episode
stringlengths
45
100
text
stringlengths
1
528
timestamp_link
stringlengths
56
56
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
It's like if you ask me, what does it
https://karpathy.ai/lexicap/0013-large.html#00:29:50.160
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
mean to understand a computer?
https://karpathy.ai/lexicap/0013-large.html#00:29:54.360
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
That's much simpler.
https://karpathy.ai/lexicap/0013-large.html#00:29:57.560
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
But in a computer, I could say, well,
https://karpathy.ai/lexicap/0013-large.html#00:29:58.640
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
I understand how to use PowerPoint.
https://karpathy.ai/lexicap/0013-large.html#00:30:01.080
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
That's my level of understanding a computer.
https://karpathy.ai/lexicap/0013-large.html#00:30:04.800
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
It is reasonable.
https://karpathy.ai/lexicap/0013-large.html#00:30:08.080
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
It gives me some power to produce slides
https://karpathy.ai/lexicap/0013-large.html#00:30:09.400
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
and beautiful slides.
https://karpathy.ai/lexicap/0013-large.html#00:30:11.760
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
Now, you can ask somebody else.
https://karpathy.ai/lexicap/0013-large.html#00:30:14.480
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
He says, well, I know how the transistors work
https://karpathy.ai/lexicap/0013-large.html#00:30:17.320
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
that are inside the computer.
https://karpathy.ai/lexicap/0013-large.html#00:30:19.840
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
I can write the equation for transistor and diodes
https://karpathy.ai/lexicap/0013-large.html#00:30:21.360
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
and circuits, logical circuits.
https://karpathy.ai/lexicap/0013-large.html#00:30:25.920
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
And I can ask this guy, do you know how to operate PowerPoint?
https://karpathy.ai/lexicap/0013-large.html#00:30:29.360
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
No idea.
https://karpathy.ai/lexicap/0013-large.html#00:30:32.440
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
So do you think if we discovered computers walking amongst us
https://karpathy.ai/lexicap/0013-large.html#00:30:34.040
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
full of these transistors that are also operating
https://karpathy.ai/lexicap/0013-large.html#00:30:39.800
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
under windows and have PowerPoint,
https://karpathy.ai/lexicap/0013-large.html#00:30:43.400
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
do you think it's digging in a little bit more?
https://karpathy.ai/lexicap/0013-large.html#00:30:45.560
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
How useful is it to understand the transistor in order
https://karpathy.ai/lexicap/0013-large.html#00:30:49.960
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
to be able to understand PowerPoint
https://karpathy.ai/lexicap/0013-large.html#00:30:53.280
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
and these higher level intelligent processes?
https://karpathy.ai/lexicap/0013-large.html#00:30:58.040
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
So I think in the case of computers,
https://karpathy.ai/lexicap/0013-large.html#00:31:00.320
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
because they were made by engineers, by us,
https://karpathy.ai/lexicap/0013-large.html#00:31:03.720
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
this different level of understanding
https://karpathy.ai/lexicap/0013-large.html#00:31:06.960
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
are rather separate on purpose.
https://karpathy.ai/lexicap/0013-large.html#00:31:09.280
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
They are separate modules so that the engineer that
https://karpathy.ai/lexicap/0013-large.html#00:31:13.280
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
designed the circuit for the chips does not
https://karpathy.ai/lexicap/0013-large.html#00:31:17.240
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
need to know what is inside PowerPoint.
https://karpathy.ai/lexicap/0013-large.html#00:31:19.640
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
And somebody can write the software translating
https://karpathy.ai/lexicap/0013-large.html#00:31:23.600
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
from one to the other.
https://karpathy.ai/lexicap/0013-large.html#00:31:27.440
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
So in that case, I don't think understanding the transistor
https://karpathy.ai/lexicap/0013-large.html#00:31:30.360
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
helps you understand PowerPoint, or very little.
https://karpathy.ai/lexicap/0013-large.html#00:31:36.960
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
If you want to understand the computer, this question,
https://karpathy.ai/lexicap/0013-large.html#00:31:41.120
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
I would say you have to understand it
https://karpathy.ai/lexicap/0013-large.html#00:31:43.960
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
at different levels.
https://karpathy.ai/lexicap/0013-large.html#00:31:45.960
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
If you really want to build one, right?
https://karpathy.ai/lexicap/0013-large.html#00:31:46.800
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
But for the brain, I think these levels of understanding,
https://karpathy.ai/lexicap/0013-large.html#00:31:51.520
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
so the algorithms, which kind of computation,
https://karpathy.ai/lexicap/0013-large.html#00:31:57.320
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
the equivalent of PowerPoint, and the circuits,
https://karpathy.ai/lexicap/0013-large.html#00:32:00.840
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
the transistors, I think they are much more
https://karpathy.ai/lexicap/0013-large.html#00:32:04.640
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
intertwined with each other.
https://karpathy.ai/lexicap/0013-large.html#00:32:07.560
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
There is not a neatly level of the software separate
https://karpathy.ai/lexicap/0013-large.html#00:32:09.560
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
from the hardware.
https://karpathy.ai/lexicap/0013-large.html#00:32:14.480
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
And so that's why I think in the case of the brain,
https://karpathy.ai/lexicap/0013-large.html#00:32:15.840
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
the problem is more difficult and more than for computers
https://karpathy.ai/lexicap/0013-large.html#00:32:20.080
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
requires the interaction, the collaboration
https://karpathy.ai/lexicap/0013-large.html#00:32:23.640
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
between different types of expertise.
https://karpathy.ai/lexicap/0013-large.html#00:32:26.560
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
The brain is a big hierarchical mess.
https://karpathy.ai/lexicap/0013-large.html#00:32:30.080
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
You can't just disentangle levels.
https://karpathy.ai/lexicap/0013-large.html#00:32:32.320
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
I think you can, but it's much more difficult.
https://karpathy.ai/lexicap/0013-large.html#00:32:35.120
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
And it's not completely obvious.
https://karpathy.ai/lexicap/0013-large.html#00:32:37.880
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
And as I said, I think it's one of the, personally,
https://karpathy.ai/lexicap/0013-large.html#00:32:40.840
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
I think is the greatest problem in science.
https://karpathy.ai/lexicap/0013-large.html#00:32:44.720
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
So I think it's fair that it's difficult.
https://karpathy.ai/lexicap/0013-large.html#00:32:47.240
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
That's a difficult one.
https://karpathy.ai/lexicap/0013-large.html#00:32:51.880
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
That said, you do talk about compositionality
https://karpathy.ai/lexicap/0013-large.html#00:32:53.320
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
and why it might be useful.
https://karpathy.ai/lexicap/0013-large.html#00:32:56.920
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
And when you discuss why these neural networks,
https://karpathy.ai/lexicap/0013-large.html#00:32:58.280
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
in artificial or biological sense, learn anything,
https://karpathy.ai/lexicap/0013-large.html#00:33:01.720
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
you talk about compositionality.
https://karpathy.ai/lexicap/0013-large.html#00:33:05.200
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
See, there's a sense that nature can be disentangled.
https://karpathy.ai/lexicap/0013-large.html#00:33:07.560
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
Or, well, all aspects of our cognition
https://karpathy.ai/lexicap/0013-large.html#00:33:13.480
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
could be disentangled to some degree.
https://karpathy.ai/lexicap/0013-large.html#00:33:19.840
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
So why do you think, first of all,
https://karpathy.ai/lexicap/0013-large.html#00:33:22.640
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
how do you see compositionality?
https://karpathy.ai/lexicap/0013-large.html#00:33:25.920
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
And why do you think it exists at all in nature?
https://karpathy.ai/lexicap/0013-large.html#00:33:27.720
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
I spoke about, I use the term compositionality
https://karpathy.ai/lexicap/0013-large.html#00:33:31.640
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
when we looked at deep neural networks, multilayers,
https://karpathy.ai/lexicap/0013-large.html#00:33:39.800
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
and trying to understand when and why they are more powerful
https://karpathy.ai/lexicap/0013-large.html#00:33:45.320
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
than more classical one layer networks,
https://karpathy.ai/lexicap/0013-large.html#00:33:50.560
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
like linear classifier, kernel machines, so called.
https://karpathy.ai/lexicap/0013-large.html#00:33:54.800
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
And what we found is that in terms
https://karpathy.ai/lexicap/0013-large.html#00:34:01.600
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
of approximating or learning or representing
https://karpathy.ai/lexicap/0013-large.html#00:34:05.360
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
a function, a mapping from an input to an output,
https://karpathy.ai/lexicap/0013-large.html#00:34:08.360
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
like from an image to the label in the image,
https://karpathy.ai/lexicap/0013-large.html#00:34:12.200
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
if this function has a particular structure,
https://karpathy.ai/lexicap/0013-large.html#00:34:16.760
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
then deep networks are much more powerful than shallow networks
https://karpathy.ai/lexicap/0013-large.html#00:34:20.840
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
to approximate the underlying function.
https://karpathy.ai/lexicap/0013-large.html#00:34:26.120
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
And the particular structure is a structure of compositionality.
https://karpathy.ai/lexicap/0013-large.html#00:34:28.880
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
If the function is made up of functions of function,
https://karpathy.ai/lexicap/0013-large.html#00:34:33.920
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
so that you need to look on when you are interpreting an image,
https://karpathy.ai/lexicap/0013-large.html#00:34:38.960
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
classifying an image, you don't need
https://karpathy.ai/lexicap/0013-large.html#00:34:45.800
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
to look at all pixels at once.
https://karpathy.ai/lexicap/0013-large.html#00:34:47.720
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
But you can compute something from small groups of pixels.
https://karpathy.ai/lexicap/0013-large.html#00:34:51.040
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
And then you can compute something
https://karpathy.ai/lexicap/0013-large.html#00:34:57.120
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
on the output of this local computation and so on,
https://karpathy.ai/lexicap/0013-large.html#00:34:59.920
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
which is similar to what you do when you read a sentence.
https://karpathy.ai/lexicap/0013-large.html#00:35:04.760
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
You don't need to read the first and the last letter.
https://karpathy.ai/lexicap/0013-large.html#00:35:07.320
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
But you can read syllables, combine them in words,
https://karpathy.ai/lexicap/0013-large.html#00:35:11.360
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
combine the words in sentences.
https://karpathy.ai/lexicap/0013-large.html#00:35:16.000
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
So this is this kind of structure.
https://karpathy.ai/lexicap/0013-large.html#00:35:18.120
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
So that's as part of a discussion
https://karpathy.ai/lexicap/0013-large.html#00:35:21.040
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
of why deep neural networks may be more
https://karpathy.ai/lexicap/0013-large.html#00:35:22.600
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
effective than the shallow methods.
https://karpathy.ai/lexicap/0013-large.html#00:35:26.120
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
And is your sense, for most things
https://karpathy.ai/lexicap/0013-large.html#00:35:27.880
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
we can use neural networks for, those problems
https://karpathy.ai/lexicap/0013-large.html#00:35:31.320
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
are going to be compositional in nature, like language,
https://karpathy.ai/lexicap/0013-large.html#00:35:37.400
Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13
like vision?
https://karpathy.ai/lexicap/0013-large.html#00:35:42.440