WEBVTT

00:00:00.000 --> 00:00:02.677
How specifically does being open-source

00:00:02.760 --> 00:00:04.517
not lead us to our end?

00:00:04.600 --> 00:00:09.737
>> Dr. Yann LeCun, Turing Award Laureate,
Vice President and Chief AI Scientist

00:00:09.820 --> 00:00:13.037
[music] of Meta. >> We need
free and diverse AI platforms

00:00:13.120 --> 00:00:15.917
for the same reason that we need free
and diverse press, so we don't have a

00:00:16.000 --> 00:00:17.957
single opinion coming to us from AI.

00:00:18.040 --> 00:00:21.997
We're still very far from matching
human and animal [music] intelligence and

00:00:22.080 --> 00:00:23.957
learning capacities with AI systems.

00:00:24.040 --> 00:00:29.160
It's not just around the corner. It's
not going to happen in the next 2 years.

00:00:30.520 --> 00:00:32.837
>> Can you talk to us, Yann, about

00:00:32.920 --> 00:00:37.877
open-source AI? Why is open-source
the way forward from a security and

00:00:37.960 --> 00:00:40.997
innovation perspective? >> So,

00:00:41.080 --> 00:00:44.957
there's going to be a a future,
not very far in the future,

00:00:45.040 --> 00:00:48.517
where all of our interaction
with the digital

00:00:48.600 --> 00:00:52.317
world are going to be
mediated by AI systems.

00:00:52.400 --> 00:00:55.197
We already see this with uh

00:00:55.280 --> 00:00:57.277
uh companies putting out

00:00:57.360 --> 00:00:59.797
basically replacement for search engines

00:00:59.880 --> 00:01:03.077
uh that are backed by AI systems.

00:01:03.160 --> 00:01:07.557
And so, in this in this future,
if all of our digital diet is mediated by

00:01:07.640 --> 00:01:09.317
AI systems,

00:01:09.400 --> 00:01:11.877
um it's going to be there's going to be

00:01:11.960 --> 00:01:15.157
a need for a a a very wide diversity of

00:01:15.240 --> 00:01:19.077
AI systems that cater to different
languages, different cultures, different

00:01:19.160 --> 00:01:22.397
value systems,
different centers of interest.

00:01:22.480 --> 00:01:26.557
And it it can't be produced by a small
number of companies on the West Coast of

00:01:26.640 --> 00:01:27.837
the US.

00:01:27.920 --> 00:01:31.797
Um um or any single country.

00:01:31.880 --> 00:01:34.677
Uh so, essentially AI

00:01:34.760 --> 00:01:39.677
tools, assistants, will become kind
of a repository of all human knowledge.

00:01:39.760 --> 00:01:43.357
And um it's going to be a a kind of

00:01:43.440 --> 00:01:46.837
infrastructure. And the history shows that

00:01:46.920 --> 00:01:50.637
infrastructure software always
end ends up being open-source

00:01:50.720 --> 00:01:54.797
because it's more secure,
it's more customizable.

00:01:54.880 --> 00:01:58.317
Uh it enables the creation
of an ecosystem on top of it.

00:01:58.400 --> 00:02:02.877
And so the the entire infrastructure of
our communication systems, the internet,

00:02:02.960 --> 00:02:06.837
um even our cell phones, uh

00:02:06.920 --> 00:02:10.357
software stack for the, you know,
communication communicating with cell

00:02:10.440 --> 00:02:12.517
phone towers, it's all open source.

00:02:12.600 --> 00:02:15.877
For very good reasons. It's not
by design, it's really the the market

00:02:15.960 --> 00:02:19.757
forces that push it.
So there's a similar phenomenon which

00:02:19.840 --> 00:02:22.757
is definitely happening now in the last uh

00:02:22.840 --> 00:02:25.677
uh few months that Llama 2 has uh has

00:02:25.760 --> 00:02:29.117
been available and more open source uh

00:02:29.200 --> 00:02:32.757
uh and then it's been available
that people are using those open source

00:02:32.840 --> 00:02:37.037
models and and customizing them for
vertical applications in businesses, but

00:02:37.120 --> 00:02:41.677
also uh training them for,
you know, a particular types of uh

00:02:41.760 --> 00:02:42.317
uh

00:02:42.400 --> 00:02:45.237
languages or or cultures. Uh and so it's

00:02:45.320 --> 00:02:49.157
a necessity. We we need it. We need

00:02:49.240 --> 00:02:52.757
free and diverse AI platforms
for the same reason that we need free and

00:02:52.840 --> 00:02:55.797
diverse press. Um so we
don't have a single opinion

00:02:55.880 --> 00:02:58.197
coming to us from from AI. Getting

00:02:58.280 --> 00:03:00.717
towards the the title, will AI lead us

00:03:00.800 --> 00:03:05.437
to our end? How specifically
does being open source

00:03:05.520 --> 00:03:09.317
not lead us to our end? Is there
something fundamental that makes it

00:03:09.400 --> 00:03:12.077
easier to secure?

00:03:12.160 --> 00:03:16.397
Yes. Uh so it's it's certainly been
the case historically that uh open source

00:03:16.480 --> 00:03:20.237
software has been more
more secure simply because

00:03:20.320 --> 00:03:25.437
there's more eyeballs on it, right?
So the entire world today runs on Linux.

00:03:25.520 --> 00:03:28.117
Um maybe not your desktop,

00:03:28.200 --> 00:03:31.077
uh but probably more more than half of

00:03:31.160 --> 00:03:35.317
your phones run on Linux.
Whatever is not an iPhone runs on Linux.

00:03:35.400 --> 00:03:39.437
Uh the web servers around the world run
on Linux. All the cloud services run on

00:03:39.520 --> 00:03:43.357
Linux. Your car has several
microcontrollers that probably run

00:03:43.440 --> 00:03:46.557
Linux. The uh helicopter
on Mars runs Linux. I

00:03:46.640 --> 00:03:50.397
mean, Linux runs the world.
And the reason is because

00:03:50.480 --> 00:03:53.637
it's open source, so you can port it
to anything you want, and you can make it

00:03:53.720 --> 00:03:55.797
more secure. And and so, you know,

00:03:55.880 --> 00:03:58.397
there's kind of a similar story uh to be

00:03:58.480 --> 00:04:02.477
uh to be made for uh for for AI uh

00:04:02.560 --> 00:04:04.597
business models. More eyeballs, more

00:04:04.680 --> 00:04:11.080
ways to uh make it safe, um uh etc. Now,

00:04:11.440 --> 00:04:14.597
the only reason why you might think that

00:04:14.680 --> 00:04:18.837
it's not a good idea to open source
is if you have this belief that somehow AI

00:04:18.920 --> 00:04:22.357
is so powerful that it
is going to take over the world

00:04:22.440 --> 00:04:26.357
at some point. Okay. And uh

00:04:26.440 --> 00:04:29.757
and this is this is not happening.

00:04:29.840 --> 00:04:33.877
Um so, first of all, we're still very far

00:04:33.960 --> 00:04:37.077
despite all the wonderful
things that AI can do today,

00:04:37.160 --> 00:04:39.237
uh we're still very far from matching

00:04:39.320 --> 00:04:44.517
human and animal intelligence
and learning capacities with AI systems.

00:04:44.600 --> 00:04:47.397
We're very far. It's not
just around the corner. It's

00:04:47.480 --> 00:04:49.597
not going to happen in the next 2 years.

00:04:49.680 --> 00:04:52.397
Um we still some need some scientific

00:04:52.480 --> 00:04:55.117
breakthroughs to get to the type of

00:04:55.200 --> 00:04:57.637
learning abilities that we observe in uh

00:04:57.720 --> 00:05:00.597
not just humans, but even your your cat.

00:05:00.680 --> 00:05:03.957
Um your cat has a much better
understanding of the physical world than

00:05:04.040 --> 00:05:07.237
uh the most powerful
AI systems that we have.

00:05:07.320 --> 00:05:12.797
>> I've seen my dog run into the glass
door in my house. Y- Y- I I I I I debate

00:05:12.880 --> 00:05:15.077
that, but yeah. >> Sure,
but you know, I've run into glass

00:05:15.160 --> 00:05:19.917
doors. Um so, the the uh

00:05:20.000 --> 00:05:23.477
H- Here is an interesting uh statistic. Um

00:05:23.560 --> 00:05:27.277
so, first of all, before we get
to human-level AI, uh we're going to need

00:05:27.360 --> 00:05:30.717
some breakthroughs that allow machines
to understand how the world works, which

00:05:30.800 --> 00:05:35.437
they can't do at the moment,
uh can remember facts, which they can't

00:05:35.520 --> 00:05:38.677
really do at the moment. They don't
really have persistent memory, at least

00:05:38.760 --> 00:05:40.077
the LLMs don't.

00:05:40.160 --> 00:05:43.237
Um they can reason, and they can plan.

00:05:43.320 --> 00:05:48.277
And those are characteristics that are
essential to intelligent behavior.

00:05:48.360 --> 00:05:52.157
Current large language models cannot do
any of those things, at least not to the

00:05:52.240 --> 00:05:56.397
level that uh we'd like them to to do.

00:05:56.480 --> 00:05:58.557
And the reason is is relatively simple.

00:05:58.640 --> 00:06:00.877
Um if you

00:06:00.960 --> 00:06:05.637
if you look at how much data is used
to train those LLMs, it sounds staggering.

00:06:05.720 --> 00:06:09.557
It's uh typically 10 trillion
tokens, right? A token is

00:06:09.640 --> 00:06:14.637
like a word, so it's 10 to the 13,
right? A one with 13 zeros. Uh

00:06:14.720 --> 00:06:18.477
words or word equivalents.

00:06:18.560 --> 00:06:22.277
Each word or each of those tokens is
stored in two bytes, so that's 2 * 10 to

00:06:22.360 --> 00:06:26.597
the 13 bytes. Seems staggering.
If you evaluate how

00:06:26.680 --> 00:06:29.677
long it would take for humans to read
this, it would take, you know, anywhere

00:06:29.760 --> 00:06:34.637
between 150,000 and 200,000 years
for a single person to read this reading 8

00:06:34.720 --> 00:06:38.597
hours a day. Uh enormous. So,

00:06:38.680 --> 00:06:42.157
the conclusion is those systems are
incredibly smart, right? Because they've

00:06:42.240 --> 00:06:47.840
they're capable of digesting
this enormous amount of knowledge.