)
and to interpret and reason,
Lex Fridman (35:43.100)
to interpret and reason over the specifics
David Ferrucci (35:46.780)
in ways that other humans would understand.
Lex Fridman (35:49.220)
What about the specifics?
David Ferrucci (35:51.240)
You know, you acquire the specifics by reading
Lex Fridman (35:53.980)
and by talking to other people.
Lex Fridman (35:55.540)
So I'm mostly actually just even,
Lex Fridman (35:57.700)
if we can focus on even the beginning,
David Ferrucci (36:00.240)
the common sense stuff,
Lex Fridman (36:01.500)
the stuff that doesn't even require reading,
David Ferrucci (36:03.420)
or it almost requires playing around with the world
Lex Fridman (36:06.860)
or something, just being able to sort of manipulate objects,
David Ferrucci (36:10.820)
drink water and so on, all of that.
Lex Fridman (36:13.900)
Every time we try to do that kind of thing
David Ferrucci (36:16.140)
in robotics or AI, it seems to be like an onion.
Lex Fridman (36:21.060)
You seem to realize how much knowledge
David Ferrucci (36:23.240)
is really required to perform
Lex Fridman (36:24.620)
even some of these basic tasks.
Lex Fridman (36:27.060)
Do you have that sense as well?
Lex Fridman (36:30.340)
And if so, how do we get all those details?
Lex Fridman (36:33.820)
Are they written down somewhere?
Lex Fridman (36:35.700)
Do they have to be learned through experience?
Lex Fridman (36:39.220)
So I think when, like, if you're talking about
Lex Fridman (36:41.340)
sort of the physics, the basic physics around us,
David Ferrucci (36:44.700)
for example, acquiring information about,
Lex Fridman (36:46.580)
acquiring how that works.
David Ferrucci (36:49.720)
Yeah, I mean, I think there's a combination of things going,
Lex Fridman (36:52.220)
I think there's a combination of things going on.
David Ferrucci (36:54.620)
I think there is like fundamental pattern matching,
Lex Fridman (36:57.780)
like what we were talking about before,
David Ferrucci (36:59.660)
where you see enough examples,
Lex Fridman (37:01.060)
enough data about something and you start assuming that.
Lex Fridman (37:03.840)
And with similar input,
Lex Fridman (37:05.480)
I'm gonna predict similar outputs.
David Ferrucci (37:07.720)
You can't necessarily explain it at all.
Lex Fridman (37:10.100)
You may learn very quickly that when you let something go,
David Ferrucci (37:14.640)
it falls to the ground.
Lex Fridman (37:16.500)
But you can't necessarily explain that.
Lex Fridman (37:19.760)
But that's such a deep idea,
Lex Fridman (37:22.340)
that if you let something go, like the idea of gravity.
David Ferrucci (37:26.120)
I mean, people are letting things go
Lex Fridman (37:27.900)
and counting on them falling
David Ferrucci (37:29.100)
well before they understood gravity.
Lex Fridman (37:30.760)
But that seems to be, that's exactly what I mean,
David Ferrucci (37:33.860)
is before you take a physics class
Lex Fridman (37:36.080)
or study anything about Newton,
David Ferrucci (37:39.540)
just the idea that stuff falls to the ground
Lex Fridman (37:42.540)
and then you'd be able to generalize
David Ferrucci (37:45.300)
that all kinds of stuff falls to the ground.
Lex Fridman (37:49.540)
It just seems like a non, without encoding it,
David Ferrucci (37:53.420)
like hard coding it in,
Lex Fridman (37:55.220)
it seems like a difficult thing to pick up.
David Ferrucci (37:57.420)
It seems like you have to have a lot of different knowledge
Lex Fridman (38:01.380)
to be able to integrate that into the framework,
David Ferrucci (38:05.340)
sort of into everything else.
Lex Fridman (38:07.700)
So both know that stuff falls to the ground
Lex Fridman (38:10.340)
and start to reason about sociopolitical discourse.
Lex Fridman (38:16.340)
So both, like the very basic
Lex Fridman (38:18.540)
and the high level reasoning decision making.
Lex Fridman (38:22.540)
I guess my question is, how hard is this problem?
Lex Fridman (38:26.420)
And sorry to linger on it because again,
Lex Fridman (38:29.060)
and we'll get to it for sure,
David Ferrucci (38:31.100)
as what Watson with Jeopardy did is take on a problem
Lex Fridman (38:34.340)
that's much more constrained
Lex Fridman (38:35.500)
but has the same hugeness of scale,
Lex Fridman (38:38.260)
at least from the outsider's perspective.
Lex Fridman (38:40.660)
So I'm asking the general life question
Lex Fridman (38:42.900)
of to be able to be an intelligent being
Lex Fridman (38:45.580)
and reason in the world about both gravity and politics,
Lex Fridman (38:50.900)
how hard is that problem?
Lex Fridman (38:53.900)
So I think it's solvable.
Lex Fridman (38:59.440)
Okay, now beautiful.
Lex Fridman (39:00.700)
So what about time travel?
Lex Fridman (39:04.820)
Okay, I'm just saying the same answer.
David Ferrucci (39:08.700)
Not as convinced.
Lex Fridman (39:09.700)
Not as convinced yet, okay.
David Ferrucci (39:11.100)
No, I think it is solvable.
Lex Fridman (39:14.260)
I mean, I think that it's a learn,
David Ferrucci (39:16.500)
first of all, it's about getting machines to learn.
Lex Fridman (39:18.440)
Learning is fundamental.
Lex Fridman (39:21.380)
And I think we're already in a place that we understand,
Lex Fridman (39:24.420)
for example, how machines can learn in various ways.
David Ferrucci (39:28.620)
Right now, our learning stuff is sort of primitive
Lex Fridman (39:32.460)
in that we haven't sort of taught machines
David Ferrucci (39:38.040)
to learn the frameworks.
Lex Fridman (39:39.260)
We don't communicate our frameworks
David Ferrucci (39:41.160)
because of how shared they are, in some cases we do,
Lex Fridman (39:42.860)
but we don't annotate, if you will,
David Ferrucci (39:46.380)
all the data in the world with the frameworks
Lex Fridman (39:48.960)
that are inherent or underlying our understanding.
David Ferrucci (39:53.120)
Instead, we just operate with the data.
Lex Fridman (39:56.180)
So if we wanna be able to reason over the data
David Ferrucci (39:59.100)
in similar terms in the common frameworks,
Lex Fridman (40:02.300)
we need to be able to teach the computer,
David Ferrucci (40:03.740)
or at least we need to program the computer
Lex Fridman (40:06.300)
to acquire, to have access to and acquire,
David Ferrucci (40:10.480)
learn the frameworks as well
Lex Fridman (40:12.860)
and connect the frameworks to the data.
David Ferrucci (40:15.740)
I think this can be done.
Lex Fridman (40:18.420)
I think we can start, I think machine learning,
David Ferrucci (40:22.980)
for example, with enough examples,
Lex Fridman (40:26.100)
can start to learn these basic dynamics.
Lex Fridman (40:28.920)
Will they relate them necessarily to the gravity?
Lex Fridman (40:32.240)
Not unless they can also acquire those theories as well
Lex Fridman (40:38.320)
and put the experiential knowledge
Lex Fridman (40:40.940)
and connect it back to the theoretical knowledge.
David Ferrucci (40:43.400)
I think if we think in terms of these class of architectures
Lex Fridman (40:47.200)
that are designed to both learn the specifics,
David Ferrucci (40:51.020)
find the patterns, but also acquire the frameworks
Lex Fridman (40:54.220)
and connect the data to the frameworks.
David Ferrucci (40:56.340)
If we think in terms of robust architectures like this,
Lex Fridman (40:59.700)
I think there is a path toward getting there.
David Ferrucci (41:03.420)
In terms of encoding architectures like that,
Lex Fridman (41:06.220)
do you think systems that are able to do this
David Ferrucci (41:10.300)
will look like neural networks or representing,
Lex Fridman (41:14.940)
if you look back to the 80s and 90s with the expert systems,
David Ferrucci (41:18.740)
they're more like graphs, systems that are based in logic,
Lex Fridman (41:24.540)
able to contain a large amount of knowledge
David Ferrucci (41:26.500)
where the challenge was the automated acquisition
Lex Fridman (41:28.500)
of that knowledge.
David Ferrucci (41:29.860)
I guess the question is when you collect both the frameworks
Lex Fridman (41:33.820)
and the knowledge from the data,
Lex Fridman (41:35.300)
what do you think that thing will look like?
Lex Fridman (41:37.260)
Yeah, so I mean, I think asking the question,
David Ferrucci (41:39.340)
they look like neural networks is a bit of a red herring.
Lex Fridman (41:41.260)
I mean, I think that they will certainly do inductive
David Ferrucci (41:45.180)
or pattern match based reasoning.
Lex Fridman (41:46.720)
And I've already experimented with architectures
David Ferrucci (41:49.000)
that combine both that use machine learning
Lex Fridman (41:52.700)
and neural networks to learn certain classes of knowledge,
David Ferrucci (41:55.340)
in other words, to find repeated patterns
Lex Fridman (41:57.300)
in order for it to make good inductive guesses,
Lex Fridman (42:01.540)
but then ultimately to try to take those learnings
Lex Fridman (42:05.260)
and marry them, in other words, connect them to frameworks
Lex Fridman (42:09.540)
so that it can then reason over that
Lex Fridman (42:11.500)
in terms other humans understand.
Lex Fridman (42:13.660)
So for example, at elemental cognition, we do both.
Lex Fridman (42:16.100)
We have architectures that do both, both those things,
Lex Fridman (42:19.820)
but also have a learning method
Lex Fridman (42:21.660)
for acquiring the frameworks themselves and saying,
David Ferrucci (42:24.400)
look, ultimately, I need to take this data.
Lex Fridman (42:27.280)
I need to interpret it in the form of these frameworks
Lex Fridman (42:30.020)
so they can reason over it.
Lex Fridman (42:30.860)
So there is a fundamental knowledge representation,
David Ferrucci (42:33.340)
like what you're saying,
Lex Fridman (42:34.220)
like these graphs of logic, if you will.
David Ferrucci (42:36.780)
There are also neural networks
Lex Fridman (42:39.280)
that acquire a certain class of information.
David Ferrucci (42:43.100)
Then they then align them with these frameworks,
Lex Fridman (42:45.900)
but there's also a mechanism
David Ferrucci (42:47.140)
to acquire the frameworks themselves.
Lex Fridman (42:49.180)
Yeah, so it seems like the idea of frameworks
David Ferrucci (42:52.540)
requires some kind of collaboration with humans.
Lex Fridman (42:55.380)
Absolutely.
Lex Fridman (42:56.300)
So do you think of that collaboration as direct?
Lex Fridman (42:59.340)
Well, and let's be clear.
David Ferrucci (43:01.900)
Only for the express purpose that you're designing,
Lex Fridman (43:06.060)
you're designing an intelligence
David Ferrucci (43:09.500)
that can ultimately communicate with humans
Lex Fridman (43:12.500)
in the terms of frameworks that help them understand things.
Lex Fridman (43:17.060)
So to be really clear,
Lex Fridman (43:19.380)
you can independently create a machine learning system,
David Ferrucci (43:24.340)
an intelligence that I might call an alien intelligence
Lex Fridman (43:28.460)
that does a better job than you with some things,
Lex Fridman (43:31.140)
but can't explain the framework to you.
Lex Fridman (43:33.500)
That doesn't mean it might be better than you at the thing.
David Ferrucci (43:36.720)
It might be that you cannot comprehend the framework
Lex Fridman (43:39.500)
that it may have created for itself that is inexplicable
David Ferrucci (43:42.780)
to you.
Lex Fridman (43:43.900)
That's a reality.
Lex Fridman (43:45.260)
But you're more interested in a case where you can.
Lex Fridman (43:48.780)
I am, yeah.
David Ferrucci (43:51.060)
My sort of approach to AI is because
Lex Fridman (43:54.260)
I've set the goal for myself.
David Ferrucci (43:55.900)
I want machines to be able to ultimately communicate,
Lex Fridman (44:00.320)
understanding with humans.
David Ferrucci (44:01.160)
I want them to be able to acquire and communicate,
Lex Fridman (44:03.460)
acquire knowledge from humans
Lex Fridman (44:04.700)
and communicate knowledge to humans.
Lex Fridman (44:06.980)
They should be using what inductive
David Ferrucci (44:11.580)
machine learning techniques are good at,
Lex Fridman (44:13.700)
which is to observe patterns of data,
David Ferrucci (44:16.780)
whether it be in language or whether it be in images
Lex Fridman (44:19.260)
or videos or whatever,
David Ferrucci (44:23.100)
to acquire these patterns,
Lex Fridman (44:25.420)
to induce the generalizations from those patterns,
Lex Fridman (44:29.340)
but then ultimately to work with humans
Lex Fridman (44:31.220)
to connect them to frameworks, interpretations, if you will,
David Ferrucci (44:34.640)
that ultimately make sense to humans.
Lex Fridman (44:36.700)
Of course, the machine is gonna have the strength
David Ferrucci (44:38.460)
that it has, the richer, longer memory,
Lex Fridman (44:41.420)
but it has the more rigorous reasoning abilities,
David Ferrucci (44:45.380)
the deeper reasoning abilities,
Lex Fridman (44:47.040)
so it'll be an interesting complementary relationship
David Ferrucci (44:51.060)
between the human and the machine.
Lex Fridman (44:53.180)
Do you think that ultimately needs explainability
Lex Fridman (44:55.100)
like a machine?
Lex Fridman (44:55.980)
So if we look, we study, for example,
David Ferrucci (44:57.860)
Tesla autopilot a lot, where humans,
Lex Fridman (45:00.820)
I don't know if you've driven the vehicle,
David Ferrucci (45:02.780)
are aware of what it is.
Lex Fridman (45:04.360)
So you're basically the human and machine
David Ferrucci (45:09.100)
are working together there,
Lex Fridman (45:10.300)
and the human is responsible for their own life
David Ferrucci (45:12.500)
to monitor the system,
Lex Fridman (45:14.220)
and the system fails every few miles,
Lex Fridman (45:18.380)
and so there's hundreds,
Lex Fridman (45:20.500)
there's millions of those failures a day,
Lex Fridman (45:23.620)
and so that's like a moment of interaction.
Lex Fridman (45:25.780)
Do you see?
David Ferrucci (45:26.620)
Yeah, that's exactly right.
Lex Fridman (45:27.900)
That's a moment of interaction
David Ferrucci (45:29.900)
where the machine has learned some stuff,
Lex Fridman (45:34.820)
it has a failure, somehow the failure's communicated,
David Ferrucci (45:38.720)
the human is now filling in the mistake, if you will,
Lex Fridman (45:41.880)
or maybe correcting or doing something
David Ferrucci (45:43.620)
that is more successful in that case,
Lex Fridman (45:45.860)
the computer takes that learning.
Lex Fridman (45:47.900)
So I believe that the collaboration
Lex Fridman (45:50.260)
between human and machine,
David Ferrucci (45:52.300)
I mean, that's sort of a primitive example
Lex Fridman (45:53.900)
and sort of a more,
David Ferrucci (45:56.920)
another example is where the machine's literally talking
Lex Fridman (45:59.220)
to you and saying, look, I'm reading this thing.
David Ferrucci (46:02.740)
I know that the next word might be this or that,
Lex Fridman (46:06.580)
but I don't really understand why.
David Ferrucci (46:08.900)
I have my guess.
Lex Fridman (46:09.940)
Can you help me understand the framework that supports this
Lex Fridman (46:14.060)
and then can kind of acquire that,
Lex Fridman (46:16.060)
take that and reason about it and reuse it
David Ferrucci (46:18.140)
the next time it's reading to try to understand something,
Lex Fridman (46:20.520)
not unlike a human student might do.
David Ferrucci (46:24.760)
I mean, I remember when my daughter was in first grade
Lex Fridman (46:27.480)
and she had a reading assignment about electricity
Lex Fridman (46:32.280)
and somewhere in the text it says,
Lex Fridman (46:35.600)
and electricity is produced by water flowing over turbines
David Ferrucci (46:38.620)
or something like that.
Lex Fridman (46:39.900)
And then there's a question that says,
Lex Fridman (46:41.240)
well, how is electricity created?
Lex Fridman (46:43.140)
And so my daughter comes to me and says,
David Ferrucci (46:45.180)
I mean, I could, you know,
Lex Fridman (46:46.500)
created and produced are kind of synonyms in this case.
Lex Fridman (46:49.200)
So I can go back to the text
Lex Fridman (46:50.620)
and I can copy by water flowing over turbines,
Lex Fridman (46:53.660)
but I have no idea what that means.
Lex Fridman (46:56.120)
Like I don't know how to interpret
David Ferrucci (46:57.620)
water flowing over turbines and what electricity even is.
Lex Fridman (47:00.380)
I mean, I can get the answer right by matching the text,
Lex Fridman (47:04.000)
but I don't have any framework for understanding
Lex Fridman (47:06.140)
what this means at all.
Lex Fridman (47:07.860)
And framework really is, I mean, it's a set of,
Lex Fridman (47:10.500)
not to be mathematical, but axioms of ideas
David Ferrucci (47:14.140)
that you bring to the table and interpreting stuff
Lex Fridman (47:16.340)
and then you build those up somehow.
David Ferrucci (47:18.380)
You build them up with the expectation
Lex Fridman (47:20.460)
that there's a shared understanding of what they are.
David Ferrucci (47:23.780)
Sure, yeah, it's the social, that us humans,
Lex Fridman (47:28.900)
do you have a sense that humans on earth in general
Lex Fridman (47:32.060)
share a set of, like how many frameworks are there?
Lex Fridman (47:36.500)
I mean, it depends on how you bound them, right?
Lex Fridman (47:38.200)
So in other words, how big or small,
Lex Fridman (47:39.900)
like their individual scope,
Lex Fridman (47:42.640)
but there's lots and there are new ones.
Lex Fridman (47:44.220)
I think the way I think about it is kind of in a layer.
David Ferrucci (47:47.620)
I think that the architectures are being layered in that.
Lex Fridman (47:50.020)
There's a small set of primitives.
David Ferrucci (47:53.560)
They allow you the foundation to build frameworks.
Lex Fridman (47:56.260)
And then there may be many frameworks,
Lex Fridman (47:58.360)
but you have the ability to acquire them.
Lex Fridman (48:00.580)
And then you have the ability to reuse them.
David Ferrucci (48:03.020)
I mean, one of the most compelling ways
Lex Fridman (48:04.940)
of thinking about this is a reasoning by analogy,
David Ferrucci (48:07.220)
where I can say, oh, wow,
Lex Fridman (48:08.180)
I've learned something very similar.
David Ferrucci (48:11.340)
I never heard of this game soccer,
Lex Fridman (48:15.240)
but if it's like basketball in the sense
David Ferrucci (48:17.820)
that the goal's like the hoop
Lex Fridman (48:19.580)
and I have to get the ball in the hoop
Lex Fridman (48:20.980)
and I have guards and I have this and I have that,
Lex Fridman (48:23.500)
like where are the similarities
Lex Fridman (48:26.460)
and where are the differences?
Lex Fridman (48:27.740)
And I have a foundation now
David Ferrucci (48:29.120)
for interpreting this new information.
Lex Fridman (48:31.340)
And then the different groups,
David Ferrucci (48:33.260)
like the millennials will have a framework.
Lex Fridman (48:36.380)
And then, you know, the Democrats and Republicans.
David Ferrucci (48:41.660)
Millennials, nobody wants that framework.
Lex Fridman (48:43.820)
Well, I mean, I think, right,
David Ferrucci (48:45.860)
I mean, you're talking about political and social ways
Lex Fridman (48:48.100)
of interpreting the world around them.
Lex Fridman (48:49.860)
And I think these frameworks are still largely,
Lex Fridman (48:51.980)
largely similar.
David Ferrucci (48:52.800)
I think they differ in maybe
Lex Fridman (48:54.540)
what some fundamental assumptions and values are.
David Ferrucci (48:57.380)
Now, from a reasoning perspective,
Lex Fridman (48:59.860)
like the ability to process the framework,
David Ferrucci (49:01.620)
it might not be that different.
Lex Fridman (49:04.160)
The implications of different fundamental values
David Ferrucci (49:06.560)
or fundamental assumptions in those frameworks
Lex Fridman (49:09.460)
may reach very different conclusions.
Lex Fridman (49:12.160)
So from a social perspective,
Lex Fridman (49:14.780)
the conclusions may be very different.
David Ferrucci (49:16.900)
From an intelligence perspective,
Lex Fridman (49:18.420)
I just followed where my assumptions took me.
David Ferrucci (49:21.620)
Yeah, the process itself will look similar.
Lex Fridman (49:23.420)
But that's a fascinating idea
David Ferrucci (49:25.580)
that frameworks really help carve
Lex Fridman (49:30.820)
how a statement will be interpreted.
David Ferrucci (49:33.740)
I mean, having a Democrat and a Republican framework
Lex Fridman (49:40.360)
and then read the exact same statement
Lex Fridman (49:42.180)
and the conclusions that you derive
Lex Fridman (49:44.200)
will be totally different
David Ferrucci (49:45.460)
from an AI perspective is fascinating.
Lex Fridman (49:47.620)
What we would want out of the AI
David Ferrucci (49:49.460)
is to be able to tell you
Lex Fridman (49:51.140)
that this perspective, one perspective,
David Ferrucci (49:53.740)
one set of assumptions is gonna lead you here,
Lex Fridman (49:55.540)
another set of assumptions is gonna lead you there.
Lex Fridman (49:58.700)
And in fact, to help people reason and say,
Lex Fridman (50:01.420)
oh, I see where our differences lie.
David Ferrucci (50:05.220)
I have this fundamental belief about that.
Lex Fridman (50:06.940)
I have this fundamental belief about that.
David Ferrucci (50:09.200)
Yeah, that's quite brilliant.
Lex Fridman (50:10.100)
From my perspective, NLP,
David Ferrucci (50:12.620)
there's this idea that there's one way
Lex Fridman (50:14.140)
to really understand a statement,
Lex Fridman (50:16.100)
but that probably isn't.
Lex Fridman (50:18.780)
There's probably an infinite number of ways
David Ferrucci (50:20.140)
to understand a statement, depending on the question.
Lex Fridman (50:21.980)
There's lots of different interpretations,
Lex Fridman (50:23.420)
and the broader the content, the richer it is.
Lex Fridman (50:31.460)
And so you and I can have very different experiences
David Ferrucci (50:35.260)
with the same text, obviously.
Lex Fridman (50:37.420)
And if we're committed to understanding each other,
David Ferrucci (50:42.300)
we start, and that's the other important point,
Lex Fridman (50:45.260)
if we're committed to understanding each other,
David Ferrucci (50:47.740)
we start decomposing and breaking down our interpretation
Lex Fridman (50:51.860)
to its more and more primitive components
David Ferrucci (50:54.020)
until we get to that point where we say,
Lex Fridman (50:55.900)
oh, I see why we disagree.
Lex Fridman (50:58.260)
And we try to understand how fundamental
Lex Fridman (51:00.500)
that disagreement really is.
Lex Fridman (51:02.220)
But that requires a commitment
Lex Fridman (51:04.580)
to breaking down that interpretation
David Ferrucci (51:06.540)
in terms of that framework in a logical way.
Lex Fridman (51:08.940)
Otherwise, and this is why I think of AI
David Ferrucci (51:12.780)
as really complimenting and helping human intelligence
Lex Fridman (51:16.020)
to overcome some of its biases and its predisposition
David Ferrucci (51:19.860)
to be persuaded by more shallow reasoning
Lex Fridman (51:25.060)
in the sense that we get over this idea,
David Ferrucci (51:26.980)
well, I'm right because I'm Republican,
Lex Fridman (51:29.980)
or I'm right because I'm Democratic,
Lex Fridman (51:31.380)
and someone labeled this as Democratic point of view,
Lex Fridman (51:33.380)
or it has the following keywords in it.
Lex Fridman (51:35.420)
And if the machine can help us break that argument down
Lex Fridman (51:38.500)
and say, wait a second, what do you really think
Lex Fridman (51:41.660)
about this, right?
Lex Fridman (51:42.500)
So essentially holding us accountable
David Ferrucci (51:45.460)
to doing more critical thinking.
Lex Fridman (51:47.540)
We're gonna have to sit and think about this fast.
David Ferrucci (51:49.500)
That's, I love that.
Lex Fridman (51:50.940)
I think that's really empowering use of AI
David Ferrucci (51:53.580)
for the public discourse is completely disintegrating
Lex Fridman (51:57.180)
currently as we learn how to do it on social media.
David Ferrucci (52:00.460)
That's right.
Lex Fridman (52:02.460)
So one of the greatest accomplishments
David Ferrucci (52:05.860)
in the history of AI is Watson competing
Lex Fridman (52:12.140)
in the game of Jeopardy against humans.
Lex Fridman (52:14.700)
And you were a lead in that, a critical part of that.
Lex Fridman (52:18.940)
Let's start at the very basics.
Lex Fridman (52:20.620)
What is the game of Jeopardy?
Lex Fridman (52:22.860)
The game for us humans, human versus human.
David Ferrucci (52:25.860)
Right, so it's to take a question and answer it.
Lex Fridman (52:33.900)
The game of Jeopardy.
David Ferrucci (52:34.740)
It's just the opposite.
Lex Fridman (52:35.580)
Actually, well, no, but it's not, right?
David Ferrucci (52:38.780)
It's really not.
Lex Fridman (52:39.620)
It's really to get a question and answer,
Lex Fridman (52:41.860)
but it's what we call a factoid question.
Lex Fridman (52:43.940)
So this notion of like, it really relates to some fact
David Ferrucci (52:46.860)
that two people would argue
Lex Fridman (52:49.260)
whether the facts are true or not.
David Ferrucci (52:50.580)
In fact, most people wouldn't.
Lex Fridman (52:51.580)
Jeopardy kind of counts on the idea
David Ferrucci (52:53.060)
that these statements have factual answers.
Lex Fridman (52:57.660)
And the idea is to, first of all,
David Ferrucci (53:02.020)
determine whether or not you know the answer,
Lex Fridman (53:03.780)
which is sort of an interesting twist.
Lex Fridman (53:06.100)
So first of all, understand the question.
Lex Fridman (53:07.860)
You have to understand the question.
Lex Fridman (53:08.860)
What is it asking?
Lex Fridman (53:09.860)
And that's a good point
Lex Fridman (53:10.740)
because the questions are not asked directly, right?
Lex Fridman (53:14.460)
They're all like,
David Ferrucci (53:15.540)
the way the questions are asked is nonlinear.
Lex Fridman (53:18.340)
It's like, it's a little bit witty.
David Ferrucci (53:20.660)
It's a little bit playful sometimes.
Lex Fridman (53:22.460)
It's a little bit tricky.
David Ferrucci (53:25.940)
Yeah, they're asked in exactly numerous witty, tricky ways.
Lex Fridman (53:30.580)
Exactly what they're asking is not obvious.
David Ferrucci (53:32.540)
It takes inexperienced humans a while
Lex Fridman (53:35.060)
to go, what is it even asking?
Lex Fridman (53:36.900)
And it's sort of an interesting realization that you have
Lex Fridman (53:39.620)
when somebody says, oh, what's,
David Ferrucci (53:40.980)
Jeopardy is a question answering show.
Lex Fridman (53:42.420)
And then he's like, oh, like, I know a lot.
Lex Fridman (53:43.860)
And then you read it and you're still trying
Lex Fridman (53:45.980)
to process the question and the champions have answered
Lex Fridman (53:48.300)
and moved on.
Lex Fridman (53:49.140)
There are three questions ahead
David Ferrucci (53:51.180)
by the time you figured out what the question even meant.
Lex Fridman (53:54.060)
So there's definitely an ability there
David Ferrucci (53:56.460)
to just parse out what the question even is.
Lex Fridman (53:59.500)
So that was certainly challenging.
David Ferrucci (54:00.820)
It's interesting historically though,
Lex Fridman (54:02.220)
if you look back at the Jeopardy games much earlier,
David Ferrucci (54:05.460)
you know, early games. Like 60s, 70s, that kind of thing.
Lex Fridman (54:08.140)
The questions were much more direct.
David Ferrucci (54:10.180)
They weren't quite like that.
Lex Fridman (54:11.300)
They got sort of more and more interesting,
David Ferrucci (54:13.660)
the way they asked them that sort of got more
Lex Fridman (54:15.340)
and more interesting and subtle and nuanced
Lex Fridman (54:18.340)
and humorous and witty over time,
Lex Fridman (54:20.780)
which really required the human
David Ferrucci (54:22.500)
to kind of make the right connections
Lex Fridman (54:24.260)
in figuring out what the question was even asking.
Lex Fridman (54:26.860)
So yeah, you have to figure out the questions even asking.
Lex Fridman (54:29.940)
Then you have to determine whether
David Ferrucci (54:31.700)
or not you think you know the answer.
Lex Fridman (54:34.500)
And because you have to buzz in really quickly,
David Ferrucci (54:37.380)
you sort of have to make that determination
Lex Fridman (54:39.820)
as quickly as you possibly can.
David Ferrucci (54:41.220)
Otherwise you lose the opportunity to buzz in.
Lex Fridman (54:43.460)
You mean...
David Ferrucci (54:44.300)
Even before you really know if you know the answer.
Lex Fridman (54:46.140)
I think a lot of humans will assume,
David Ferrucci (54:48.660)
they'll process it very superficially.
Lex Fridman (54:53.020)
In other words, what's the topic?
Lex Fridman (54:54.940)
What are some keywords?
Lex Fridman (54:55.980)
And just say, do I know this area or not
Lex Fridman (54:58.660)
before they actually know the answer?
Lex Fridman (55:00.820)
Then they'll buzz in and think about it.
Lex Fridman (55:03.220)
So it's interesting what humans do.
Lex Fridman (55:04.700)
Now, some people who know all things,
David Ferrucci (55:06.940)
like Ken Jennings or something,
Lex Fridman (55:08.460)
or the more recent big Jeopardy player,
David Ferrucci (55:11.460)
I mean, they'll just buzz in.
Lex Fridman (55:12.420)
They'll just assume they know all of Jeopardy
Lex Fridman (55:14.100)
and they'll just buzz in.
Lex Fridman (55:15.900)
Watson, interestingly, didn't even come close
Lex Fridman (55:18.380)
to knowing all of Jeopardy, right?
Lex Fridman (55:20.140)
Watson really...
David Ferrucci (55:20.980)
Even at the peak, even at its best.
Lex Fridman (55:22.700)
Yeah, so for example, I mean,
David Ferrucci (55:24.580)
we had this thing called recall,
Lex Fridman (55:25.980)
which is like how many of all the Jeopardy questions,
Lex Fridman (55:29.420)
how many could we even find the right answer for anywhere?
Lex Fridman (55:34.420)
Like, can we come up with, we had a big body of knowledge,
David Ferrucci (55:38.220)
something in the order of several terabytes.
Lex Fridman (55:39.780)
I mean, from a web scale, it was actually very small,
Lex Fridman (55:42.900)
but from like a book scale,
Lex Fridman (55:44.340)
we're talking about millions of books, right?
Lex Fridman (55:46.260)
So the equivalent of millions of books,
Lex Fridman (55:48.260)
encyclopedias, dictionaries, books,
David Ferrucci (55:50.340)
it's still a ton of information.
Lex Fridman (55:52.260)
And I think it was only 85% was the answer
David Ferrucci (55:55.820)
anywhere to be found.
Lex Fridman (55:57.580)
So you're already down at that level
Lex Fridman (56:00.340)
just to get started, right?
Lex Fridman (56:02.060)
So, and so it was important to get a very quick sense
Lex Fridman (56:07.900)
of do you think you know the right answer to this question?
Lex Fridman (56:10.060)
So we had to compute that confidence
David Ferrucci (56:12.180)
as quickly as we possibly could.
Lex Fridman (56:14.300)
So in effect, we had to answer it
Lex Fridman (56:16.460)
and at least spend some time essentially answering it
Lex Fridman (56:22.020)
and then judging the confidence that our answer was right
Lex Fridman (56:26.660)
and then deciding whether or not
Lex Fridman (56:28.060)
we were confident enough to buzz in.
Lex Fridman (56:30.020)
And that would depend on what else was going on in the game.
Lex Fridman (56:31.940)
Because there was a risk.
Lex Fridman (56:33.380)
So like if you're really in a situation
Lex Fridman (56:35.060)
where I have to take a guess, I have very little to lose,
David Ferrucci (56:38.340)
then you'll buzz in with less confidence.
Lex Fridman (56:40.220)
So that was accounted for the financial standings
David Ferrucci (56:42.940)
of the different competitors.
Lex Fridman (56:44.300)
Correct.
Lex Fridman (56:45.420)
How much of the game was left?
Lex Fridman (56:46.620)
How much time was left?
David Ferrucci (56:48.260)
Where you were in the standing, things like that.
Lex Fridman (56:50.740)
How many hundreds of milliseconds
Lex Fridman (56:52.860)
that we're talking about here?
Lex Fridman (56:53.900)
Do you have a sense of what is?
David Ferrucci (56:55.980)
We targeted, yeah, we targeted.
Lex Fridman (56:58.420)
So, I mean, we targeted answering
David Ferrucci (57:01.180)
in under three seconds and.
Lex Fridman (57:04.660)
Buzzing in.
Lex Fridman (57:05.500)
So the decision to buzz in and then the actual answering
Lex Fridman (57:09.940)
are those two different stages?
David Ferrucci (57:10.980)
Yeah, they were two different things.
Lex Fridman (57:12.660)
In fact, we had multiple stages,
David Ferrucci (57:14.540)
whereas like we would say, let's estimate our confidence,
Lex Fridman (57:17.380)
which was sort of a shallow answering process.
Lex Fridman (57:21.060)
And then ultimately decide to buzz in
Lex Fridman (57:23.820)
and then we may take another second or something
David Ferrucci (57:27.420)
to kind of go in there and do that.
Lex Fridman (57:30.900)
But by and large, we were saying like,
David Ferrucci (57:32.180)
we can't play the game.
Lex Fridman (57:33.940)
We can't even compete if we can't on average
David Ferrucci (57:37.620)
answer these questions in around three seconds or less.
Lex Fridman (57:40.380)
So you stepped in.
Lex Fridman (57:41.740)
So there's these three humans playing a game
Lex Fridman (57:45.340)
and you stepped in with the idea that IBM Watson
David Ferrucci (57:47.980)
would be one of, replace one of the humans
Lex Fridman (57:49.980)
and compete against two.
Lex Fridman (57:52.020)
Can you tell the story of Watson taking on this game?
Lex Fridman (57:56.740)
Sure.
David Ferrucci (57:57.580)
It seems exceptionally difficult.
Lex Fridman (57:58.700)
Yeah, so the story was that it was coming up,
David Ferrucci (58:03.500)
I think to the 10 year anniversary of Big Blue,
Lex Fridman (58:06.940)
not Big Blue, Deep Blue.
David Ferrucci (58:08.780)
IBM wanted to do sort of another kind of really
Lex Fridman (58:11.940)
fun challenge, public challenge that can bring attention
David Ferrucci (58:15.260)
to IBM research and the kind of the cool stuff
Lex Fridman (58:17.180)
that we were doing.
David Ferrucci (58:19.740)
I had been working in AI at IBM for some time.
Lex Fridman (58:23.740)
I had a team doing what's called
David Ferrucci (58:26.460)
open domain factoid question answering,
Lex Fridman (58:28.620)
which is, we're not gonna tell you what the questions are.
David Ferrucci (58:31.020)
We're not even gonna tell you what they're about.
Lex Fridman (58:33.100)
Can you go off and get accurate answers to these questions?
Lex Fridman (58:36.860)
And it was an area of AI research that I was involved in.
Lex Fridman (58:41.420)
And so it was a very specific passion of mine.
David Ferrucci (58:44.300)
Language understanding had always been a passion of mine.
Lex Fridman (58:47.100)
One sort of narrow slice on whether or not
David Ferrucci (58:49.660)
you could do anything with language
Lex Fridman (58:51.020)
was this notion of open domain and meaning
David Ferrucci (58:52.900)
I could ask anything about anything.
Lex Fridman (58:54.620)
Factoid meaning it essentially had an answer
Lex Fridman (58:57.900)
and being able to do that accurately and quickly.
Lex Fridman (59:00.940)
So that was a research area
David Ferrucci (59:02.420)
that my team had already been in.
Lex Fridman (59:03.980)
And so completely independently,
Lex Fridman (59:06.340)
several IBM executives, like what are we gonna do?
Lex Fridman (59:09.060)
What's the next cool thing to do?
Lex Fridman (59:11.060)
And Ken Jennings was on his winning streak.
Lex Fridman (59:13.900)
This was like, whatever it was, 2004, I think,
David Ferrucci (59:16.660)
was on his winning streak.
Lex Fridman (59:18.780)
And someone thought, hey, that would be really cool
David Ferrucci (59:20.900)
if the computer can play Jeopardy.
Lex Fridman (59:23.940)
And so this was like in 2004,
David Ferrucci (59:25.740)
they were shopping this thing around
Lex Fridman (59:28.020)
and everyone was telling the research execs, no way.
David Ferrucci (59:33.540)
Like, this is crazy.
Lex Fridman (59:35.180)
And we had some pretty senior people in the field
Lex Fridman (59:37.020)
and they're saying, no, this is crazy.
Lex Fridman (59:38.180)
And it would come across my desk and I was like,
Lex Fridman (59:40.180)
but that's kind of what I'm really interested in doing.
Lex Fridman (59:44.700)
But there was such this prevailing sense of this is nuts.
David Ferrucci (59:47.460)
We're not gonna risk IBM's reputation on this.
Lex Fridman (59:49.380)
We're just not doing it.
Lex Fridman (59:50.500)
And this happened in 2004, it happened in 2005.
Lex Fridman (59:53.140)
At the end of 2006, it was coming around again.
Lex Fridman (59:59.260)
And I was coming off of a,
🔗 相关节目