WEBVTT

00:00:00.000 --> 00:00:01.757
foreshadowed a couple times already that

00:00:01.840 --> 00:00:04.597
you have some good models on the way.

00:00:04.680 --> 00:00:08.317
What is Spud? Uh the information said uh

00:00:08.400 --> 00:00:11.677
that you finished uh pre-training Spud,

00:00:11.760 --> 00:00:16.757
and Sam Altman, the CEO of OpenAI,
has told the staff that uh they should

00:00:16.840 --> 00:00:20.637
expect to have a very strong model
in a few weeks. This was a few weeks ago.

00:00:20.720 --> 00:00:24.117
And the team believes it can really
accelerate the economy, and things are

00:00:24.200 --> 00:00:26.757
moving faster than many of us expect

00:00:26.840 --> 00:00:29.677
expected. So, what's Spud?

00:00:29.760 --> 00:00:33.957
It's a good model. Um but I think
that it's really not about any one model.

00:00:34.040 --> 00:00:37.957
Okay, right? The the the way that our
development process works is you have

00:00:38.040 --> 00:00:42.077
pre-training, so you produce a new base
model that then is the foundation that

00:00:42.160 --> 00:00:45.757
we build further improvements
on top of. And that

00:00:45.840 --> 00:00:49.477
is always a huge effort across many
people in the company, and that's where

00:00:49.560 --> 00:00:52.797
I've actually been spending most of my
efforts over the past 18 months has been

00:00:52.880 --> 00:00:56.837
really focused on our GPU
infrastructure, on supporting the

00:00:56.920 --> 00:00:59.997
teams that do all of the training
frameworks to scale up at these big

00:01:00.080 --> 00:01:04.077
runs. But then there's a reinforcement
learning process. So, you take this AI

00:01:04.160 --> 00:01:08.077
that has learned lots of things
about the world, and it applies that

00:01:08.160 --> 00:01:10.397
knowledge. And then we do
a post-training process where you really

00:01:10.480 --> 00:01:13.717
say, "Okay, now you know how to solve
problems. You practice it in in all

00:01:13.800 --> 00:01:16.637
these different contexts." And then
here's kind of the last mile of of

00:01:16.720 --> 00:01:18.557
behavior and usability. So, I think of

00:01:18.640 --> 00:01:22.037
Spud as a new base, as a new pre-train,

00:01:22.120 --> 00:01:25.917
and that we have had this uh

00:01:26.000 --> 00:01:30.117
I'd say it's like we have maybe 2 years
worth of research that is coming to

00:01:30.200 --> 00:01:33.717
fruition in this model.
It's going to be very exciting.

00:01:33.800 --> 00:01:37.357
And I think that the way that the world
will experience it is just improved

00:01:37.440 --> 00:01:41.397
capabilities, and that for me
it's never about any one release,

00:01:41.480 --> 00:01:45.797
because as soon as we have this one
release, it'll be an early version of

00:01:45.880 --> 00:01:49.117
what we have coming. We'll do much
more of each of these steps of the

00:01:49.200 --> 00:01:53.997
improvement process. And so, I think
that where we're going is is almost just

00:01:54.080 --> 00:01:58.237
we have this engine of progress that
just moves faster and faster, and that

00:01:58.320 --> 00:02:01.237
Spud is just one step along the way.
So, what do you think it'll be able to do

00:02:01.320 --> 00:02:04.357
that today's models can't? I So, I think

00:02:04.440 --> 00:02:06.237
it's going to be able to solve both much

00:02:06.320 --> 00:02:11.437
harder problems. I think it will be
much more nuanced. It'll understand

00:02:11.520 --> 00:02:15.077
instructions better. It'll understand
the context much better. That there's

00:02:15.160 --> 00:02:18.317
this thing called big model smell that
people talk about, where it's just like

00:02:18.400 --> 00:02:19.757
there's something about like when these

00:02:19.840 --> 00:02:22.157
models are just actually just much

00:02:22.240 --> 00:02:25.837
smarter, much more capable, that they

00:02:25.920 --> 00:02:28.357
bend to you much more, and you feel it.

00:02:28.440 --> 00:02:31.637
Right? When you ask a question and the AI
doesn't quite get it, it's always so

00:02:31.720 --> 00:02:34.437
disappointing. Right? When you have
to like explain it, you're like, "You

00:02:34.520 --> 00:02:37.517
really should be able to figure this out."
And so, I would just think of it

00:02:37.600 --> 00:02:42.517
as in some ways just
qualitatively there will be

00:02:42.600 --> 00:02:45.757
like quantitatively lots of shifts,
right? And qualitatively there will just

00:02:45.840 --> 00:02:48.757
be new things where you would be
frustrated before you never use an AI

00:02:48.840 --> 00:02:51.917
for it, and now you just use it without
without thinking very much. And I think

00:02:52.000 --> 00:02:55.557
that that that is what we're going
to see across the board. I'm super excited

00:02:55.640 --> 00:02:57.397
to see how it raises the ceiling, right?

00:02:57.480 --> 00:03:00.077
We've already seen these physics
applications, things like that. And I

00:03:00.160 --> 00:03:04.357
think we will be able to just solve like
way more open-ended problems, way longer

00:03:04.440 --> 00:03:07.997
time horizons. And then also
very excited to see how it raises

00:03:08.080 --> 00:03:11.557
the floor, where just for anything you
want to do, it's just so much useful

00:03:11.640 --> 00:03:14.880
more useful for you.