WEBVTT
00:01.620 --> 00:05.500
Okay.
So the first reason you might be convinced
00:05.580 --> 00:09.500
is coming in the next five years is
that a lot of existing people in the space
00:09.580 --> 00:13.120
believe this.
And I think some of them actually
00:13.160 --> 00:16.329
genuinely believe this. Like, this
is not entirely just hype or
00:16.360 --> 00:19.580
marketing.
00:26.720 --> 00:30.360
In between 5 and 20 years from now,
there's a good chance, a
00:30.500 --> 00:33.540
50% chance, we'll get AI smarter than us.
00:34.880 --> 00:38.760
Oh, yeah. Yes. In fact,
I have reduced my expectations for
00:38.900 --> 00:42.060
when we will get to AGI or whatever, human
00:42.120 --> 00:45.960
level. Uh,
I used to think it would be 5 to 20
00:45.980 --> 00:49.560
years, and now I think it could be,
you know, 2 to 10 years.
00:51.200 --> 00:55.080
And so as a result of that, I, I...
You can find it on my blog
00:55.300 --> 00:59.050
from about 2009, I think
is the first time I publicly talked about
00:59.080 --> 01:02.020
I predicted a 50% chance of, um, AGI by
01:02.160 --> 01:03.560
2028.
01:03.620 --> 01:03.790
How did-
01:03.800 --> 01:05.920
I still, I still believe that today.
01:07.380 --> 01:10.970
But still, if you,
if you extrapolate this kind of progress,
01:11.020 --> 01:12.220
very different place
01:13.140 --> 01:15.680
in 5 years, let alone 10 years.
01:15.690 --> 01:19.140
... has been added. If you just kind of,
like, eyeball the rate at which these
01:19.180 --> 01:23.160
capabilities are increasing,
it does make you think that we'll
01:23.200 --> 01:25.640
get there by 2026 or 2027.
01:25.700 --> 01:29.420
In the essay,
you actually say a really big
01:29.520 --> 01:33.410
thing, which is ASI, superintelligence, is
01:33.480 --> 01:36.960
actually thousands of days away.
01:37.020 --> 01:38.550
Maybe. I mean, that's our hope, our guess.
01:38.580 --> 01:39.680
Yeah.
01:41.960 --> 01:43.800
Like, I think,
I think we're quite close to digital
01:43.860 --> 01:46.520
superintelligence. It, it may
01:46.580 --> 01:50.500
happen this year, and
if it doesn't happen this year, next year
01:50.560 --> 01:51.660
for sure.
01:54.120 --> 01:57.230
The second reason you might be convinced,
uh,
01:58.200 --> 02:00.660
is just by trying the models yourself.
02:00.680 --> 02:03.940
Like, instead of trusting what, you know,
other people say about the various AI
02:03.980 --> 02:07.900
models, try the models yourself.
And don't just try, you know,
02:07.940 --> 02:09.979
one AI model. Try at least two or three.
02:10.000 --> 02:13.850
Try the model, you know, 2019,
try a model from 2022, try a
02:13.880 --> 02:17.260
model from 2025. And then once you have,
like, at least two or three data points
02:17.380 --> 02:21.160
from, you know,
three different time periods,
02:21.240 --> 02:24.440
then you can use that to, you know,
extrapolate, you know,
02:24.480 --> 02:28.300
next. Uh, I think a common issue
02:28.580 --> 02:32.200
is that a lot of people who
are using AI have only been using it since
02:32.360 --> 02:36.200
2023 or 2024.
That simply does not give you enough data
02:36.420 --> 02:37.280
about the past.
02:38.380 --> 02:41.730
Like,
AI has been a thing since at least 2019.
02:41.820 --> 02:44.230
Uh,
transformers in particular have been a
02:44.440 --> 02:46.680
2019.
03:49.780 --> 03:53.600
The third reason you might be convinced
is more of a speculative
03:53.660 --> 03:57.390
argument.
There isn't a lot of actual data backing
03:57.420 --> 04:01.300
something I believe. Uh, so today's, uh,
AI
04:01.320 --> 04:05.030
models can do tasks
that can be done quickly.
04:05.060 --> 04:07.840
Like, if you can do a task that will take,
you know, one minute or 10 minutes, you
04:07.880 --> 04:11.720
know, the AI can do it for you.
If you want to do a task
04:11.740 --> 04:14.640
know, one hour of your time
and lots of failed attempts, you know,
04:14.700 --> 04:18.660
probably not do it for you. Uh, however,
the tasks
04:18.680 --> 04:21.480
which it can do,
those it can do way faster than you.
04:21.540 --> 04:24.740
Like, if it will take, you know,
10 minutes, the AI will do it in,
04:24.820 --> 04:28.800
seconds. Uh, AIs operate, at least as
04:28.840 --> 04:32.420
of today, between 10 to, you know,
100 times the speed at which you're
04:32.560 --> 04:34.180
operating.
04:35.760 --> 04:39.640
Also, the time horizon of the tasks
04:39.720 --> 04:43.560
is increasing. So, you know,
AI six months ago could do tasks
04:43.620 --> 04:44.000
that were
04:44.900 --> 04:48.140
maybe one minute long,
and the AI today can do tasks that are,
04:48.160 --> 04:50.760
long. There is a trend that
is increasing in that direction.
04:50.780 --> 04:54.650
The technical reason behind this is, uh,
increasing
04:54.740 --> 04:58.220
compute being given to reinforcement
learning on top of transformers,
04:58.760 --> 05:02.160
and that's something I'll get to later,
but this is a specific technical reason
05:02.200 --> 05:06.060
this is happening. Uh,
my view of intelligence, at
05:06.100 --> 05:06.320
least
05:07.140 --> 05:10.940
to some extent, is
that you get better at a thing
05:11.000 --> 05:14.170
by doing it a lot of times. You know,
you get better at giving speeches by
05:14.180 --> 05:17.000
lot of speeches.
You get better at writing code by writing
05:17.020 --> 05:19.780
code. You get better
05:21.040 --> 05:24.280
at, you know,
producing movies by producing lots of
05:24.340 --> 05:27.240
And if you can run faster,
you can run more
05:27.320 --> 05:31.120
iterations. Uh, all scientific discovery,
in my
05:31.140 --> 05:34.860
view, works like this. Uh, every lab
05:34.880 --> 05:38.580
experiment you run
is yet another iteration
05:38.620 --> 05:42.460
system. You know,
it could be microscopes
05:42.500 --> 05:46.120
biology. It could be, uh, you know,
05:46.140 --> 05:49.010
telescopes that help you study stars.
05:49.460 --> 05:53.400
It could be any field. And
if you can run more iterations,
05:53.440 --> 05:57.080
you can make progress faster. Uh,
there are some people who
05:57.120 --> 06:01.040
think that the amount of AI-driven
scientific progress
06:01.100 --> 06:04.984
we can get is limited because...The
physical world
06:05.003 --> 06:07.883
is slow. You know, labs are slow. Labs
are expensive.
06:07.904 --> 06:10.664
Running experiments takes time,
and this is
06:10.804 --> 06:14.623
true. However, A, there
are domains where the
06:14.664 --> 06:16.183
labs are cheap. Uh,
06:19.003 --> 06:21.344
learning how to get good speeches
is very cheap.
06:21.383 --> 06:24.104
Learning how to get good at math
and software is cheap.
06:24.204 --> 06:27.244
Uh, like, it does not cost a lot of money,
and I
06:27.284 --> 06:30.962
think that is partly why we have gotten a
lot of good
06:31.003 --> 06:34.964
capabilities from reinforcement learning
in math and software.
06:34.974 --> 06:38.484
Like, AI is extremely good at math
and software in particular as of
06:38.584 --> 06:39.464
2025.
06:41.183 --> 06:44.484
I think this also carries over to other
fields of scientific discovery.
06:46.024 --> 06:49.683
Maybe the speed up is not that much,
but there will be a
06:49.724 --> 06:51.943
speed up. Uh,
06:52.804 --> 06:55.524
you can still get super intelligence this
way.
06:55.703 --> 06:59.424
If, you know,
humanity would otherwise have taken,
06:59.464 --> 07:03.344
know, invent a certain technology,
if AI can do
07:03.384 --> 07:06.443
that in ten years, that
is still super intelligence.
07:06.464 --> 07:10.323
Like, imagine if you
were sitting in 1800 and you could get
07:10.404 --> 07:14.264
entire 1800 to 1900, you know,
technological inventions to you in
07:14.304 --> 07:18.114
ten years. Or imagine you were in 1900,
you know,
07:18.123 --> 07:21.964
before World War I, you know,
people used sticks and stones to literally
07:22.024 --> 07:25.703
fight,
and by the end of the century people used,
07:25.864 --> 07:28.164
and, you know, space satellites
07:28.484 --> 07:31.904
and, uh, bomber planes and so
07:32.024 --> 07:33.284
on.
07:35.304 --> 07:39.203
If you can speed that up,
you can compress the amount of time
07:39.244 --> 07:40.024
requires to
07:40.864 --> 07:44.784
invent things,
that gets you super intelligence, and
07:44.943 --> 07:46.104
third reason.
07:48.884 --> 07:52.683
The fourth reason I believe we might get
super intelligence in the next
07:52.744 --> 07:56.244
five years is, uh, model scaling. This
is a
07:56.284 --> 07:58.724
well-known trend among machine learning
researchers.
07:58.744 --> 08:02.703
However, if you are not very in the space,
you may not be aware of it.
08:02.804 --> 08:06.544
Uh,
most of the AI progress in the last six
08:06.623 --> 08:10.404
2025 has been driven by one trend
and one trend only,
08:10.464 --> 08:13.743
which is you take a transformer,
you give it more data, more
08:13.803 --> 08:17.703
computation,
and it will produce better results.
08:17.724 --> 08:19.943
People don't know how this happens,
but this happens.
08:19.963 --> 08:23.063
This predictability repeatedly happens.
08:23.183 --> 08:26.044
Uh, the two ingredients are compute
and data.
08:26.134 --> 08:30.073
Uh,
data is just the entire internet's data
08:30.083 --> 08:33.644
the model. Compute
is just more computers that people have
08:33.664 --> 08:36.764
purchased. Uh,
there's a special type of computer used.
08:36.774 --> 08:39.223
They're called graphics processing units.
08:39.284 --> 08:42.683
They are good at this operation of matrix
multiplication in
08:42.723 --> 08:45.884
particular.
But basically it's just more computers.
08:45.904 --> 08:46.964
We have gone from
08:48.084 --> 08:51.084
spending, you know, like,
08:51.104 --> 08:54.704
$100,000,
which could be literally just one machine
08:54.764 --> 08:57.743
person's room, you know, $100 billion
08:58.784 --> 09:02.684
worth of compute to train a single AI
model that is, you know, an
09:02.764 --> 09:06.284
entire data center spanning, you know,
thousands of
09:06.324 --> 09:10.284
acres. This scale up in, you know,
09:10.324 --> 09:14.153
six orders of magnitude has just been
driven because every
09:14.243 --> 09:18.164
time we notice we use more compute
and more data, we get more AI
09:18.184 --> 09:21.784
performance.
We can now even predict this up to three
09:21.844 --> 09:25.243
decimal places. There
is a law called the Chinchilla Scaling
09:25.304 --> 09:28.024
Law, uh, I'll put
that on the screen right now,
09:29.364 --> 09:33.243
where N is the number of parameters,
like how big the model
09:33.304 --> 09:36.504
is,
and D is how much compute we give to it.
09:36.564 --> 09:37.223
Assume
09:38.153 --> 09:42.144
we s-
show the model every token in the data
09:42.204 --> 09:45.844
exactly once,
this L gives you the training loss.
09:45.904 --> 09:49.104
This is a proxy for how good the model
09:49.243 --> 09:53.144
performs. There is one important
09:53.184 --> 09:56.924
nitpick here that's really important,
which is that the training
09:56.983 --> 10:00.424
loss correlates with your actual
performance that you care about,
10:00.464 --> 10:03.344
but it is not the performance itself.
10:03.404 --> 10:07.304
It is possible for training loss to reduce
and for capabilities to not increase
10:07.344 --> 10:10.824
that much.
It is possible for training loss to reduce
10:10.864 --> 10:14.684
capabilities to increase a lot.
Predicting this
10:14.743 --> 10:18.564
relationship is hard between training loss
and capabilities,
10:18.624 --> 10:21.464
but there is a relationship.
More training loss does mean more
10:21.483 --> 10:25.464
capabilities.
Most AI researchers across the field have
10:25.524 --> 10:29.404
failed to predict exactly what training
loss leads to exactly
10:29.444 --> 10:33.404
what capabilities. There
are few people who have predicted it,
10:33.483 --> 10:37.243
and these people are, for example,
Ilya Sutskever at OpenAI was very
10:37.284 --> 10:41.104
bullish on this, and he predicts,
you know, we will get
10:41.144 --> 10:44.844
super intelligence in the next five to ten
years.
10:44.884 --> 10:48.664
People who have actually believed in this
law have generally predicted that AI will
10:48.723 --> 10:51.324
come sooner than what most other people in
the field
10:51.483 --> 10:56.024
predict.
10:58.503 --> 11:02.404
Uh,
we still have at least two orders of
11:02.444 --> 11:06.284
up left. Uh, we have spent, like I said,
you know, $10 billion has
11:06.324 --> 11:10.124
already been spent and the next set of ra-
data centers we're making are being
11:10.144 --> 11:13.304
made so that we spend $100 billion on the
training of a single
11:13.384 --> 11:17.304
model. We might scale
that up another order of magnitude,
11:17.364 --> 11:19.344
trillion dollars on a single model.
11:19.364 --> 11:23.223
It seems likely we will not go beyond a
trillion dollars on a single model
11:23.944 --> 11:26.844
unless we get, you know,
really exceptional capabilities.
11:26.884 --> 11:30.824
This alone means, you know,
two more orders of magnitude from 10
11:30.944 --> 11:32.404
trillion,
11:33.304 --> 11:36.223
which you can compare to, you know,
the six orders of magnitude they have
11:36.304 --> 11:39.884
gone through. There
are some people who say, you know,
11:39.944 --> 11:43.723
model, which was GPT-4.5 was not that
11:43.884 --> 11:47.723
impressive, meaning, you know,
this law might be weakening and that
11:47.764 --> 11:51.684
might be true,
but just remember we still have two more
11:53.964 --> 11:57.584
The fifth reason you might be convinced
super intelligence might be coming in the
11:57.644 --> 12:00.983
next five years is reinforcement learning
scaling.
12:01.024 --> 12:03.864
Reinforcement learning
is an old technique in machine learning.
12:03.904 --> 12:07.764
However,
it has been applied to transformers in
12:07.804 --> 12:11.464
Literally only one year ago we figured out
how
12:11.644 --> 12:15.144
to apply reinforcement learning on a
transformer to get
12:15.204 --> 12:18.843
even more capabilities.Uh, reinforcement
12:18.904 --> 12:22.843
learning allows the model to try
something, fail, try something, fail,
12:22.883 --> 12:25.584
try again,
and actually gets the correct answer.
12:25.684 --> 12:29.564
Uh,
this is different from the base model
12:29.584 --> 12:31.973
know, cached answer which it gives out.
12:31.983 --> 12:35.483
Reinforcement learning allows us to spend
more compute on a single
12:35.623 --> 12:39.454
task, uh,
whereas earlier it could just give the
12:39.524 --> 12:43.223
answer for the task. There are multiple,
you
12:43.304 --> 12:47.103
know,
tasks which people consider difficult
12:47.164 --> 12:49.684
would never crack, you know,
as of 2023 and
12:49.723 --> 12:53.483
2024. Uh, researchers said that
12:53.623 --> 12:55.603
certain benchmarks would not be cracked.
12:55.644 --> 12:56.544
People said, you know,
12:57.524 --> 12:59.623
Humanities Last Exam would not be cracked.
12:59.644 --> 13:01.983
People said, you know,
ARC AGI would not be cracked.
13:02.024 --> 13:05.164
People said, you know,
mathematics would not be cracked.
13:05.243 --> 13:07.404
All three of these have been cracked.
13:07.483 --> 13:11.304
Uh,
Humanities Last Exam consists of lot of
13:11.363 --> 13:15.003
answers by PhDs in various fields of both
science
13:15.243 --> 13:18.024
and social science. This includes biology,
chemistry,
13:18.034 --> 13:22.023
archeology, uh, journalism, law and so on.
13:22.123 --> 13:25.723
Uh, all these PhDs agreed that, you know,
if these questions get
13:25.763 --> 13:29.723
cracked, that means the model
is as good as a PhD in those fields,
13:29.743 --> 13:32.584
seeing AI has already solved, like,
almost half of that data
13:32.684 --> 13:34.763
set.
13:35.924 --> 13:39.644
We have started seeing a loss to predict,
you
13:39.723 --> 13:40.243
know,
13:41.123 --> 13:44.623
how much computation leads to how much
performance,
13:45.123 --> 13:49.103
like we had with model scaling where we
have a very clear empirical law which
13:49.123 --> 13:53.084
gives you up to three decimal places
accuracy predictions.
13:53.204 --> 13:57.003
Uh,
we do not have as accurate predictions for
13:57.084 --> 14:00.424
scaling of reinforcement learning,
because it is relatively new.
14:00.444 --> 14:03.523
We have less data points,
we have tried this for less longer.
14:03.544 --> 14:07.404
However,
people are already starting to make some
14:07.444 --> 14:08.804
are starting to see that
14:09.784 --> 14:13.743
every time we increase the compute by,
you know, 10x, we get a
14:13.863 --> 14:17.743
little bit of improvement. So
if you spend, you know, $10 per task,
14:17.824 --> 14:21.343
spend $100, then you spend $1000 a task,
you get little improvement, little
14:21.384 --> 14:23.103
improvement, little improvement.
14:23.144 --> 14:26.824
We have gone all the way,
at least as per public data of OpenAI,
14:26.843 --> 14:30.684
the way up to $500,000 per task,
and we also know
14:30.723 --> 14:34.404
that, uh, AI has been able to get,
you know, the gold
14:34.483 --> 14:38.134
medal on the International Mathematics
Olympiad, which is like top olympiad
14:38.464 --> 14:42.294
for mathematics worldwide.
We still have many,
14:42.324 --> 14:45.373
many,
many more orders of magnitude to go from,
14:45.384 --> 14:49.363
$100,000, and in theory what we can spend
14:49.404 --> 14:52.284
is,
which is at least a billion dollars per
14:52.304 --> 14:55.674
Like, if you were really solving,
you know, novel physics,
14:55.684 --> 14:59.564
know, novel chemistry, if you were,
you know, trying to solve something like
14:59.584 --> 15:03.084
cancer or trying to, you know,
solve math theorems that have, you know,
15:03.123 --> 15:06.824
unsolved for, you know, the past century,
like, why would you not spend a billion
15:06.884 --> 15:10.843
dollars on those tasks?
So we have many more orders
15:10.863 --> 15:14.424
of magnitude to go here.
We have a lot more compute we
15:14.464 --> 15:18.324
this,
and we don't know how far this law scales,
15:18.363 --> 15:21.824
find out very soon.
15:24.023 --> 15:27.684
The sixth and final reason you might be
convinced superintelligence might be
15:27.704 --> 15:31.304
by 2030 is that reinforcement learning has
worked
15:31.343 --> 15:34.603
before. Like I said, this
is a not new technique.
15:34.644 --> 15:38.343
We are applying it to transformers newly
because transformers can, you know,
15:38.363 --> 15:42.184
understand English
and generalize across so many,
15:42.223 --> 15:46.023
tasks. Whereas earlier,
before transformers, we had, you know,
15:46.103 --> 15:49.623
doing specific things. Even
when we just had narrow models doing
15:49.723 --> 15:53.603
specific things,
like we had just one model
15:53.664 --> 15:56.863
we knew that reinforcement learning
actually works pretty well.
15:56.983 --> 16:00.704
Uh, we have, uh, an AI model that is,
you know, the world
16:00.763 --> 16:02.964
champion in chess
that knows nothing about chess.
16:02.983 --> 16:06.743
Like, no chess experts had to tell
that model how to get good at chess
16:07.003 --> 16:10.623
the best at chess. Uh, this
is obviously by DeepMind.
16:10.664 --> 16:14.363
This is, uh, AlphaZero. We
16:14.483 --> 16:18.424
also have DeepMind's model
that beat the best
16:18.483 --> 16:21.523
Go player in the world. This happened,
again, before transformers were even
16:21.584 --> 16:25.284
invented. Like, this is very old news,
but we have gotten
16:25.363 --> 16:27.444
superhuman performance in narrow domains.
16:27.454 --> 16:31.164
You know, poker has been solved.
Dota has been solved.
16:31.204 --> 16:35.103
StarCraft has been solved. People,
AI researchers moved on from
16:35.163 --> 16:38.804
games because they found them too easy,
they found it too easy to make an AI that
16:38.843 --> 16:41.164
would just completely obliterate all the
human players in a
16:41.223 --> 16:45.064
game. And that is why we
are now working with general
16:45.103 --> 16:48.804
domains.
So we know point number A on narrow
16:48.824 --> 16:51.723
domains,
reinforcement learning gets superhuman
16:51.743 --> 16:55.543
Point number B,
transformers can work on general domains.
16:55.584 --> 16:59.464
Point number C,
reinforcement learning plus transformers
16:59.483 --> 17:03.444
domains actually works. Point number D,
this thing scales up.
17:03.464 --> 17:05.304
Point number E, we are going to scale it
17:05.424 --> 17:07.743
up.
Enter "Subscribe: " followed by Signal username, email address or whatsapp phone number. You will receive atmost one update per month. View previous updates here.
Alternatively, you can use RSS, or a third-party service like blogtrottr.
Enter "Comment: " followed by comment. Alternatively, contact me