00:00:00
two things. One is how quickly starting
00:00:05
in March of 2023, so we're talking about
00:00:09
less than two and a half years, what
00:00:12
this team has accomplished
00:00:15
and how far ahead they are of everybody
00:00:18
else as demonstrated by this. But the
00:00:20
second is a fundamental architectural
00:00:23
decision that Elon made which I think we
00:00:27
didn't fully appreciate until now. And
00:00:30
it maps to an architectural decision he
00:00:33
made at Tesla as well. And for all we
00:00:35
know, we'll figure out that he made an
00:00:37
equivalent decision at SpaceX. And that
00:00:40
decision is really well encapsulated by
00:00:43
this essay, the bitter lesson by Rich
00:00:46
Sutton. And Nick, if you can just throw
00:00:48
this up here, but just to summarize what
00:00:50
this says, it basically says in a
00:00:52
nutshell
00:00:53
that you're always better off when
00:00:56
you're trying to solve an AI problem
00:00:59
taking a general learning approach that
00:01:01
can scale with computation because it
00:01:03
ultimately proves to be the most
00:01:05
effective. And the alternative would be
00:01:08
something that's much more human labored
00:01:10
and human involved that requires human
00:01:13
knowledge. And so the first method, what
00:01:16
it essentially allows you to do is view
00:01:20
any problem as an endless scalable
00:01:24
search or learning task. And as it's
00:01:27
turned out, whether it's chess or go or
00:01:31
speech recognition or computer vision,
00:01:34
whenever there was two competing
00:01:35
approaches, one that used general
00:01:37
computation and one that used human
00:01:39
knowledge, the general computation
00:01:42
problem always won. And so it creates
00:01:46
this bitter lesson for humans that want
00:01:48
to think that we are at the center of
00:01:50
all of this critical learning and all of
00:01:52
these leaps. And what these results show
00:01:55
is a general computational approach that
00:01:58
doesn't require as much human labeling
00:02:00
can actually get to the answer and
00:02:02
better answers faster. That has huge
00:02:05
implications because if you think about
00:02:07
all these other companies, what is Llama
00:02:09
been doing? They just spent 15 billion
00:02:11
to buy 49% of scale AI. That's exactly a
00:02:14
bet on human knowledge. What is Gemini
00:02:17
doing? What is OpenAI doing? What is
00:02:18
Anthropic doing? So all these things
00:02:20
come into question. I just think it's an
00:02:22
incredible moment in technology where we
00:02:25
see so many examples. Travis is another
00:02:27
one. What he's just talked about, you
00:02:29
know, the bitter lesson is you could
00:02:32
believe that, you know, food is this
00:02:34
immutable thing that's made meticulously
00:02:36
by hand by these individuals. Or you can
00:02:38
take this general purpose computer
00:02:40
approach, which is what he took, waited
00:02:42
for these cost curves to come into play,
00:02:44
and now you can scale food to every
00:02:46
human on Earth. I ju I just think it's a
00:02:48
it's so profoundly important.
