1
00:00:01,340 --> 00:00:04,940
In this video,
I am going to talk about four technical

2
00:00:05,000 --> 00:00:08,880
why I believe super intelligence may be
possible by 2030.

3
00:00:08,890 --> 00:00:11,940
The first reason is super intelligence via
speed.

4
00:00:11,960 --> 00:00:15,440
Then we have model scaling,
then we have RL scaling on top of

5
00:00:15,500 --> 00:00:19,410
transformers. Lastly, we have RL scaling,
uh, outside of

6
00:00:19,500 --> 00:00:23,420
transformers. The

7
00:00:23,460 --> 00:00:27,420
third reason you might be convinced
is more of a speculative argument.

8
00:00:27,440 --> 00:00:31,320
There isn't a lot of actual data backing
this, but this is something I

9
00:00:31,360 --> 00:00:35,220
believe. Uh, so today's, uh, AI models

10
00:00:35,540 --> 00:00:38,340
can do tasks that can be done quickly.

11
00:00:38,380 --> 00:00:41,160
Like, if you can do a task that will take,
you know, one minute or 10 minutes, you

12
00:00:41,200 --> 00:00:45,040
know the AI can do it for you.
If you want to do a task

13
00:00:45,060 --> 00:00:47,960
know, one hour of your time
and lots of failed attempts,

14
00:00:48,020 --> 00:00:51,980
probably not do it for you. Uh, however,
the tasks

15
00:00:52,000 --> 00:00:54,810
which it can do,
those it can do way faster than you.

16
00:00:54,840 --> 00:00:58,070
Like, if it will take, you know,
10 minutes, the AI will do it in,

17
00:00:58,140 --> 00:01:02,120
seconds. Uh, AIs operate, at least as

18
00:01:02,160 --> 00:01:05,760
of today, between 10 to, you know,
100 times the speed at which you are

19
00:01:05,880 --> 00:01:07,500
operating.

20
00:01:09,020 --> 00:01:12,960
Also, the time horizon of the tasks

21
00:01:13,040 --> 00:01:16,880
is increasing. So, you know,
AI six months ago could do tasks

22
00:01:16,940 --> 00:01:17,320
that were

23
00:01:18,240 --> 00:01:21,460
maybe one minute long
and the AI today can do tasks that are,

24
00:01:21,500 --> 00:01:24,080
long. There is a trend that
is increasing in that direction.

25
00:01:24,100 --> 00:01:28,020
The technical reason behind this is, uh,
increasing

26
00:01:28,060 --> 00:01:31,550
compute being given to reinforcement
learning on top of transformers.

27
00:01:32,060 --> 00:01:35,480
And that's something I'll get to later,
but this is specific technical reason why

28
00:01:35,520 --> 00:01:39,380
this is happening. Uh,
my view of intelligence, at

29
00:01:39,420 --> 00:01:39,660
least

30
00:01:40,480 --> 00:01:44,440
to some extent, is
that you get better at a thing by

31
00:01:44,480 --> 00:01:47,480
doing it a lot of times. You know,
you get better at giving speeches by

32
00:01:47,500 --> 00:01:50,300
lot of speeches.
You get better at writing code by writing

33
00:01:50,360 --> 00:01:53,100
code. You get better

34
00:01:54,360 --> 00:01:57,340
at, you know,
producing movies by producing lots of

35
00:01:57,660 --> 00:02:00,560
And if you can run faster,
you can run more

36
00:02:00,640 --> 00:02:04,440
iterations. Uh, all scientific discovery,
in my

37
00:02:04,480 --> 00:02:08,180
view, works like this. Uh, every lab

38
00:02:08,220 --> 00:02:11,880
experiment you run
is yet another iteration

39
00:02:11,940 --> 00:02:15,800
system. You know,
it could be microscopes

40
00:02:15,840 --> 00:02:19,440
biology, it could be, uh, you know,

41
00:02:19,460 --> 00:02:22,740
telescopes that help you study stars.

42
00:02:22,760 --> 00:02:26,720
It could be any field. And
if you can run more iterations,

43
00:02:26,760 --> 00:02:30,400
you can make progress faster. Uh,
there are some people who

44
00:02:30,440 --> 00:02:34,380
think that the amount of AI-driven
scientific progress

45
00:02:34,420 --> 00:02:36,720
we can get is limited because

46
00:02:37,580 --> 00:02:41,480
the physical world is slow. You know,
labs are slow, labs are expensive, running

47
00:02:41,540 --> 00:02:45,440
experiments takes time. And this is true.

48
00:02:45,480 --> 00:02:49,100
However, A, there
are domains where the labs are cheap.

49
00:02:49,220 --> 00:02:49,410
Uh,

50
00:02:52,340 --> 00:02:54,660
learning how to get good speeches
is very cheap.

51
00:02:54,720 --> 00:02:57,420
Learning how to get good at math
and software is cheap.

52
00:02:57,520 --> 00:03:00,200
Uh, like, it does not cost a lot of money.

53
00:03:00,210 --> 00:03:00,820
And I think

54
00:03:01,800 --> 00:03:05,120
that is partly why we have gotten a lot of
good capabilities

55
00:03:05,460 --> 00:03:08,290
from reinforcement learning in math
and software.

56
00:03:08,300 --> 00:03:11,820
Like, AI is extremely good at math
and software in particular as of

57
00:03:11,940 --> 00:03:15,880
2025. I think this also carries over to

58
00:03:15,960 --> 00:03:19,320
other fields of scientific discovery.

59
00:03:19,360 --> 00:03:22,980
Maybe the speed up is not that much,
but there will be a

60
00:03:23,060 --> 00:03:26,920
speed up. Uh, you can still get

61
00:03:26,960 --> 00:03:30,609
super intelligence this way. If, you know,
humanity would

62
00:03:30,640 --> 00:03:33,800
otherwise have taken, you know,
100 years to, you know, invent a certain

63
00:03:33,840 --> 00:03:35,600
technology, if AI can

64
00:03:36,540 --> 00:03:39,760
do that in 10 years, that
is still super intelligence.

65
00:03:39,780 --> 00:03:43,640
Like, imagine if you
were sitting in 1800 and you could get

66
00:03:43,700 --> 00:03:47,580
entire 1800 to 1900, you know,
technological inventions to you in

67
00:03:47,620 --> 00:03:51,440
10 years. Or imagine you were in 1900,
you know,

68
00:03:51,460 --> 00:03:55,300
before World War I, you know,
people used sticks and stones to literally

69
00:03:55,340 --> 00:03:59,180
fight.
And by the end of the century people used,

70
00:03:59,260 --> 00:04:02,300
and, you know, space satellites and,

71
00:04:03,740 --> 00:04:06,600
uh, bomber planes and so on.

72
00:04:08,620 --> 00:04:12,540
If you can speed that up,
you can compress the amount of time

73
00:04:12,560 --> 00:04:16,100
requires to invent things,
that gets you super

74
00:04:16,140 --> 00:04:18,640
intelligence, and that is my third

75
00:04:18,860 --> 00:04:22,760
reason. The fourth

76
00:04:22,800 --> 00:04:26,720
reason I believe we might get super
intelligence in the next five years

77
00:04:26,960 --> 00:04:30,800
is, uh, model scaling. This
is a well-known trend among

78
00:04:30,820 --> 00:04:34,440
machine learning researchers. However,
if you are not very in this space, you may

79
00:04:34,480 --> 00:04:38,240
not be aware of it. Uh,
most of the AI progress in the last

80
00:04:38,280 --> 00:04:42,170
six years from 2019 to 2025 has been
driven by one trend and

81
00:04:42,200 --> 00:04:46,000
one trend only, which
is you take a transformer, you give it

82
00:04:46,060 --> 00:04:49,620
more data, more computation,
and it will produce better

83
00:04:49,680 --> 00:04:53,260
results. This is magic.
People don't know how this happens,

84
00:04:53,280 --> 00:04:56,380
This predictability repeatedly happens.

85
00:04:56,520 --> 00:04:59,500
Uh, the two ingredients are compute
and data.

86
00:04:59,520 --> 00:05:03,460
Uh,
data is just the entire internet's data

87
00:05:03,520 --> 00:05:06,940
model. Compute is just more computers
that people have

88
00:05:06,980 --> 00:05:10,080
purchased. Uh,
there's a special type of computer used.

89
00:05:10,100 --> 00:05:12,540
They're called graphics processing units.

90
00:05:12,560 --> 00:05:16,000
They are good at this operation of matrix
multiplication in

91
00:05:16,040 --> 00:05:19,220
particular. But basically,
it's just more computers.

92
00:05:19,260 --> 00:05:20,280
We have gone from

93
00:05:21,360 --> 00:05:24,420
spending, you know, like,

94
00:05:24,430 --> 00:05:28,040
$100,000,
which could be literally just one machine

95
00:05:28,100 --> 00:05:31,980
person's room, to, you know, $100 billion

96
00:05:32,100 --> 00:05:36,020
worth of compute to train a single AI
model that is, you know, an

97
00:05:36,100 --> 00:05:39,620
entire data center spanning, you know,
thousands of

98
00:05:39,700 --> 00:05:40,190
acres.

99
00:05:41,880 --> 00:05:45,240
This scale up in, you know,
six orders of magnitude

100
00:05:45,780 --> 00:05:49,740
has just been driven,
because every time we notice we use more

101
00:05:49,780 --> 00:05:52,260
more data, we get more AI performance.

102
00:05:52,960 --> 00:05:56,260
We can now even predict this up to three
decimal places.

103
00:05:56,280 --> 00:05:59,420
There is a law called the Chinchilla
Scaling Law.

104
00:05:59,520 --> 00:06:02,657
Uh,
I'll put that on the screen right now....

105
00:06:02,668 --> 00:06:06,548
that N is the number of parameters,
like how big the model

106
00:06:06,608 --> 00:06:09,828
is,
and D is how much compute we give to it.

107
00:06:09,888 --> 00:06:10,488
Assume

108
00:06:11,428 --> 00:06:15,407
we s-
show the model every token in the data

109
00:06:15,548 --> 00:06:19,188
exactly once.
This L gives you the training loss.

110
00:06:19,227 --> 00:06:22,428
This is a proxy for how good the model

111
00:06:22,587 --> 00:06:23,248
performs.

112
00:06:25,087 --> 00:06:28,828
There is one important metric here that's
really important, which is

113
00:06:28,907 --> 00:06:32,788
that the training loss correlates with
your actual performance

114
00:06:32,847 --> 00:06:36,668
that you care about, but it
is not the performance itself.

115
00:06:36,727 --> 00:06:40,628
It is possible for training loss to reduce
and for capabilities to not increase

116
00:06:40,668 --> 00:06:44,147
that much.
It is possible for training loss to reduce

117
00:06:44,188 --> 00:06:48,008
capabilities to increase a lot.
Predicting this

118
00:06:48,068 --> 00:06:51,888
relationship is hard between training loss
and capabilities,

119
00:06:51,948 --> 00:06:54,788
but there is a relationship.
More training loss does mean more

120
00:06:54,808 --> 00:06:58,788
capabilities.
Most AI researchers across the field have

121
00:06:58,847 --> 00:07:02,727
failed to predict exactly what training
loss leads to exactly

122
00:07:02,768 --> 00:07:06,727
what capabilities. There
are few people who have predicted it,

123
00:07:06,808 --> 00:07:10,568
and these people are, for example,
Ilya Sutskever at OpenAI was very

124
00:07:10,587 --> 00:07:14,428
bullish on this, and he predicts,
you know, we will get

125
00:07:14,448 --> 00:07:17,347
super intelligence in the next 5 to 10
years.

126
00:07:18,168 --> 00:07:21,988
People who have actually believed in this
law have generally predicted that AI will

127
00:07:22,048 --> 00:07:24,647
come sooner than what most other people in
the field

128
00:07:24,788 --> 00:07:29,668
predict.

129
00:07:31,828 --> 00:07:35,727
Uh,
we still have at least two orders of

130
00:07:35,768 --> 00:07:39,607
up left. Uh, we have spent, like I said,
you know, $10 billion has

131
00:07:39,647 --> 00:07:43,448
already been spent and the next set of la-
data centers we're making are being

132
00:07:43,488 --> 00:07:46,628
made so that we spend $100 trillion on the
training of a single

133
00:07:46,707 --> 00:07:50,628
model. We might scale
that up another order of magnitude,

134
00:07:50,688 --> 00:07:52,308
trillion dollars on a single model.

135
00:07:52,688 --> 00:07:56,548
It seems likely we will not go beyond a
trillion dollars on a single model

136
00:07:57,268 --> 00:08:00,168
unless we get, you know,
really exceptional capabilities.

137
00:08:00,227 --> 00:08:04,147
This alone means, you know,
two more orders of magnitude from 10

138
00:08:04,248 --> 00:08:08,128
trillion. Which you can compare to,
you know, the six

139
00:08:08,168 --> 00:08:10,448
orders of magnitude they have already gone
through.

140
00:08:10,467 --> 00:08:13,948
There are some people who say, you know,
the latest model, which was

141
00:08:13,987 --> 00:08:17,048
GPT-4.5, was not that

142
00:08:17,207 --> 00:08:21,048
impressive, meaning, you know,
this law might be weakening, and that

143
00:08:21,107 --> 00:08:25,028
might be true,
but just remember we still have two more

144
00:08:27,268 --> 00:08:30,908
The fifth reason you might be convinced
super intelligence might be coming in the

145
00:08:30,948 --> 00:08:34,308
next five years is reinforcement learning
scaling.

146
00:08:34,328 --> 00:08:37,188
Reinforcement learning
is an old technique in machine learning.

147
00:08:37,227 --> 00:08:41,087
However,
it has been applied to transformers in

148
00:08:41,127 --> 00:08:44,928
Literally only one year ago we figured out
how

149
00:08:44,968 --> 00:08:48,508
to apply reinforcement learning on a
transformer to get

150
00:08:48,568 --> 00:08:52,188
even more capabilities. Uh, reinforcement

151
00:08:52,248 --> 00:08:56,168
learning allows the model to try
something, fail, try something, fail,

152
00:08:56,208 --> 00:08:58,887
try again,
and until it gets the correct answer.

153
00:08:59,008 --> 00:09:02,887
Uh,
this is different from the base model

154
00:09:02,928 --> 00:09:05,288
know, cached answer which it gives out.

155
00:09:05,308 --> 00:09:08,807
Reinforcement learning allows us to spend
more compute on a single

156
00:09:08,948 --> 00:09:12,768
task. Uh,
whereas earlier it would just give the

157
00:09:12,808 --> 00:09:14,088
answer for the task.

158
00:09:15,068 --> 00:09:18,908
There are multiple, you know, tasks
which people consider

159
00:09:18,948 --> 00:09:22,848
difficult, which, you know,
AI would never crack, you know, as of 2023

160
00:09:22,887 --> 00:09:26,808
and 2024. Uh, researchers said that

161
00:09:26,948 --> 00:09:28,948
certain benchmarks would not be cracked.

162
00:09:28,968 --> 00:09:29,867
People said, you know,

163
00:09:30,848 --> 00:09:32,938
humanity's last exam would not be cracked.

164
00:09:32,968 --> 00:09:35,308
People said, you know,
ARC AGI would not be cracked.

165
00:09:35,348 --> 00:09:38,488
People said, you know,
mathematics olympiads would not be

166
00:09:38,568 --> 00:09:40,728
All three of these have been cracked.

167
00:09:40,828 --> 00:09:44,627
Uh,
humanity's last exam consists of lot of

168
00:09:44,668 --> 00:09:48,328
answers by PhDs in various fields of both
science

169
00:09:48,568 --> 00:09:51,268
and social science. This includes biology,
chemistry,

170
00:09:51,387 --> 00:09:55,348
archeology, uh, journalism, law and so on.

171
00:09:55,448 --> 00:09:59,028
Uh, all these PhDs agreed that, you know,
if these questions get

172
00:09:59,088 --> 00:10:03,028
cracked, that means the model
is as good as a PhD in those fields,

173
00:10:03,088 --> 00:10:05,928
seeing AI is already solved like almost
half of that data

174
00:10:05,988 --> 00:10:09,877
set. We have

175
00:10:10,028 --> 00:10:13,137
started seeing a loss to predict,
you know,

176
00:10:14,448 --> 00:10:18,387
how much computation leads to how much
performance.

177
00:10:18,448 --> 00:10:22,428
Like where we had with model scaling where
we have a very clear empirical law which

178
00:10:22,548 --> 00:10:26,407
gives you up to three decimal places
accuracy predictions.

179
00:10:26,528 --> 00:10:30,328
Uh,
we do not have as accurate predictions for

180
00:10:30,407 --> 00:10:33,718
scaling of reinforcement learning because
it is relatively new.

181
00:10:33,768 --> 00:10:36,848
We have less data points,
we have tried this for less longer.

182
00:10:36,867 --> 00:10:40,728
However,
people are already starting to make some

183
00:10:40,768 --> 00:10:42,127
are starting to see that

184
00:10:43,088 --> 00:10:47,068
every time we increase the compute by,
you know, 10x, we get a

185
00:10:47,188 --> 00:10:51,068
little bit of improvement. So
if you spend, you know, $10 per task,

186
00:10:51,147 --> 00:10:54,668
spend $100,
then you spend $1,000 per task,

187
00:10:54,788 --> 00:10:56,428
improvement, little improvement.

188
00:10:56,468 --> 00:11:00,147
We have gone all the way,
at least as per public data of OpenAI,

189
00:11:00,168 --> 00:11:03,948
the way up to $500,000 per task,
and we also know

190
00:11:04,008 --> 00:11:07,748
that, uh, AI has been able to get,
you know, the gold

191
00:11:07,808 --> 00:11:11,708
medal on the International Mathematics
Olympiad, which is like the olympiad

192
00:11:11,788 --> 00:11:15,617
for mathematics worldwide.
We still have many,

193
00:11:15,647 --> 00:11:18,718
many,
many more orders of magnitude to go from,

194
00:11:18,718 --> 00:11:22,348
$100,000 and in theory what we can

195
00:11:22,448 --> 00:11:25,588
spend is,
which is at least a billion dollars per

196
00:11:25,627 --> 00:11:28,998
Like if you were really solving, you know,
novel physics, novel mathematics, you

197
00:11:29,008 --> 00:11:32,887
know, novel chemistry. If you were,
you know, trying to solve something like

198
00:11:32,907 --> 00:11:36,407
cancer or trying to, you know,
solve math theorems that have, you know,

199
00:11:36,448 --> 00:11:40,147
unsolved for, you know, the past century,
like, why would you not spend a billion

200
00:11:40,188 --> 00:11:44,168
dollars on this task?
So we have many more orders

201
00:11:44,188 --> 00:11:48,088
of magnitude to go here.
We have a lot more compute we're going to

202
00:11:48,488 --> 00:11:52,428
and we don't know how far this law scales,
but we are going to find out very

203
00:11:52,508 --> 00:11:55,147
soon.

204
00:11:57,348 --> 00:12:00,688
The sixth and final reason you might be
convinced super intelligence might be

205
00:12:00,748 --> 00:12:04,627
coming by 2030 is
that reinforcement learning has worked

206
00:12:04,668 --> 00:12:07,928
before. Like I said, this
is a not new technique.

207
00:12:07,968 --> 00:12:11,668
We are applying it to transformers newly
because transformers can, you know,

208
00:12:11,688 --> 00:12:15,528
understand English
and generalize across so many,

209
00:12:15,548 --> 00:12:19,387
tasks.
Whereas earlier before transformers we

210
00:12:19,407 --> 00:12:22,948
doing specific things. Even
when we just had narrow models doing

211
00:12:22,968 --> 00:12:25,328
specific things,
like we had just one model and

212
00:12:25,338 --> 00:12:26,068
(laughs)

213
00:12:26,088 --> 00:12:29,028
... one task. Even back then we knew
that reinforcement learning actually works

214
00:12:29,068 --> 00:12:32,728
pretty well. Uh, we have, uh, an AI

215
00:12:32,768 --> 00:12:35,988
model that is, you know,
the world champion in chess

216
00:12:36,008 --> 00:12:39,897
chess.
Like no chess experts had to tell

217
00:12:39,928 --> 00:12:43,348
it's the best at chess. Uh, this
is obviously by

218
00:12:43,387 --> 00:12:46,387
DeepMind. This is, uh, AlphaZero.

219
00:12:46,448 --> 00:12:46,668
We

220
00:12:47,768 --> 00:12:51,708
also have DeepMind's model
that beat the best

221
00:12:51,788 --> 00:12:55,548
Go player in the world.
This happened again before transformers

222
00:12:55,588 --> 00:12:59,228
Like this is very old news,
but we have gotten superhuman

223
00:12:59,268 --> 00:13:02,968
performance in narrow domains. You know,
poker has been solved.

224
00:13:03,008 --> 00:13:06,268
Dota has been solved.
StarCraft has been solved.

225
00:13:06,308 --> 00:13:10,228
People,
AI researchers moved on from games because

226
00:13:10,248 --> 00:13:13,778
They found it too easy to make an AI
that will just completely obliterate all

227
00:13:13,808 --> 00:13:17,588
human players in a game. And that
is why we are now

228
00:13:17,608 --> 00:13:21,528
working with general domains.
So we know point number A

229
00:13:21,608 --> 00:13:25,048
on narrow domains,
reinforcement learning gets superhuman

230
00:13:25,068 --> 00:13:28,867
Point number B,
transformers can work on general domains.

231
00:13:28,907 --> 00:13:32,788
Point number C,
reinforcement learning plus transformers

232
00:13:32,808 --> 00:13:36,768
domains actually works. Point number D,
this thing scales up.

233
00:13:36,808 --> 00:13:38,627
Point number E, we are going to scale it

234
00:13:38,748 --> 00:13:41,147
up.