WEBVTT

00:00:00.000 --> 00:00:03.357
And so places where I think that's about
to happen and it hasn't happened yet.

00:00:03.440 --> 00:00:05.597
Um, one of them is interpretability. I

00:00:05.680 --> 00:00:07.837
think interpretability is both the key

00:00:07.920 --> 00:00:11.837
to steering and making safe AI systems

00:00:11.920 --> 00:00:13.677
um and and we're about to understand and

00:00:13.760 --> 00:00:17.117
interpretability contains insights about

00:00:17.200 --> 00:00:21.837
intelligent optimization problems
and about how the human brain works. Um,

00:00:21.920 --> 00:00:24.077
I've said and I I'm really not joking.

00:00:24.160 --> 00:00:26.317
Chris Ola is going to be a future Nobel

00:00:26.400 --> 00:00:28.797
Medicine Laurette. I'm serious. I'm

00:00:28.880 --> 00:00:33.597
serious because a lot of these I used
to be a neuroscientist. A lot of these

00:00:33.680 --> 00:00:37.757
mental illnesses, the ones we haven't
figured out, right? Schizophrenia or the

00:00:37.840 --> 00:00:42.557
mood disorders, I suspect there's some
higher level system thing going on and

00:00:42.640 --> 00:00:46.797
that it's hard to make sense of those
with brains because brains are so mushy

00:00:46.880 --> 00:00:50.477
and hard to hard to open up and interact
with. Neural nets are not like this.

00:00:50.560 --> 00:00:53.277
They're not a perfect analogy,
but as time goes on, they will be a better

00:00:53.360 --> 00:00:56.477
analogy. In AI, it's really something
where like in three three to six months,

00:00:56.560 --> 00:00:59.757
you can really learn a lot about how
these systems work. It's not not so

00:00:59.840 --> 00:01:04.717
complex. And I think uh I think it's
probably worth learning if you think

00:01:04.800 --> 00:01:07.277
it's plausible that these systems are
going to become that powerful because I

00:01:07.360 --> 00:01:09.437
think we need our best and brightest to

00:01:09.520 --> 00:01:11.437
kind of I don't know have their game

00:01:11.520 --> 00:01:14.637
face on or something for uh the

00:01:14.720 --> 00:01:18.797
challenges and the opportunities that
that the technology might bring. where I

00:01:18.880 --> 00:01:20.957
had like my personal runway. I I ran out

00:01:21.040 --> 00:01:23.277
and so I was like, "Okay, I think that

00:01:23.360 --> 00:01:26.237
I'm going to need six months of stealth

00:01:26.320 --> 00:01:28.717
study to have a shot at getting a job."

00:01:28.800 --> 00:01:32.557
At that point, it was Deep Mind
or Google Brain were the two places to do

00:01:32.640 --> 00:01:36.317
work there. Or Merie Merie was the third
one that I was like looking at. So, I

00:01:36.400 --> 00:01:39.837
was like, if I want to help out with that,
those are the three places to look

00:01:39.920 --> 00:01:41.597
at. I don't have any of the skills yet.

00:01:41.680 --> 00:01:45.357
I need six months of self-study to feel
like I would not be a drag on them and

00:01:45.440 --> 00:01:50.317
like actually be helping instead. Can
you um maybe explain a bit what was that

00:01:50.400 --> 00:01:53.197
cell study like because I'm sure there's
a lot of software engineers right now in

00:01:53.280 --> 00:01:58.237
their 20s are looking tool to become
AI researchers. What was what was that 6

00:01:58.320 --> 00:02:02.397
months like even though as you said
you had uh gotten a B minus in linear

00:02:02.480 --> 00:02:07.917
algebra which is like core >> might have
been C plus I'm checking going to give

00:02:08.000 --> 00:02:09.437
>> that's pretty impressive
where where you got to.

00:02:09.520 --> 00:02:13.037
>> Yeah. Yeah. It turned out okay. First
I did a contract actually with Twitch um

00:02:13.120 --> 00:02:17.117
and like earned like enough to have
that six months of runway. So I did like

00:02:17.200 --> 00:02:21.357
three month contract with Twitch
and then I made a a plan to self-study. I

00:02:21.440 --> 00:02:24.717
don't think it's the right plan now
for people too because at least the 2015.

00:02:24.800 --> 00:02:25.997
What did it look like? It was like take

00:02:26.080 --> 00:02:28.797
a Corsera course on machine learning,

00:02:28.880 --> 00:02:32.557
try to solve some Kaggle projects, read

00:02:32.640 --> 00:02:35.997
linear algebra done right and uh I had a

00:02:36.080 --> 00:02:38.637
statistics textbook. I think I had YC

00:02:38.720 --> 00:02:41.837
alumni credits and so I bought like a

00:02:41.920 --> 00:02:45.917
GPU and I would like SSH into the GPU to

00:02:46.000 --> 00:02:47.597
like work through my courses for it.

00:02:47.680 --> 00:02:51.677
>> And this is right after Yeah,
it was already after Alex Knack, right?

00:02:51.760 --> 00:02:55.357
>> This after Alex Net. Yeah. So I was
mostly doing image image classification

00:02:55.440 --> 00:02:58.717
stuff that I was trying to learn
was like the thing that all the courses

00:02:58.800 --> 00:03:01.917
would teach you to do. >> How did
you get the open AI job? Because

00:03:02.000 --> 00:03:05.437
you were one of the few engineers.
it was mostly researchers and they had

00:03:05.520 --> 00:03:07.277
pretty stacked team of researchers.

00:03:07.360 --> 00:03:10.477
>> I messaged Greg um
as soon as OpenAI was

00:03:10.560 --> 00:03:14.477
announced and I was like, I'd love to help
out in some way. I got a B minus in

00:03:14.560 --> 00:03:18.077
my linear algebra, but I know some
engineering. I've done a bit of

00:03:18.160 --> 00:03:23.920
distributed systems work. If you guys need
help, I'm like happy to mop floors if if