1
00:00:00,000 --> 00:00:02,839
Okay, let's unpack this. You know, thousands

2
00:00:02,839 --> 00:00:06,320
of years ago, Plato introduced this incredible

3
00:00:06,320 --> 00:00:08,320
thought experiment that philosophers have, well,

4
00:00:08,460 --> 00:00:10,339
they've wrestled with it ever since. Right, the

5
00:00:10,339 --> 00:00:12,460
allegory of the cave. Exactly, the allegory of

6
00:00:12,460 --> 00:00:14,720
the cave. And the premise is really striking.

7
00:00:15,140 --> 00:00:18,019
So imagine prisoners who have been chained in

8
00:00:18,019 --> 00:00:21,579
this deep cave for their entire lives, right?

9
00:00:21,780 --> 00:00:24,579
Just facing a wall. Yeah, facing a blank stone

10
00:00:24,579 --> 00:00:29,980
wall. And behind them is this raging fire. Between

11
00:00:29,980 --> 00:00:31,859
the fire and the prisoners, there are people

12
00:00:31,859 --> 00:00:34,859
carrying various objects past. But the prisoners

13
00:00:34,859 --> 00:00:38,039
can't see the object. No, exactly. All they can

14
00:00:38,039 --> 00:00:40,219
see are the shadows being cast on the wall in

15
00:00:40,219 --> 00:00:44,509
front of them. To these people, those flat, flickering,

16
00:00:44,649 --> 00:00:48,049
two -dimensional shadows are the absolute entirety

17
00:00:48,049 --> 00:00:51,750
of reality. Right, they have no concept of a

18
00:00:51,750 --> 00:00:54,350
three -dimensional world. None. No concept of

19
00:00:54,350 --> 00:00:59,789
depth or light or actual color. But Plato asked

20
00:00:59,789 --> 00:01:02,149
this really profound question with this setup.

21
00:01:03,240 --> 00:01:05,819
He asked, if different observers are looking

22
00:01:05,819 --> 00:01:07,879
at completely different shadows on that wall,

23
00:01:08,239 --> 00:01:11,519
so maybe one sees a shadow of a vase and another

24
00:01:11,519 --> 00:01:15,469
sees a shadow of a dog, could they ever, purely

25
00:01:15,469 --> 00:01:18,329
by studying their own respective shadows, converge

26
00:01:18,329 --> 00:01:21,489
on the exact same understanding of the true multi

27
00:01:21,489 --> 00:01:23,590
-dimensional reality that's actually casting

28
00:01:23,590 --> 00:01:25,650
them? It's a foundational question about the

29
00:01:25,650 --> 00:01:28,170
nature of perception itself. I mean, historically,

30
00:01:28,310 --> 00:01:30,430
we've applied it to human consciousness, like

31
00:01:30,430 --> 00:01:33,290
asking if any two people ever truly experienced

32
00:01:33,290 --> 00:01:36,329
the exact same reality. Right. But what's fascinating

33
00:01:36,329 --> 00:01:39,969
here is that a thought experiment designed millennia

34
00:01:39,969 --> 00:01:43,030
ago to probe human philosophy has suddenly become

35
00:01:43,030 --> 00:01:46,049
arguably the most urgent high stakes question

36
00:01:46,049 --> 00:01:48,730
in modern computer science. Oh, absolutely. We

37
00:01:48,730 --> 00:01:50,709
are no longer just wondering about humans in

38
00:01:50,709 --> 00:01:53,319
a cave. We are asking this exact question about

39
00:01:53,319 --> 00:01:55,840
artificial intelligence. Because we have built

40
00:01:55,840 --> 00:02:00,060
these massive sprawling AI models that are essentially

41
00:02:00,060 --> 00:02:03,079
trapped in their own caves. Exactly. I mean,

42
00:02:03,180 --> 00:02:04,840
some of these models are trained exclusively

43
00:02:04,840 --> 00:02:07,280
on text. They are fed billions of books, articles,

44
00:02:07,540 --> 00:02:11,500
websites, but they never actually see a sunset.

45
00:02:11,900 --> 00:02:14,180
They only see the shadows of human language describing

46
00:02:14,180 --> 00:02:17,139
a sunset, you know? Yeah, just text on the screen.

47
00:02:17,580 --> 00:02:19,520
And then on the other side of the spectrum, we

48
00:02:19,520 --> 00:02:21,939
have vision models trained exclusively on images

49
00:02:21,939 --> 00:02:24,800
and video. Right. They don't have language. They

50
00:02:24,800 --> 00:02:27,699
only see the shadows of pixels, light, and geometry.

51
00:02:28,400 --> 00:02:30,840
Yet despite being locked in entirely different

52
00:02:30,840 --> 00:02:33,180
sensory caves, they are all ultimately observing

53
00:02:33,180 --> 00:02:36,360
the exact same underlying world. And that shared

54
00:02:36,360 --> 00:02:39,139
underlying world is really the crux of the research

55
00:02:39,139 --> 00:02:42,000
we are doing a deep dive into today. It is. We

56
00:02:42,000 --> 00:02:44,620
are unpacking a genuinely groundbreaking paper

57
00:02:44,620 --> 00:02:47,469
today. It's titled Proving and Improving the

58
00:02:47,469 --> 00:02:50,349
Platonic Representation Hypothesis with Jeppas.

59
00:02:50,530 --> 00:02:53,150
And that's authored by Randall Balastriero and

60
00:02:53,150 --> 00:02:55,949
Jan Lacoon. Yes. Two very heavy hitters in the

61
00:02:55,949 --> 00:02:58,550
field. So let's lay out the mission for this

62
00:02:58,550 --> 00:03:00,689
deep dive, because the scope of what they've

63
00:03:00,689 --> 00:03:02,710
accomplished here is just massive. It really

64
00:03:02,710 --> 00:03:05,150
is. We are going to explore how these authors

65
00:03:05,150 --> 00:03:07,650
take a concept that was previously just, well,

66
00:03:07,729 --> 00:03:10,599
a very compelling theory. This idea that as AI

67
00:03:10,599 --> 00:03:13,020
models get bigger and smarter, their internal

68
00:03:13,020 --> 00:03:15,219
understanding of the world naturally converges

69
00:03:15,219 --> 00:03:18,520
into one shared reality, and they mathematically

70
00:03:18,520 --> 00:03:21,360
prove it. Right. They turn it from a hunch into

71
00:03:21,360 --> 00:03:25,120
an ironclad theorem. Exactly. But in the process

72
00:03:25,120 --> 00:03:27,979
of proving that, they accidentally expose this

73
00:03:27,979 --> 00:03:30,520
massive, highly uncomfortable inefficiency in

74
00:03:30,520 --> 00:03:33,280
how today's biggest and most famous AI models

75
00:03:33,280 --> 00:03:35,740
are fundamentally built. And the implications

76
00:03:35,740 --> 00:03:37,780
of that exposed inefficiency are staggering.

77
00:03:38,099 --> 00:03:40,560
I mean, it challenges the entire brute force

78
00:03:40,560 --> 00:03:42,680
scaling paradigm that currently dominates the

79
00:03:42,680 --> 00:03:44,620
tech industry. Which is all about more, right?

80
00:03:44,800 --> 00:03:47,460
Exactly. We are living in an era where the prevailing

81
00:03:47,460 --> 00:03:50,360
wisdom is simply more. More data. Larger data

82
00:03:50,360 --> 00:03:52,840
sets, more parameters, bigger data centers. We

83
00:03:52,840 --> 00:03:54,919
assume that intelligence is just a byproduct

84
00:03:54,919 --> 00:03:57,479
of scale. Just feed the machine more and it'll

85
00:03:57,479 --> 00:04:01,180
get smarter. Right. But... Ballastroero and Lacoon

86
00:04:01,180 --> 00:04:03,680
are mathematically demonstrating that the objective

87
00:04:03,680 --> 00:04:05,639
function we are using to train these massive

88
00:04:05,639 --> 00:04:08,699
models is inherently flawed for the goal of actually

89
00:04:08,699 --> 00:04:12,039
understanding reality. Which is exactly why you,

90
00:04:12,379 --> 00:04:14,629
the listener, need to care about this. You are

91
00:04:14,629 --> 00:04:16,370
constantly bombarded with headlines about the

92
00:04:16,370 --> 00:04:19,209
latest AI model requiring trillions of words,

93
00:04:19,509 --> 00:04:21,649
demanding data centers the size of small cities.

94
00:04:22,009 --> 00:04:23,769
Consuming electrical power on the scale of whole

95
00:04:23,769 --> 00:04:27,269
nations, yeah. Exactly. We are treating intelligence

96
00:04:27,269 --> 00:04:29,870
like this massive industrial mining operation.

97
00:04:30,610 --> 00:04:33,509
But this research suggests there is a fundamentally

98
00:04:33,509 --> 00:04:37,430
smarter way. A vastly smarter way. A method that

99
00:04:37,430 --> 00:04:39,709
isn't just a little better, but is mathematically

100
00:04:39,709 --> 00:04:43,410
proven to be up to 42 times more efficient at

101
00:04:43,410 --> 00:04:45,550
building an AI that genuinely understands the

102
00:04:45,550 --> 00:04:50,490
shape of the world. 42 times. It's wild. To appreciate

103
00:04:50,490 --> 00:04:53,029
how they achieve that efficiency, we first have

104
00:04:53,029 --> 00:04:55,269
to understand the mathematical proof they constructed.

105
00:04:55,769 --> 00:04:58,089
We need to look at how they demonstrated that

106
00:04:58,089 --> 00:05:00,790
different AI modalities, those separate models

107
00:05:00,790 --> 00:05:02,829
trapped in their respective text and image caves,

108
00:05:03,269 --> 00:05:06,730
can actually be proven to learn the exact same

109
00:05:06,730 --> 00:05:10,329
reality. So to do that, we need to properly define

110
00:05:10,329 --> 00:05:13,209
the platonic representation hypothesis first.

111
00:05:13,269 --> 00:05:16,569
Right. Now, I recall this hypothesis gaining

112
00:05:16,569 --> 00:05:19,269
a lot of traction recently. specifically around

113
00:05:19,269 --> 00:05:23,170
a paper by He and colleagues back in 2024. But

114
00:05:23,170 --> 00:05:25,589
up until now, it was really just an observation,

115
00:05:25,730 --> 00:05:28,990
right? Yeah, purely empirical. Researchers were

116
00:05:28,990 --> 00:05:31,069
basically looking under the hood of these massive

117
00:05:31,069 --> 00:05:33,709
models, say a giant language model and a giant

118
00:05:33,709 --> 00:05:36,089
vision model, and they noticed something weird.

119
00:05:36,589 --> 00:05:39,209
What was it? Well, as these models scaled up

120
00:05:39,209 --> 00:05:42,149
and ingested more data, the way they mathematically

121
00:05:42,149 --> 00:05:45,230
mapped out information internally, started to

122
00:05:45,230 --> 00:05:48,810
look eerily similar. Oh, interesting. Yeah. So

123
00:05:48,810 --> 00:05:51,550
if you plotted the way a vision model clustered

124
00:05:51,550 --> 00:05:54,329
images of dogs and plotted the way a language

125
00:05:54,329 --> 00:05:57,810
model clustered text descriptions of dogs, the

126
00:05:57,810 --> 00:06:00,310
shapes of those clusters were converging. So

127
00:06:00,310 --> 00:06:03,629
they were forming the same internal shapes? Exactly.

128
00:06:04,209 --> 00:06:06,310
He and his team hypothesized that these models

129
00:06:06,310 --> 00:06:09,870
were independently converging on a shared statistical

130
00:06:09,870 --> 00:06:13,149
model of reality. But critically, an observation

131
00:06:13,149 --> 00:06:17,509
is not a guarantee. Right. It's like noticing

132
00:06:17,509 --> 00:06:19,649
that two different trees are growing toward the

133
00:06:19,649 --> 00:06:21,730
same patch of sunlight. That's a great way to

134
00:06:21,730 --> 00:06:24,029
put it. It doesn't give you the genetic blueprint

135
00:06:24,029 --> 00:06:26,589
explaining why they must do so, nor does it give

136
00:06:26,589 --> 00:06:29,870
you a reliable recipe to force it to happen every

137
00:06:29,870 --> 00:06:33,250
time. It was empirical, not fundamental. Like

138
00:06:33,250 --> 00:06:35,810
seeing that every time we build a bigger telescope,

139
00:06:35,889 --> 00:06:38,250
these two different star clusters look a bit

140
00:06:38,250 --> 00:06:41,430
more similar, but lacking the actual physics

141
00:06:41,430 --> 00:06:43,550
equations to prove they are identical. Precisely.

142
00:06:43,819 --> 00:06:46,519
And what Balastrero and Lacoon do in this paper

143
00:06:46,519 --> 00:06:49,660
is provide those physics equations. They turn

144
00:06:49,660 --> 00:06:52,060
the Platonic representation hypothesis into the

145
00:06:52,060 --> 00:06:55,459
Platonic representation theorem. Meaning it's

146
00:06:55,459 --> 00:06:58,240
mathematically locked in. Yes. And the engine

147
00:06:58,240 --> 00:07:00,699
they use to prove this is a specific type of

148
00:07:00,699 --> 00:07:03,939
AI architecture called a JEPA. JEPA? Yeah. Okay.

149
00:07:04,000 --> 00:07:06,319
What does that stand for? It stands for Joint

150
00:07:06,319 --> 00:07:09,819
Embedding Predictive Architecture. Joint Embedding

151
00:07:09,819 --> 00:07:12,779
Predictive Architecture. Right. Huh. It is a

152
00:07:12,779 --> 00:07:15,279
highly specific framework within self -supervised

153
00:07:15,279 --> 00:07:17,740
learning. And unlike generative models which

154
00:07:17,740 --> 00:07:20,500
try to recreate the exact pixels of an image

155
00:07:20,500 --> 00:07:23,720
or predict the exact next word in a sentence,

156
00:07:24,279 --> 00:07:27,019
a JEPA works entirely in the abstract space of

157
00:07:27,019 --> 00:07:29,199
representations. So it's not guessing words or

158
00:07:29,199 --> 00:07:32,610
pixels? No. And to understand the mechanics of

159
00:07:32,610 --> 00:07:35,170
a JEPA, we have to look at the concept of positive

160
00:07:35,170 --> 00:07:38,850
pairs. In training, a JEPA is fed pairs of inputs

161
00:07:38,850 --> 00:07:41,750
that is explicitly told our views of the exact

162
00:07:41,750 --> 00:07:45,490
same underlying thing. OK, let me test an analogy

163
00:07:45,490 --> 00:07:48,389
here to ground this, because abstract representation

164
00:07:48,389 --> 00:07:51,610
space can get heavy quickly. Go for it. Imagine

165
00:07:51,610 --> 00:07:54,009
we have two detectives who are trying to reconstruct

166
00:07:54,009 --> 00:07:56,490
a bank robbery. OK. And they are interviewing

167
00:07:56,490 --> 00:07:59,310
two completely different witnesses. Witness A.

168
00:07:59,420 --> 00:08:01,939
was unfortunately locked in a closet during the

169
00:08:01,939 --> 00:08:04,060
robbery. Not a great place to be. No. It was

170
00:08:04,060 --> 00:08:06,500
pitch black, so they saw nothing. But they heard

171
00:08:06,500 --> 00:08:08,839
everything. They heard the exact sound of the

172
00:08:08,839 --> 00:08:11,459
vault mechanism, the heavy footsteps on the marble

173
00:08:11,459 --> 00:08:14,399
floor, the specific phrasing the robbers used

174
00:08:14,399 --> 00:08:17,379
when giving orders. OK, so that's our audio -based

175
00:08:17,379 --> 00:08:20,920
AI model. Exactly. Now, Witness B is completely

176
00:08:20,920 --> 00:08:23,500
deaf. But they were hiding on a balcony overlooking

177
00:08:23,500 --> 00:08:26,300
the lobby. So they saw the masks, the weapons,

178
00:08:26,540 --> 00:08:28,680
the precise layout of where the hostages were

179
00:08:28,680 --> 00:08:32,889
moved. And that's our vision model. Right. So

180
00:08:32,889 --> 00:08:35,730
a positive pair is the fact that both witnesses,

181
00:08:35,769 --> 00:08:38,129
despite experiencing completely different sensory

182
00:08:38,129 --> 00:08:41,210
data, are describing the exact same bank robbery.

183
00:08:41,850 --> 00:08:44,250
That analogy maps perfectly to the architecture.

184
00:08:44,629 --> 00:08:47,870
The robbery itself is the world state. The audio

185
00:08:47,870 --> 00:08:50,950
and the visual accounts are the modalities. OK.

186
00:08:51,149 --> 00:08:54,830
So the question is if these two detectives take

187
00:08:54,830 --> 00:08:57,190
these completely different formats of information.

188
00:08:57,649 --> 00:09:00,230
do they eventually draw the exact same floor

189
00:09:00,230 --> 00:09:02,769
plan of the crime scene? And the paper proves

190
00:09:02,769 --> 00:09:05,909
mathematically that the answer is yes. Wow. Yeah.

191
00:09:06,370 --> 00:09:09,450
In what they formalize as Theorem 1, the Platonic

192
00:09:09,450 --> 00:09:11,990
Representation Theorem, they demonstrate that

193
00:09:11,990 --> 00:09:14,450
if any number of JEPAs, whether they are processing

194
00:09:14,450 --> 00:09:17,690
text, images, or audio, are trained completely

195
00:09:17,690 --> 00:09:20,509
separately, but they are trained on positive

196
00:09:20,509 --> 00:09:22,950
pairs that relate to the exact same world state.

197
00:09:23,169 --> 00:09:26,730
They will inevitably learn the exact same mathematical

198
00:09:26,730 --> 00:09:29,289
embedding of the world. Wait, I need to push

199
00:09:29,289 --> 00:09:31,309
on the mechanics of that. Sure. Because it sounds

200
00:09:31,309 --> 00:09:34,210
almost like magic. I mean, an image is a dense

201
00:09:34,210 --> 00:09:36,570
grid of millions of colored pixels governed by

202
00:09:36,570 --> 00:09:39,649
spatial geometry, right? Right. And a text document

203
00:09:39,649 --> 00:09:41,549
is a linear sequence of letters, spaces, and

204
00:09:41,549 --> 00:09:44,610
punctuation governed by grammar. They are fundamentally

205
00:09:44,610 --> 00:09:47,259
different languages. So How can a mathematical

206
00:09:47,259 --> 00:09:49,580
formula guarantee that they arrive at the same

207
00:09:49,580 --> 00:09:52,159
internal map? It comes down to a mathematical

208
00:09:52,159 --> 00:09:55,419
concept the authors utilize called an injective

209
00:09:55,419 --> 00:09:58,879
rendering. Injective rendering. Right. An injective

210
00:09:58,879 --> 00:10:00,899
rendering essentially means that the process

211
00:10:00,899 --> 00:10:03,679
of capturing the world state, like taking a photograph

212
00:10:03,679 --> 00:10:06,860
of a room or writing a highly detailed descriptive

213
00:10:06,860 --> 00:10:10,299
caption of that same room, preserves the distinctness

214
00:10:10,299 --> 00:10:13,480
of the original reality. So it doesn't lose the

215
00:10:13,480 --> 00:10:16,759
core facts. Exactly. It doesn't scramble or destroy

216
00:10:16,759 --> 00:10:19,019
the core information. It merely translates it.

217
00:10:19,299 --> 00:10:21,759
The photograph and the caption are just different

218
00:10:21,759 --> 00:10:24,059
coordinate systems used to describe the exact

219
00:10:24,059 --> 00:10:27,299
same underlying structure. Oh, I see. So if I

220
00:10:27,299 --> 00:10:29,879
have a physical cube on my desk, I can describe

221
00:10:29,879 --> 00:10:32,600
its location using Cartesian coordinates, like

222
00:10:32,600 --> 00:10:35,399
x, y, and z axes. Or I can describe it using

223
00:10:35,399 --> 00:10:37,419
polar coordinates, using angles and distances

224
00:10:37,419 --> 00:10:40,299
from a center point. Perfect example. The numbers

225
00:10:40,299 --> 00:10:42,220
in my notebook will look completely different,

226
00:10:42,379 --> 00:10:45,340
but the cube hasn't moved an inch. The reality

227
00:10:45,340 --> 00:10:48,559
is static, only the coordinates changed. Precisely.

228
00:10:49,039 --> 00:10:51,139
And this is where the paper dives deep into the

229
00:10:51,139 --> 00:10:53,720
underlying mechanics, specifically in Proposition

230
00:10:53,720 --> 00:10:57,679
1, which deals with whitened embeddings and the

231
00:10:57,679 --> 00:11:00,320
spectral gap. Okay, let's break that down without

232
00:11:00,320 --> 00:11:02,980
needing a whiteboard. Let's do that. So when

233
00:11:02,980 --> 00:11:05,759
an AI model processes data, it maps it into a

234
00:11:05,759 --> 00:11:08,299
high dimensional space. The goal of a JEPA is

235
00:11:08,299 --> 00:11:10,700
to make the representations of positive pairs,

236
00:11:11,019 --> 00:11:13,500
the two views of the same thing, as close together

237
00:11:13,500 --> 00:11:15,960
in that space as possible. OK, but there's a

238
00:11:15,960 --> 00:11:18,659
danger here, right? Huge danger. What stops the

239
00:11:18,659 --> 00:11:21,539
model from just cheating? Like, if its only job

240
00:11:21,539 --> 00:11:24,940
is to make the math for two inputs match, why

241
00:11:24,940 --> 00:11:27,240
doesn't it just map every single image and every

242
00:11:27,240 --> 00:11:28,779
single sentence in the universe to the number

243
00:11:28,779 --> 00:11:31,470
zero? Yeah. just map everything to zero. Then

244
00:11:31,470 --> 00:11:33,750
they all match perfectly, and the model technically

245
00:11:33,750 --> 00:11:37,029
did its job with zero effort. You've just described

246
00:11:37,029 --> 00:11:39,230
what machine learning researchers call collapse.

247
00:11:39,730 --> 00:11:41,590
It's a very real problem where the model takes

248
00:11:41,590 --> 00:11:43,850
the lazy way out and learns absolutely nothing.

249
00:11:44,129 --> 00:11:47,309
Right. So how do they stop that? To prevent this,

250
00:11:47,509 --> 00:11:49,789
JEPAs use something called second order anti

251
00:11:49,789 --> 00:11:53,629
-collapse. Second order anti -collapse. Basically,

252
00:11:53,669 --> 00:11:56,330
it's a mathematical constraint forcing the model's

253
00:11:56,330 --> 00:11:58,850
representations to spread out. It demands that

254
00:11:58,850 --> 00:12:01,610
the dimensions of the embedding space vary independently

255
00:12:01,610 --> 00:12:04,490
and carry distinct information. So it forces

256
00:12:04,490 --> 00:12:07,610
it to use all its brain power. Exactly. When

257
00:12:07,610 --> 00:12:09,649
you apply this anti -collapse mechanism, the

258
00:12:09,649 --> 00:12:11,870
model is forced to actually map the structure

259
00:12:11,870 --> 00:12:14,809
of the data. It has to do the homework. Exactly.

260
00:12:15,190 --> 00:12:17,750
We can't just write zero for every answer. Exactly.

261
00:12:18,009 --> 00:12:20,470
And the paper proves that if the data has a clear

262
00:12:20,470 --> 00:12:22,929
underlying structure, what they term a spectral

263
00:12:22,929 --> 00:12:25,549
gap, meaning the true concepts are mathematically

264
00:12:25,549 --> 00:12:28,250
distinct from random noise, there is only one

265
00:12:28,250 --> 00:12:30,730
unique mathematical solution for how to map that

266
00:12:30,730 --> 00:12:33,149
data while satisfying both the positive pairs

267
00:12:33,149 --> 00:12:35,669
matching and the anti -collapse constraint. Wait,

268
00:12:36,090 --> 00:12:40,029
only one solution? only one valid mathematical

269
00:12:40,029 --> 00:12:42,570
solution to perfectly describe that world state.

270
00:12:43,009 --> 00:12:45,350
So both the text model and the vision model,

271
00:12:45,389 --> 00:12:47,570
even though they are trained in total isolation

272
00:12:47,570 --> 00:12:50,309
from each other in their respective caves, will

273
00:12:50,309 --> 00:12:52,970
inevitably land on that exact same solution.

274
00:12:53,509 --> 00:12:56,769
Because math is universal. They are solving the

275
00:12:56,769 --> 00:12:58,750
same complex equation from different starting

276
00:12:58,750 --> 00:13:01,610
points. And there is only one correct answer.

277
00:13:01,830 --> 00:13:04,490
Right. The paper has a very specific caveat.

278
00:13:04,889 --> 00:13:08,000
It says they will arrive at the identical mathematical

279
00:13:08,000 --> 00:13:11,779
representation up to a single geometric rotation.

280
00:13:12,240 --> 00:13:15,460
Right, up to a single rotation. What does that

281
00:13:15,460 --> 00:13:17,879
actually mean in practical terms? Let's return

282
00:13:17,879 --> 00:13:19,639
to your two detectives and their floor plans.

283
00:13:19,919 --> 00:13:22,379
Let's assume both detectives are incredibly thorough.

284
00:13:22,899 --> 00:13:25,659
The layout, the distances between the teller

285
00:13:25,659 --> 00:13:28,419
desks, the location of the vault, they are 100

286
00:13:28,419 --> 00:13:31,519
% identical on both floor plans. Perfect copies

287
00:13:31,519 --> 00:13:35,529
of reality. Exactly. But detective A, the audio

288
00:13:35,529 --> 00:13:37,850
detective, happened to draw his map on a piece

289
00:13:37,850 --> 00:13:41,309
of paper facing north. Detective B, the visual

290
00:13:41,309 --> 00:13:44,549
detective, drew his map facing east. Oh, I see.

291
00:13:45,090 --> 00:13:47,389
So if you place the two pieces of paper next

292
00:13:47,389 --> 00:13:50,490
to each other, they look different. But the underlying

293
00:13:50,490 --> 00:13:53,649
geometry is identical. If you just apply a single

294
00:13:53,649 --> 00:13:56,909
geometric rotation, just spinning Detective B's

295
00:13:56,909 --> 00:14:00,009
paper 90 degrees, the maps overlay perfectly.

296
00:14:00,210 --> 00:14:03,309
The desks align, the doors align, the vault aligns.

297
00:14:03,450 --> 00:14:07,509
That is a staggering realization. I mean, it

298
00:14:07,509 --> 00:14:10,769
means that true alignment between vision AI and

299
00:14:10,769 --> 00:14:13,870
language AI isn't some artificial bridge we have

300
00:14:13,870 --> 00:14:15,990
to forcefully engineer by mashing them together

301
00:14:15,990 --> 00:14:19,129
with massive amounts of compute. No, it's a natural

302
00:14:19,129 --> 00:14:21,330
mathematical consequence of the geometry of the

303
00:14:21,330 --> 00:14:23,850
universe, provided the models are looking at

304
00:14:23,850 --> 00:14:26,210
the exact same world states. Okay, but if we

305
00:14:26,210 --> 00:14:28,870
accept this theorem is mathematically sound...

306
00:14:28,879 --> 00:14:31,480
It opens up a massive contradiction in the real

307
00:14:31,480 --> 00:14:33,639
world. It does. Because if theorem one proves

308
00:14:33,639 --> 00:14:36,320
perfect alignment is a guaranteed mathematical

309
00:14:36,320 --> 00:14:39,149
consequence, why are we struggling so hard to

310
00:14:39,149 --> 00:14:41,330
achieve it? That is the million dollar question.

311
00:14:41,590 --> 00:14:43,690
Like, why does state -of -the -art vision and

312
00:14:43,690 --> 00:14:46,450
language models today still suffer from hallucinations,

313
00:14:47,169 --> 00:14:49,750
misunderstandings, and just a fundamental inability

314
00:14:49,750 --> 00:14:53,129
to flawlessly map words to images without requiring

315
00:14:53,129 --> 00:14:55,669
trillions of parameters to brute force the connection?

316
00:14:55,850 --> 00:14:58,769
Right. If the math dictates. that they should

317
00:14:58,769 --> 00:15:01,750
naturally align and reality shows us they are

318
00:15:01,750 --> 00:15:04,230
struggling to align, then something in the way

319
00:15:04,230 --> 00:15:06,870
we are building them is fundamentally disrupting

320
00:15:06,870 --> 00:15:09,549
that math. The ingredients we are using must

321
00:15:09,549 --> 00:15:11,950
be tainted. And that leads directly into the

322
00:15:11,950 --> 00:15:14,190
most critical distinction in the entire paper.

323
00:15:14,769 --> 00:15:18,389
The viewpoint matters immensely. Okay. Ballastriero

324
00:15:18,389 --> 00:15:20,649
and Lacun diagnosed the problem by categorizing

325
00:15:20,649 --> 00:15:23,409
the training data itself. They distinguish between

326
00:15:23,409 --> 00:15:26,149
what they call good views and bad views. Good

327
00:15:26,149 --> 00:15:29,000
views and bad views. Yes. And this single distinction

328
00:15:29,000 --> 00:15:32,200
explains why modern AI is so incredibly inefficient.

329
00:15:32,519 --> 00:15:34,639
OK, let's unpack the mechanics of these views

330
00:15:34,639 --> 00:15:37,299
then. This seems to be the fulcrum of their entire

331
00:15:37,299 --> 00:15:40,100
argument. It is. According to the paper, a good

332
00:15:40,100 --> 00:15:42,759
view is defined as providing the model with two

333
00:15:42,759 --> 00:15:45,320
completely different observations of the same

334
00:15:45,320 --> 00:15:47,580
scene. Let's ground that in a tangible example.

335
00:15:47,759 --> 00:15:50,240
A good view for an image model would be providing

336
00:15:50,240 --> 00:15:52,659
it with two completely distinct photographs of

337
00:15:52,659 --> 00:15:55,139
the exact same living room taken by two different

338
00:15:55,139 --> 00:15:57,820
cameras at slightly different angles. Okay. For

339
00:15:57,820 --> 00:16:00,320
a text model, a good view would be providing

340
00:16:00,320 --> 00:16:02,399
it with two completely different paragraphs written

341
00:16:02,399 --> 00:16:04,879
by two different human beings, both describing

342
00:16:04,879 --> 00:16:08,000
that exact same living room. So they are completely

343
00:16:08,000 --> 00:16:10,080
different renderings, different pixels, different

344
00:16:10,080 --> 00:16:12,720
words, but they are anchored to the same underlying

345
00:16:12,720 --> 00:16:15,799
reality. Right. The positive pair is the reality

346
00:16:15,799 --> 00:16:18,080
of the room, not the medium used to capture it.

347
00:16:18,460 --> 00:16:20,539
But every medium inherently carries baggage.

348
00:16:20,779 --> 00:16:23,460
Ah, yes. And the authors formalized this in Theorem

349
00:16:23,460 --> 00:16:27,299
3, introducing the concept of nuisance. Nuisance?

350
00:16:28,019 --> 00:16:30,379
They state mathematically that every observation,

351
00:16:30,679 --> 00:16:32,659
every single piece of data we feed into a model

352
00:16:32,659 --> 00:16:36,059
has a nuisance specific to its modality. What

353
00:16:36,059 --> 00:16:39,009
is nuisance exactly? Nuisance is the information

354
00:16:39,009 --> 00:16:41,269
in the rendering that belongs entirely to the

355
00:16:41,269 --> 00:16:43,129
medium and has absolutely nothing to do with

356
00:16:43,129 --> 00:16:45,690
the world state itself. So if I take a photograph

357
00:16:45,690 --> 00:16:47,830
of that living room, the nuisance would be things

358
00:16:47,830 --> 00:16:50,789
like the specific lens distortion of my camera,

359
00:16:51,009 --> 00:16:53,169
right? Yes, or a speck of dust on the sensor.

360
00:16:53,629 --> 00:16:55,970
Or the exact angle I was holding my phone, or

361
00:16:55,970 --> 00:16:58,769
the specific shadow cast by a cloud passing outside

362
00:16:58,769 --> 00:17:01,870
of that exact millisecond. None of those things

363
00:17:01,870 --> 00:17:04,250
define what a living room actually is. Right.

364
00:17:04,619 --> 00:17:06,839
They only define how the living room was captured

365
00:17:06,839 --> 00:17:10,220
in that specific instance. That is a perfect

366
00:17:10,220 --> 00:17:13,059
explanation of visual nuisance. And text has

367
00:17:13,059 --> 00:17:15,920
its own distinct nuisance, I'd imagine. Oh, definitely.

368
00:17:16,339 --> 00:17:18,920
The specific grammatical structures chosen, a

369
00:17:18,920 --> 00:17:22,680
writer's idiosyncratic vocabulary, a typo, the

370
00:17:22,680 --> 00:17:24,720
decision to use a passive voice instead of an

371
00:17:24,720 --> 00:17:28,720
active voice. OK, let's say I write, the crimson

372
00:17:28,720 --> 00:17:31,940
sphere rests upon the mahogany table. And you

373
00:17:31,940 --> 00:17:34,660
write, there's a red ball on a brown desk. We

374
00:17:34,660 --> 00:17:37,579
are describing the exact same world state. But

375
00:17:37,579 --> 00:17:39,779
my pretentious vocabulary in your straightforward

376
00:17:39,779 --> 00:17:43,299
syntax are the text's nuisance. They are artifacts

377
00:17:43,299 --> 00:17:45,740
of us, the authors, not the ball or the desk.

378
00:17:46,099 --> 00:17:48,940
Exactly. And what's fascinating here is the mathematical

379
00:17:48,940 --> 00:17:51,480
implication of this nuisance. The authors prove

380
00:17:51,480 --> 00:17:54,380
that no rotation in a mathematical space can

381
00:17:54,380 --> 00:17:57,039
ever align the nuisance of one modality with

382
00:17:57,039 --> 00:17:59,859
the nuisance of another. Really? Yeah. They are

383
00:17:59,859 --> 00:18:02,799
entirely separate orthogonal dimensions of reality.

384
00:18:03,210 --> 00:18:05,990
You cannot mathematically rotate the concept

385
00:18:05,990 --> 00:18:08,769
of pretentious vocabulary to match the concept

386
00:18:08,769 --> 00:18:12,349
of lens flare. They simply do not map onto each

387
00:18:12,349 --> 00:18:15,910
other. Which makes perfect sense. I mean, going

388
00:18:15,910 --> 00:18:19,190
back to the detectives. My detective, who is

389
00:18:19,190 --> 00:18:21,410
listening to the bank robbery, might be distracted

390
00:18:21,410 --> 00:18:24,569
by a ringing phone nearby. Right. And the detective

391
00:18:24,569 --> 00:18:26,250
watching from the balcony might be distracted

392
00:18:26,250 --> 00:18:30,089
by a smudge on the glass. You can't align a ringing

393
00:18:30,089 --> 00:18:32,710
phone with a glass smudge. They are totally irrelevant

394
00:18:32,710 --> 00:18:36,250
to the floor plan of the bank. Exactly. So if

395
00:18:36,250 --> 00:18:38,210
good views are different observations of the

396
00:18:38,210 --> 00:18:40,329
same scene that allow the model to ignore the

397
00:18:40,329 --> 00:18:42,950
nuisance and focus on the floor plan... What

398
00:18:42,950 --> 00:18:46,589
exactly constitutes a bad view? A bad view is

399
00:18:46,589 --> 00:18:48,890
defined as a corruption of a single observation.

400
00:18:49,170 --> 00:18:52,130
A corruption of a single observation. What does

401
00:18:52,130 --> 00:18:54,529
that look like? Instead of giving the model two

402
00:18:54,529 --> 00:18:56,890
different photographs of the room, you take one

403
00:18:56,890 --> 00:18:59,289
single photograph and you mathematically alter

404
00:18:59,289 --> 00:19:02,279
it. Like photoshopping it. Kind of. You crop

405
00:19:02,279 --> 00:19:04,839
it twice. So the model is comparing the top half

406
00:19:04,839 --> 00:19:06,519
of the photo to the bottom half of the photo.

407
00:19:06,960 --> 00:19:09,500
Or in the realm of text, you take one single

408
00:19:09,500 --> 00:19:12,200
sentence and you mask out different words. Oh,

409
00:19:12,559 --> 00:19:14,700
you mean like taking my sentence and creating

410
00:19:14,700 --> 00:19:18,039
one view that says the blank sphere rests upon

411
00:19:18,039 --> 00:19:20,599
the mahogany table and pairing it with a second

412
00:19:20,599 --> 00:19:23,019
view that says the crimson sphere rests upon

413
00:19:23,019 --> 00:19:27,160
the blank table. Yes, that is. a bad view. You

414
00:19:27,160 --> 00:19:29,539
are forcing the model to compare a single observation

415
00:19:29,539 --> 00:19:32,140
against a corrupted version of itself. OK, I

416
00:19:32,140 --> 00:19:33,839
have to pause here because this feels like a

417
00:19:33,839 --> 00:19:36,380
direct attack on the entire modern machine learning

418
00:19:36,380 --> 00:19:39,880
industry. It kind of is. I mean, masking text

419
00:19:39,880 --> 00:19:41,960
where you hide a word and force the model to

420
00:19:41,960 --> 00:19:44,740
predict it and contrast of cropping of images

421
00:19:44,740 --> 00:19:47,380
where you take a picture of a dog, crop the ear

422
00:19:47,380 --> 00:19:49,619
and crop the tail and tell the model they are

423
00:19:49,619 --> 00:19:52,920
the same dog. Isn't that exactly how almost every

424
00:19:52,920 --> 00:19:56,009
single major AI model is trained today? Absolutely

425
00:19:56,009 --> 00:19:58,630
is. That is the foundational architecture of

426
00:19:58,630 --> 00:20:01,569
models like BERT and modern masked autoencoders.

427
00:20:01,829 --> 00:20:03,450
Yet the vast majority of the industry relies

428
00:20:03,450 --> 00:20:05,529
heavily on corruptions of single observations.

429
00:20:06,150 --> 00:20:08,849
And Proposition 3 in this paper explains exactly

430
00:20:08,849 --> 00:20:11,250
why doing so permanently cripples the model's

431
00:20:11,250 --> 00:20:13,670
ability to achieve perfect alignment. The authors

432
00:20:13,670 --> 00:20:17,170
prove mathematically that if an AI uses bad views,

433
00:20:17,349 --> 00:20:19,849
it is forced to memorize the nuisance. Okay,

434
00:20:20,109 --> 00:20:22,380
break down the mechanics of that. Why does cropping

435
00:20:22,380 --> 00:20:25,420
an image force the model to memorize the camera

436
00:20:25,420 --> 00:20:28,039
lens flare? Because the AI is an optimization

437
00:20:28,039 --> 00:20:30,480
machine. When you give it a top half -crop and

438
00:20:30,480 --> 00:20:32,539
a bottom half -crop of the exact same photograph,

439
00:20:33,000 --> 00:20:35,059
the easiest mathematical way for the model to

440
00:20:35,059 --> 00:20:38,039
link those two images is not to deduce the complex

441
00:20:38,039 --> 00:20:40,599
3D geometry of the living room. Right, that would

442
00:20:40,599 --> 00:20:43,559
take actual reasoning. Exactly. The easiest way

443
00:20:43,559 --> 00:20:45,920
is to notice the consistent pixel noise, the

444
00:20:45,920 --> 00:20:48,559
exact lighting gradient, or the specific compression

445
00:20:48,559 --> 00:20:51,140
artifacts that exist uniformly across that single

446
00:20:51,140 --> 00:20:54,390
digital file. Oh wow. It latches onto the nuisance,

447
00:20:54,589 --> 00:20:56,869
because the nuisance is the strongest, most immediate

448
00:20:56,869 --> 00:20:59,250
mathematical link between the two halves. So

449
00:20:59,250 --> 00:21:02,589
instead of learning what a room is, it just learns

450
00:21:02,589 --> 00:21:05,690
the specific digital signature of that one JPEG

451
00:21:05,690 --> 00:21:08,470
file. Yep. And what happens to the AI's internal

452
00:21:08,470 --> 00:21:11,109
map when it does that? It permanently wastes

453
00:21:11,109 --> 00:21:14,390
its capacity. How so? Well, an AI model has a

454
00:21:14,390 --> 00:21:16,509
finite number of dimensions in its embedding

455
00:21:16,509 --> 00:21:18,910
space. Think of them as distinct cognitive slots

456
00:21:18,910 --> 00:21:21,650
it can use to store concepts. If it is trained

457
00:21:21,650 --> 00:21:24,309
on bad views, it is mathematically forced to

458
00:21:24,309 --> 00:21:26,470
spend a massive percentage of those precious

459
00:21:26,470 --> 00:21:29,190
dimensions memorizing camera angles, lighting

460
00:21:29,190 --> 00:21:31,529
glitches, and specific grammatical quirks. It's

461
00:21:31,529 --> 00:21:34,430
just filling up its brain with garbage. Precisely.

462
00:21:34,789 --> 00:21:37,150
And because we established earlier that visual

463
00:21:37,150 --> 00:21:39,809
nuisance cannot be rotated to align with textual

464
00:21:39,809 --> 00:21:42,730
nuisance, those wasted dimensions can never be

465
00:21:42,730 --> 00:21:45,170
aligned across modalities. Right. You can't line

466
00:21:45,170 --> 00:21:48,890
a typo with a pixel error. Right. By using bad

467
00:21:48,890 --> 00:21:51,150
views, we are building models that permanently

468
00:21:51,150 --> 00:21:53,789
encode irrelevant artifacts into their core understanding

469
00:21:53,789 --> 00:21:56,089
of the world. This places a hard mathematical

470
00:21:56,089 --> 00:21:58,369
ceiling on how well a vision model and a language

471
00:21:58,369 --> 00:22:01,150
model can ever truly understand each other. Wow.

472
00:22:01,750 --> 00:22:04,950
Okay, so if bad views inherently force a model

473
00:22:04,950 --> 00:22:07,369
to memorize nuisance and completely ruin the

474
00:22:07,369 --> 00:22:09,549
possibility of a perfect platonic representation,

475
00:22:10,049 --> 00:22:12,930
we have to turn this lens onto the absolute dominant

476
00:22:12,930 --> 00:22:15,430
force in AI today. We do. We have to talk about

477
00:22:15,430 --> 00:22:18,509
next token prediction. We really do. Because

478
00:22:18,509 --> 00:22:21,150
next token prediction isn't just a popular technique.

479
00:22:21,250 --> 00:22:24,710
It is the fundamental engine driving the entire

480
00:22:24,710 --> 00:22:29,049
generative AI boom. JetGBT, Claude, Llama, Gemini.

481
00:22:29,430 --> 00:22:31,730
All of them. All of these massive large language

482
00:22:31,730 --> 00:22:34,750
models are, at their core, built on a surprisingly

483
00:22:34,750 --> 00:22:37,150
simple objective. Given a sequence of words,

484
00:22:37,329 --> 00:22:40,549
predict the very next word. And the authors target

485
00:22:40,549 --> 00:22:43,470
this specifically in Proposition 4. And they

486
00:22:43,470 --> 00:22:46,150
make a rather shocking claim. Very shocking.

487
00:22:46,609 --> 00:22:48,809
They state that next -token prediction isn't

488
00:22:48,809 --> 00:22:51,170
some entirely different paradigm. Mathematically,

489
00:22:51,769 --> 00:22:54,369
it is actually just a JEPA in disguise. It is.

490
00:22:54,890 --> 00:22:57,450
But how does that work? How is guessing the next

491
00:22:57,450 --> 00:23:00,210
word in a sentence the same architectural mechanism

492
00:23:00,210 --> 00:23:02,549
as comparing two views of a scene? It requires

493
00:23:02,549 --> 00:23:04,710
looking at the linear algebra operating under

494
00:23:04,710 --> 00:23:07,190
the hood. When you break down the math of next

495
00:23:07,190 --> 00:23:09,230
token predictions, specifically looking at models

496
00:23:09,230 --> 00:23:12,230
with a linear softmax head, the authors demonstrate

497
00:23:12,230 --> 00:23:15,369
it maps perfectly to a JEPA architecture. Okay,

498
00:23:15,390 --> 00:23:17,990
lay that out for me. So the context encoder of

499
00:23:17,990 --> 00:23:20,410
the JEPA processes the prefix, which is the chunk

500
00:23:20,410 --> 00:23:23,170
of text leading up to the blank. Right. The target

501
00:23:23,170 --> 00:23:27,150
encoder processes the actual next word. The model

502
00:23:27,150 --> 00:23:28,869
then predicts the inner product between those

503
00:23:28,869 --> 00:23:30,769
two embeddings. It is fundamentally treating

504
00:23:30,769 --> 00:23:33,069
the prefix and the continuation as a positive

505
00:23:33,069 --> 00:23:35,900
pair. Wait, let me make sure I am following the

506
00:23:35,900 --> 00:23:38,420
logic here. Okay. If the underlying math proves

507
00:23:38,420 --> 00:23:41,259
that next -token prediction is just a JEPA comparing

508
00:23:41,259 --> 00:23:44,720
a prefix and a continuation, and a prefix and

509
00:23:44,720 --> 00:23:47,319
a continuation are by definition always located

510
00:23:47,319 --> 00:23:49,220
right next to each other within a single document.

511
00:23:49,660 --> 00:23:52,900
Which means? Which means next -token prediction

512
00:23:52,900 --> 00:23:57,140
is fundamentally a bad view. Yes. It is the ultimate

513
00:23:57,140 --> 00:23:59,880
bad view. It operates entirely within the isolated

514
00:23:59,880 --> 00:24:02,099
vacuum of a single observation. You have hit

515
00:24:02,099 --> 00:24:04,299
the exact core of their argument. By definition,

516
00:24:04,660 --> 00:24:07,220
a prefix of a document and the immediate continuation

517
00:24:07,220 --> 00:24:09,839
of that document share the exact same nuisance.

518
00:24:09,920 --> 00:24:12,720
Wow. They show the same authors. It used syncretic

519
00:24:12,720 --> 00:24:15,079
writing style. They pull from the exact same

520
00:24:15,079 --> 00:24:17,559
limited vocabulary pool. They share the same

521
00:24:17,559 --> 00:24:20,460
formatting, the same tone, the same biases. So

522
00:24:20,460 --> 00:24:23,460
when we train a trillion parameter model to predict

523
00:24:23,460 --> 00:24:26,819
the next token, we aren't actually teaching it

524
00:24:26,819 --> 00:24:29,259
the platonic reality of the concepts it's reading

525
00:24:29,259 --> 00:24:33,359
about. We are forcing it to memorize the stylistic

526
00:24:33,359 --> 00:24:35,480
nuisance of every single author on the internet.

527
00:24:35,880 --> 00:24:38,700
We are forcing it to memorize that when a specific

528
00:24:38,700 --> 00:24:41,859
Reddit user uses a certain colloquialism, they

529
00:24:41,859 --> 00:24:43,960
usually follow it up with a specific punctuation

530
00:24:43,960 --> 00:24:46,859
mark. Exactly. And if we connect this to the

531
00:24:46,859 --> 00:24:49,319
bigger picture of the tech industry, this explains

532
00:24:49,319 --> 00:24:51,880
a mystery that has baffled researchers for years.

533
00:24:52,059 --> 00:24:55,079
Which is what? Why do large language models require

534
00:24:55,079 --> 00:24:58,380
such unfathomably massive amounts of data to

535
00:24:58,380 --> 00:25:00,890
exhibit general intelligence? We are feeding

536
00:25:00,890 --> 00:25:03,589
them the entire digitized knowledge of humanity

537
00:25:03,589 --> 00:25:06,349
over and over again. And the reason we have to

538
00:25:06,349 --> 00:25:08,630
do that is because next token prediction models

539
00:25:08,630 --> 00:25:11,509
are only figuring out the true world state accidentally.

540
00:25:11,670 --> 00:25:13,390
What do you mean accidentally? Because their

541
00:25:13,390 --> 00:25:15,710
primary mathematical objective is to memorize

542
00:25:15,710 --> 00:25:18,029
the local nuisance of individual documents. Right.

543
00:25:18,369 --> 00:25:21,210
The only way they extract genuine universal concepts

544
00:25:21,210 --> 00:25:24,869
is by finding incredibly weak, dense statistical

545
00:25:24,869 --> 00:25:27,630
intersections across millions of disparate documents.

546
00:25:27,910 --> 00:25:30,900
Give me an example of that. So the model reads

547
00:25:30,900 --> 00:25:33,019
10 ,000 completely different articles written

548
00:25:33,019 --> 00:25:36,299
by 10 ,000 different authors with 10 ,000 different

549
00:25:36,299 --> 00:25:38,779
writing styles. The model is forced to memorize

550
00:25:38,779 --> 00:25:41,599
all 10 ,000 distinct styles. That's a lot of

551
00:25:41,599 --> 00:25:44,920
nuisance. It is. But deeply buried in all that

552
00:25:44,920 --> 00:25:47,460
nuisance, it statistically averages out the fact

553
00:25:47,460 --> 00:25:50,240
that across all those different styles, the word

554
00:25:50,240 --> 00:25:53,079
apple frequently co -occurs near concepts of

555
00:25:53,079 --> 00:25:56,740
red, fruit, and tree. So it has to computationally

556
00:25:56,740 --> 00:26:00,650
sift through mountains of grammatical garbage

557
00:26:00,650 --> 00:26:04,309
just to glean tiny accidental fragments of the

558
00:26:04,309 --> 00:26:07,329
true shape of reality. Exactly. That implies

559
00:26:07,329 --> 00:26:09,390
that the giants of the AI industry are taking

560
00:26:09,390 --> 00:26:12,789
the absolute longest most computationally exhausting

561
00:26:12,789 --> 00:26:14,890
and mathematically inefficient route possible

562
00:26:14,890 --> 00:26:16,859
to understanding the world. They really are.

563
00:26:17,160 --> 00:26:19,240
They're brute forcing alignment through sheer

564
00:26:19,240 --> 00:26:21,759
overwhelming scale because their fundamental

565
00:26:21,759 --> 00:26:23,920
mechanism, predicting the next token in a single

566
00:26:23,920 --> 00:26:26,420
document, is mathematically flawed for the purpose

567
00:26:26,420 --> 00:26:29,799
of learning clean representations. It is a remarkable

568
00:26:29,799 --> 00:26:32,519
inefficiency. Yeah, the models are learning in

569
00:26:32,519 --> 00:26:34,480
spite of their objective function, not because

570
00:26:34,480 --> 00:26:37,660
of it. Yeah, they are using brute force to overcome

571
00:26:37,660 --> 00:26:40,200
a mathematical handicap we artificially imposed

572
00:26:40,200 --> 00:26:43,759
on them. Okay, this is a brilliant theoretical

573
00:26:43,759 --> 00:26:46,140
teardown. The math points to a massive flaw in

574
00:26:46,140 --> 00:26:49,180
how we build AI, but theory doesn't run data

575
00:26:49,180 --> 00:26:52,279
centers. No, it doesn't. We need empirical evidence.

576
00:26:52,599 --> 00:26:55,259
We need to see if this platonic representation

577
00:26:55,259 --> 00:26:57,880
theorem actually holds up when you compile the

578
00:26:57,880 --> 00:27:00,420
code and run it on real -world data. Does training

579
00:27:00,420 --> 00:27:02,539
on good views actually crush the standard bad

580
00:27:02,539 --> 00:27:04,980
views in next -token paradigms in practice? And

581
00:27:04,980 --> 00:27:07,339
this is where the authors transition from theorists

582
00:27:07,339 --> 00:27:10,230
to experimentalists. They designed an extensive

583
00:27:10,230 --> 00:27:12,789
suite of tests to prove this out, starting with

584
00:27:12,789 --> 00:27:15,049
a widely used standard data set called Flickr

585
00:27:15,049 --> 00:27:17,950
30K. OK, let's establish the parameters of this

586
00:27:17,950 --> 00:27:21,789
test. How do you design an experiment to definitively

587
00:27:21,789 --> 00:27:24,529
prove that one method of training builds a better

588
00:27:24,529 --> 00:27:28,009
internal map of reality than another? First,

589
00:27:28,089 --> 00:27:30,549
you need complete isolation to prevent the models

590
00:27:30,549 --> 00:27:33,529
from cheating. Right. So they utilize two separate

591
00:27:33,529 --> 00:27:36,339
models. For the vision side, they used a vision

592
00:27:36,339 --> 00:27:41,140
transformer, specifically a VIT -S16. This is

593
00:27:41,140 --> 00:27:43,839
a compact, highly efficient architecture designed

594
00:27:43,839 --> 00:27:47,619
to chop images into a 16 by 16 grid of patches,

595
00:27:48,339 --> 00:27:50,559
analyzing the geometric relationships between

596
00:27:50,559 --> 00:27:53,839
those patches. OK. And for the text? For the

597
00:27:53,839 --> 00:27:56,720
text side, they used a 33 million parameter model

598
00:27:56,720 --> 00:28:00,380
based on a standard BERT architecture, which

599
00:28:00,380 --> 00:28:02,980
processes language bidirectionally to understand

600
00:28:02,980 --> 00:28:06,140
context. Crucially, they train these two models

601
00:28:06,140 --> 00:28:08,700
completely independently. Completely isolated.

602
00:28:08,900 --> 00:28:11,539
Yes. The vision transformer was only ever fed

603
00:28:11,539 --> 00:28:14,740
pixels. The BERT text model was only ever fed

604
00:28:14,740 --> 00:28:17,460
words. Total sensory deprivation from each other.

605
00:28:17,839 --> 00:28:19,960
The text model has absolutely no idea what an

606
00:28:19,960 --> 00:28:22,339
image is. And the vision model has never encountered

607
00:28:22,339 --> 00:28:25,500
language. Exactly. Then to test their theorem,

608
00:28:25,740 --> 00:28:28,059
they created three different variations of the

609
00:28:28,059 --> 00:28:30,380
text model. They kept the architecture identical.

610
00:28:30,619 --> 00:28:33,180
the only variable they changed was the definition

611
00:28:33,180 --> 00:28:36,920
of the positive pair used during training. Setting

612
00:28:36,920 --> 00:28:39,900
up the showdown between the views. Exactly. First,

613
00:28:39,980 --> 00:28:42,980
they trained a good view text JEPA. Now, the

614
00:28:42,980 --> 00:28:46,160
Flickr 30K data set is unique because every image

615
00:28:46,160 --> 00:28:49,119
in it has five distinct captions written by five

616
00:28:49,119 --> 00:28:51,359
different human annotators. Oh, that's perfect.

617
00:28:51,480 --> 00:28:54,759
It is. So for the good view model, the positive

618
00:28:54,759 --> 00:28:57,619
pair was two completely different human captions

619
00:28:57,619 --> 00:29:00,799
describing the exact same underlying image. So

620
00:29:00,799 --> 00:29:04,319
pairing a sentence like a brown dog leaps through

621
00:29:04,319 --> 00:29:06,859
the crashing waves with a completely different

622
00:29:06,859 --> 00:29:09,500
sentence like a golden retriever playing in the

623
00:29:09,500 --> 00:29:12,140
ocean surf. Exactly. Two distinct observations,

624
00:29:12,740 --> 00:29:15,279
different vocabularies, but anchored to the exact

625
00:29:15,279 --> 00:29:19,539
same world state. A true good view. Second, they

626
00:29:19,539 --> 00:29:22,279
trained a bad view text jepa. They abandoned

627
00:29:22,279 --> 00:29:24,640
the multiple captions, they took a single caption,

628
00:29:24,740 --> 00:29:26,720
and they simply masked out different spans of

629
00:29:26,720 --> 00:29:28,960
words. Okay. The model was forced to compare

630
00:29:28,960 --> 00:29:31,200
a corrupted version of a sentence against another

631
00:29:31,200 --> 00:29:33,339
corrupted version of that exact same sentence.

632
00:29:33,559 --> 00:29:35,539
Which has been making the standard industry practice

633
00:29:35,539 --> 00:29:38,279
we discussed earlier. And third, they trained

634
00:29:38,279 --> 00:29:40,859
a standard next token prediction model on the

635
00:29:40,859 --> 00:29:43,819
exact same set of captions using the classic

636
00:29:43,819 --> 00:29:46,400
prefix and continuation method that powers models

637
00:29:46,400 --> 00:29:48,980
like chat GPT. So the playing field is completely

638
00:29:48,980 --> 00:29:52,269
level. Same raw data, same parameter count. Identical

639
00:29:52,269 --> 00:29:54,349
compute limits. Exactly. The only difference

640
00:29:54,349 --> 00:29:57,069
is the mathematical objective function dictating

641
00:29:57,069 --> 00:29:59,990
how they process the text. So how did they evaluate

642
00:29:59,990 --> 00:30:02,029
which model built the better map of reality?

643
00:30:02,509 --> 00:30:05,269
They used a zero -shot retrieval test to measure

644
00:30:05,269 --> 00:30:07,569
alignment. Because these models had never seen

645
00:30:07,569 --> 00:30:10,230
each other's data, their internal geometric maps

646
00:30:10,230 --> 00:30:12,710
were floating in completely unaligned spaces.

647
00:30:13,309 --> 00:30:16,349
The researchers calculated a single simple mathematical

648
00:30:16,349 --> 00:30:19,390
rotation matrix to try and overlay the text map

649
00:30:19,390 --> 00:30:22,109
onto the vision map. Then they fed the text model

650
00:30:22,109 --> 00:30:24,990
a caption it had never seen before, translated

651
00:30:24,990 --> 00:30:28,390
that caption into mathematical coordinates, rotated

652
00:30:28,390 --> 00:30:30,470
those coordinates into the vision space, and

653
00:30:30,470 --> 00:30:33,009
asked the vision model to pull up the image that

654
00:30:33,009 --> 00:30:35,369
matched those coordinates. So finding an exact

655
00:30:35,369 --> 00:30:38,329
match completely blind, relying solely on whether

656
00:30:38,329 --> 00:30:40,509
the geometric shapes of their respective realities

657
00:30:40,509 --> 00:30:43,319
naturally fit together. Yes. What were the results?

658
00:30:44,289 --> 00:30:46,670
The results starkly validated the mathematics.

659
00:30:47,410 --> 00:30:50,069
The good view model achieved remarkable alignment.

660
00:30:50,490 --> 00:30:54,150
It was able to retrieve the exact, correct, true

661
00:30:54,150 --> 00:30:57,829
caption, first for 4 .4 % of the test images

662
00:30:57,829 --> 00:31:00,150
in a zero -shot setting. Now, to someone outside

663
00:31:00,150 --> 00:31:02,829
of computer science, 4 .4 % might sound a bit

664
00:31:02,829 --> 00:31:04,829
low, but we have to put that in context. We do.

665
00:31:04,990 --> 00:31:08,029
In a massive, dense data set filled with thousands

666
00:31:08,029 --> 00:31:10,730
of extremely similar images of dogs, beaches,

667
00:31:10,809 --> 00:31:13,970
and people, finding the exact right mathematical

668
00:31:13,970 --> 00:31:17,190
coordinate completely blind just by rotating

669
00:31:17,190 --> 00:31:21,049
a space. It is 44 times better than random chance.

670
00:31:21,069 --> 00:31:23,789
Wow. It proves that the geometry of the text

671
00:31:23,789 --> 00:31:26,269
concepts naturally mirrors the geometry of the

672
00:31:26,269 --> 00:31:29,130
visual concepts. But the real story is the catastrophic

673
00:31:29,130 --> 00:31:31,049
failure of the other two models. What happened

674
00:31:31,049 --> 00:31:33,730
to them? When they applied the exact same rotation

675
00:31:33,730 --> 00:31:36,769
to the bad view model and the next token prediction

676
00:31:36,769 --> 00:31:40,309
model, they failed. spectacularly. On the data

677
00:31:40,309 --> 00:31:42,390
charts presented in the paper, their alignment

678
00:31:42,390 --> 00:31:44,849
scores basically overlap each other at the absolute

679
00:31:44,849 --> 00:31:47,269
bottom. So despite having the exact same processing

680
00:31:47,269 --> 00:31:49,769
power and reading the exact same words, they

681
00:31:49,769 --> 00:31:51,650
couldn't align with the visual reality. Because

682
00:31:51,650 --> 00:31:54,170
they didn't actually learn concepts. They spent

683
00:31:54,170 --> 00:31:56,170
all their mathematical dimensions memorizing

684
00:31:56,170 --> 00:31:58,589
the specific phrasing and grammar of the Flickr

685
00:31:58,589 --> 00:32:00,849
captions. They learned the nuisance and missed

686
00:32:00,849 --> 00:32:04,440
the world. Precisely. And in dissecting why the

687
00:32:04,440 --> 00:32:06,640
good view model succeeded, the authors identified

688
00:32:06,640 --> 00:32:09,319
a very specific mechanism, a knob they could

689
00:32:09,319 --> 00:32:12,039
turn that drastically improved the results. They

690
00:32:12,039 --> 00:32:14,440
call it trimming. How does trimming work in this

691
00:32:14,440 --> 00:32:17,119
context? Well, even when using two different

692
00:32:17,119 --> 00:32:19,819
human captions as a good view, there is often

693
00:32:19,819 --> 00:32:22,660
overlapping vocabulary. If both humans use the

694
00:32:22,660 --> 00:32:25,440
word dog, the model might still try to take the

695
00:32:25,440 --> 00:32:27,839
lazy route. I see. It might just mathematically

696
00:32:27,839 --> 00:32:30,019
link the word dog in caption A to the word dog

697
00:32:30,019 --> 00:32:33,079
in caption B, completely bypassing the deeper

698
00:32:33,079 --> 00:32:36,160
semantic meaning. Trimming involves algorithmically

699
00:32:36,160 --> 00:32:38,700
cutting down the captions to remove those overlapping

700
00:32:38,700 --> 00:32:42,799
words. Oh, so if caption A is the brown dog leaps

701
00:32:42,799 --> 00:32:46,900
and caption B is the dog runs in the surf, trimming

702
00:32:46,900 --> 00:32:50,279
forces the model to compare brown leaps against

703
00:32:50,279 --> 00:32:53,470
runs in the surf. Yes. mathematically forces

704
00:32:53,470 --> 00:32:56,210
the neural network to bridge this semantic gap

705
00:32:56,210 --> 00:32:58,269
between leaping and running rather than just

706
00:32:58,269 --> 00:33:01,250
matching the identical word dog. Exactly. It

707
00:33:01,250 --> 00:33:03,710
strips away the shared nuisance of the vocabulary,

708
00:33:03,730 --> 00:33:05,650
forcing the model to discover the underlying

709
00:33:05,650 --> 00:33:08,190
conceptual world state that links those disparate

710
00:33:08,190 --> 00:33:10,750
words together. You know, if we translate this

711
00:33:10,750 --> 00:33:13,089
highly technical experiment into how human beings

712
00:33:13,089 --> 00:33:15,930
actually learn, It makes perfect intuitive sense.

713
00:33:16,289 --> 00:33:18,609
Really does. We don't learn what an apple is

714
00:33:18,609 --> 00:33:20,769
by sitting in a windowless room reading a dense

715
00:33:20,769 --> 00:33:24,150
Wikipedia article about apples and trying to

716
00:33:24,150 --> 00:33:26,009
statistically predict the next word in a sentence.

717
00:33:26,250 --> 00:33:29,369
That's next -token prediction. No, we learn what

718
00:33:29,369 --> 00:33:31,890
an apple is through good views. Right. We see

719
00:33:31,890 --> 00:33:34,670
a physical apple on a kitchen table. Then we

720
00:33:34,670 --> 00:33:37,490
see a stylized two -dimensional cartoon drawing

721
00:33:37,490 --> 00:33:40,069
of an apple in a children's book. Then we hear

722
00:33:40,069 --> 00:33:43,150
our parents say the audio waveform, apple. Those

723
00:33:43,150 --> 00:33:45,589
are radically different modalities. Completely

724
00:33:45,589 --> 00:33:48,730
different observations. But they are all anchored

725
00:33:48,730 --> 00:33:52,269
to the same underlying reality. The AI models

726
00:33:52,269 --> 00:33:54,069
that learn this way naturally build the same

727
00:33:54,069 --> 00:33:57,170
map of the world that we do. That is a very elegant

728
00:33:57,170 --> 00:33:59,430
way to frame it. And it's crucial to note that

729
00:33:59,430 --> 00:34:02,529
this isn't just a quirk of text processing. The

730
00:34:02,529 --> 00:34:04,670
authors expanded their experiments to the visual

731
00:34:04,670 --> 00:34:07,829
domain using the MSR VTT dataset. What's that?

732
00:34:08,090 --> 00:34:10,329
It's a standard benchmark consisting of thousands

733
00:34:10,329 --> 00:34:13,699
of short video clips. OK, but How do you translate

734
00:34:13,699 --> 00:34:16,559
the concept of good and bad views to a video

735
00:34:16,559 --> 00:34:19,300
file? A bad view for a video model would be the

736
00:34:19,300 --> 00:34:21,719
industry standard, taking a single frozen frame

737
00:34:21,719 --> 00:34:24,699
from the video and contrastively cropping it,

738
00:34:25,019 --> 00:34:26,920
comparing the top left corner of the frame to

739
00:34:26,920 --> 00:34:28,840
the bottom right corner of the exact same frame.

740
00:34:29,480 --> 00:34:32,500
You are forcing the model to memorize the static

741
00:34:32,500 --> 00:34:36,039
pixel noise of that specific millisecond. Exactly.

742
00:34:36,800 --> 00:34:39,320
A good view, however, leverages the nature of

743
00:34:39,320 --> 00:34:42,050
time. You take two completely different frames

744
00:34:42,050 --> 00:34:44,469
from the same continuous video clip, separated

745
00:34:44,469 --> 00:34:46,690
by a few seconds. Because the cameras likely

746
00:34:46,690 --> 00:34:49,230
move slightly. The lighting is shifted, the subjects

747
00:34:49,230 --> 00:34:52,269
are in different poses. Right. They are two distinct,

748
00:34:52,369 --> 00:34:54,429
independent observations of an ongoing scene.

749
00:34:55,010 --> 00:34:57,610
The nuisance changes, but the world state remains

750
00:34:57,610 --> 00:35:00,449
the same. And when they tested this? The results

751
00:35:00,449 --> 00:35:02,750
mirrored the text experiments perfectly. When

752
00:35:02,750 --> 00:35:05,230
they utilized good views on the vision side,

753
00:35:05,400 --> 00:35:08,420
by comparing different frames and paired it with

754
00:35:08,420 --> 00:35:11,860
good views on the text side, alignment skyrocketed

755
00:35:11,860 --> 00:35:14,760
in every single test they ran. The mathematical

756
00:35:14,760 --> 00:35:17,199
alignment of the AI models flawlessly tracks

757
00:35:17,199 --> 00:35:19,880
with the purity of the positive pairs they are

758
00:35:19,880 --> 00:35:22,579
trained on. So they proved the theorem on small,

759
00:35:22,840 --> 00:35:25,519
highly curated data sets like Flickr and MSR

760
00:35:25,519 --> 00:35:28,860
VTT. But the AI industry is currently defined

761
00:35:28,860 --> 00:35:31,820
by internet scale data. Massive data, yeah. Right.

762
00:35:31,900 --> 00:35:34,420
If this hypothesis is truly a breakthrough, it

763
00:35:34,420 --> 00:35:37,679
has to survive the chaotic, messy reality of

764
00:35:37,679 --> 00:35:40,659
web scale training. Does this hyper efficient,

765
00:35:41,099 --> 00:35:44,099
mathematically pure, good view method scale up

766
00:35:44,099 --> 00:35:46,099
or does it break down when you throw millions

767
00:35:46,099 --> 00:35:49,059
of random internet images at it? It doesn't break

768
00:35:49,059 --> 00:35:52,619
down. It scales magnificently. And this brings

769
00:35:52,619 --> 00:35:56,099
us to the empirical crescendo of the paper. They

770
00:35:56,099 --> 00:35:58,599
scaled their experiments using the recap data

771
00:35:58,599 --> 00:36:01,260
comp data set. What's in that? It consists of

772
00:36:01,260 --> 00:36:04,260
massive amounts of web text and images. And the

773
00:36:04,260 --> 00:36:06,480
statistics they generated here are the ones that

774
00:36:06,480 --> 00:36:08,920
should keep executives at major AI labs awake

775
00:36:08,920 --> 00:36:11,360
at night. OK, let's get into the numbers. What

776
00:36:11,360 --> 00:36:14,380
happens when David meets Goliath? First, they

777
00:36:14,380 --> 00:36:15,960
establish a baseline metric that we can call

778
00:36:15,960 --> 00:36:19,500
the 42x efficiency rule. They tracked how much

779
00:36:19,500 --> 00:36:21,940
training data a model needed to reach a specific

780
00:36:21,940 --> 00:36:25,139
high level of mathematical alignment with visual

781
00:36:25,139 --> 00:36:28,210
reality. OK. They found that a text JEPA trained

782
00:36:28,210 --> 00:36:31,269
purely on good views reached that benchmark using

783
00:36:31,269 --> 00:36:34,889
exactly 21 .6 million training samples. OK, 21

784
00:36:34,889 --> 00:36:37,510
.6 million. And how many samples did a standard

785
00:36:37,510 --> 00:36:39,429
Next Token prediction model need to hit that

786
00:36:39,429 --> 00:36:42,650
exact same level of understanding? 900 million

787
00:36:42,650 --> 00:36:45,530
samples. Wait, really? It needed to ingest nearly

788
00:36:45,530 --> 00:36:47,849
a billion samples to learn what the JEPA learned

789
00:36:47,849 --> 00:36:51,550
in 20 million? Yeah. That is mathematically 42

790
00:36:51,550 --> 00:36:54,190
times more sample efficient. That isn't an incremental

791
00:36:54,190 --> 00:36:57,599
improvement. That is a generational leap in efficiency.

792
00:36:58,400 --> 00:37:00,280
But they pushed it even further, right? They

793
00:37:00,280 --> 00:37:02,219
set up a direct showdown against the most famous

794
00:37:02,219 --> 00:37:04,519
models in the world? They did. They trained a

795
00:37:04,519 --> 00:37:06,539
slightly larger version of their good -view text

796
00:37:06,539 --> 00:37:09,559
JEPA. It had 109 billion parameters. To give

797
00:37:09,559 --> 00:37:12,480
the listeners some context on scale, 109 million

798
00:37:12,480 --> 00:37:14,880
parameters is microscopic by today's standards.

799
00:37:15,599 --> 00:37:17,400
That is a model you could comfortably run locally

800
00:37:17,400 --> 00:37:19,699
on a mid -range smart phone. Easily, yeah. It

801
00:37:19,699 --> 00:37:21,800
is a rounding error compared to what Meta or

802
00:37:21,800 --> 00:37:24,420
OpenAI are building. Absolutely. So they trained

803
00:37:24,420 --> 00:37:28,460
this tiny 109 million parameter model on 900

804
00:37:28,460 --> 00:37:31,480
million samples, meaning it processed roughly

805
00:37:31,480 --> 00:37:34,800
116 billion text tokens during its entire lifespan.

806
00:37:35,039 --> 00:37:39,119
OK. Then they evaluated it on the PRIH benchmark,

807
00:37:39,300 --> 00:37:43,480
specifically using the WIT 1024 data set. which

808
00:37:43,480 --> 00:37:46,219
is designed to test how well a text model's internal

809
00:37:46,219 --> 00:37:48,719
representation naturally aligns with a completely

810
00:37:48,719 --> 00:37:51,219
frozen state -of -the -art vision model called

811
00:37:51,219 --> 00:37:55,079
Dino V2. What's special about Dino V2? Dino V2

812
00:37:55,079 --> 00:37:57,260
is a powerhouse vision model trained without

813
00:37:57,260 --> 00:38:00,239
labels that understands deep geometric features

814
00:38:00,239 --> 00:38:03,059
of images. So they are testing how well this

815
00:38:03,059 --> 00:38:05,699
tiny text model naturally clicks into a massive

816
00:38:05,699 --> 00:38:08,670
visual map of reality. Who are they competing

817
00:38:08,670 --> 00:38:11,559
against? They pitted their tiny Juppa against

818
00:38:11,559 --> 00:38:14,579
the absolute titans of the open source AI world.

819
00:38:15,179 --> 00:38:18,400
They tested it against Quen 2 .5, which has 72

820
00:38:18,400 --> 00:38:20,780
billion parameters, and Meta's Lama 3, which

821
00:38:20,780 --> 00:38:25,239
boasts 70 billion parameters. A 109 million parameter

822
00:38:25,239 --> 00:38:28,320
model going up against 70 billion parameter behemoths.

823
00:38:29,219 --> 00:38:31,420
And those massive models weren't trained on a

824
00:38:31,420 --> 00:38:34,119
mere 116 billion tokens, were they? Not even

825
00:38:34,119 --> 00:38:36,420
close. Those models were trained on anywhere

826
00:38:36,420 --> 00:38:40,360
from 15 to 18 trillion tokens of internet data.

827
00:38:40,679 --> 00:38:43,880
Trillion. Trillion. They have ingested a significant

828
00:38:43,880 --> 00:38:46,880
portion of all recorded human text. So we have

829
00:38:46,880 --> 00:38:49,820
a model that is 660 times smaller in physical

830
00:38:49,820 --> 00:38:54,480
size, trained on over 100 times less data. It's

831
00:38:54,480 --> 00:38:56,820
not just David versus Goliath. It's an ant against

832
00:38:56,820 --> 00:38:59,579
an elephant. What happened when they ran the

833
00:38:59,579 --> 00:39:03,280
benchmark? The tiny 109 million parameter good

834
00:39:03,280 --> 00:39:06,550
view JEPA. beat all of them. Wait, I want to

835
00:39:06,550 --> 00:39:08,550
make sure I am hearing that correctly. It didn't

836
00:39:08,550 --> 00:39:12,230
just compete, it won. It won decisively. It achieved

837
00:39:12,230 --> 00:39:15,630
an alignment score of 0 .184 on the benchmark.

838
00:39:16,090 --> 00:39:21,789
The massive QN2 .5 -72B only scored 0 .154. Wow.

839
00:39:22,369 --> 00:39:24,610
The JEPA even outperformed the dedicated text

840
00:39:24,610 --> 00:39:28,030
encoder of CLIP, which is a massive model, explicitly

841
00:39:28,030 --> 00:39:30,929
designed and trained on perfectly paired image

842
00:39:30,929 --> 00:39:34,030
and text data to solve this exact problem. The

843
00:39:34,030 --> 00:39:36,170
implications of that are genuinely staggering.

844
00:39:37,030 --> 00:39:38,849
It means that the current strategy dominating

845
00:39:38,849 --> 00:39:41,769
Silicon Valley, buying tens of thousands of GPUs

846
00:39:41,769 --> 00:39:44,429
to train ever larger models on ever larger mountains

847
00:39:44,429 --> 00:39:48,409
of data, is fundamentally masking a massive architectural

848
00:39:48,409 --> 00:39:51,829
flaw. It is. Scale alone is not the lever to

849
00:39:51,829 --> 00:39:54,869
build intelligence or a true platonic representation

850
00:39:54,869 --> 00:39:58,369
of reality. The objective function is the lever.

851
00:39:58,710 --> 00:40:01,610
What this experiment proves is that if your objective

852
00:40:01,610 --> 00:40:04,409
function, like next token prediction, forces

853
00:40:04,409 --> 00:40:07,309
your model to memorize nuisance, throwing 15

854
00:40:07,309 --> 00:40:09,630
trillion tokens at it doesn't fix the problem.

855
00:40:09,829 --> 00:40:11,670
No, it just makes the problem bigger. Right.

856
00:40:11,849 --> 00:40:14,429
It just means the model is memorizing 15 trillion

857
00:40:14,429 --> 00:40:17,730
tokens worth of useless stylistic garbage, hoping

858
00:40:17,730 --> 00:40:21,110
to eventually accidentally stumble upon enough

859
00:40:21,110 --> 00:40:23,949
statistical correlations to mimic understanding.

860
00:40:24,139 --> 00:40:27,039
Whereas the good view JEPA, by ignoring the nuisance

861
00:40:27,039 --> 00:40:29,400
and focusing purely on the underlying world state

862
00:40:29,400 --> 00:40:32,860
from day one, learns the actual geometry of reality

863
00:40:32,860 --> 00:40:36,019
with hyper -efficiency. Exactly. So we now have

864
00:40:36,019 --> 00:40:38,380
a mathematically proven, hyper -efficient way

865
00:40:38,380 --> 00:40:41,400
to align different types of AI models so they

866
00:40:41,400 --> 00:40:43,800
share the exact same internal map of reality.

867
00:40:44,039 --> 00:40:47,420
We do. The theoretical part is fascinating, but

868
00:40:47,420 --> 00:40:50,599
engineering is about building things. What does

869
00:40:50,599 --> 00:40:53,389
this shared platonic geometry actually allow

870
00:40:53,389 --> 00:40:56,530
us to do in the real world. This leads us to

871
00:40:56,530 --> 00:40:58,309
the final section of their paper, which reads

872
00:40:58,309 --> 00:41:01,010
almost like science fiction. I love science fiction.

873
00:41:01,269 --> 00:41:04,110
Well, if the maps are identical, what are the

874
00:41:04,110 --> 00:41:06,789
practical applications? Because they map out

875
00:41:06,789 --> 00:41:09,750
the exact same geometric space, you can bypass

876
00:41:09,750 --> 00:41:12,170
the incredibly expensive and tedious processes

877
00:41:12,170 --> 00:41:14,630
that currently bottleneck AI development. Like

878
00:41:14,630 --> 00:41:17,150
what? Let's look at their first major application.

879
00:41:17,500 --> 00:41:20,539
building an image captioning system without ever

880
00:41:20,539 --> 00:41:22,760
using joint training. Okay, to give context,

881
00:41:23,460 --> 00:41:25,460
normally if you want an AI to look at a picture

882
00:41:25,460 --> 00:41:28,039
and write a descriptive caption, you have to

883
00:41:28,039 --> 00:41:30,679
build a bridge. A very expensive bridge. Right.

884
00:41:30,719 --> 00:41:32,960
You have to feed a neural network millions of

885
00:41:32,960 --> 00:41:35,000
images that already have human written captions

886
00:41:35,000 --> 00:41:38,059
explicitly attached to them. So the model slowly

887
00:41:38,059 --> 00:41:41,320
learns pixel by pixel and word by word how the

888
00:41:41,320 --> 00:41:43,639
two connect. It's incredibly labor -intensive

889
00:41:43,639 --> 00:41:46,579
to build those datasets. Exactly. But applying

890
00:41:46,579 --> 00:41:49,400
the Platonic representation theorem, they bypassed

891
00:41:49,400 --> 00:41:54,079
that entirely. They took the frozen Dino V2 vision

892
00:41:54,079 --> 00:41:56,579
model, which only knows pixels, and a frozen

893
00:41:56,579 --> 00:41:59,480
text model that was fine -tuned entirely on text.

894
00:42:00,500 --> 00:42:02,800
Neither model had ever seen a paired image in

895
00:42:02,800 --> 00:42:05,199
caption during their training. To connect them,

896
00:42:05,320 --> 00:42:08,269
they relied on theorem one. Because they know

897
00:42:08,269 --> 00:42:11,570
both models map the exact same underlying reality,

898
00:42:11,670 --> 00:42:14,030
they simply calculated a single linear rotation

899
00:42:14,030 --> 00:42:16,230
matrix to align their spaces. Going back to the

900
00:42:16,230 --> 00:42:18,389
detectives, they just spun one floor plan 90

901
00:42:18,389 --> 00:42:20,090
degrees so it overlaid the other floor plan.

902
00:42:20,110 --> 00:42:22,630
That's it. A simple mathematical rotation fitted

903
00:42:22,630 --> 00:42:25,690
on a very small set of anchor pairs. Just a rotation?

904
00:42:25,909 --> 00:42:28,449
Yes. Now, they understood that because these

905
00:42:28,449 --> 00:42:30,190
models are trained empirically, the alignment

906
00:42:30,190 --> 00:42:31,989
isn't going to be mathematically perfect down

907
00:42:31,989 --> 00:42:34,650
to the last decimal. There will be a tiny bit

908
00:42:34,650 --> 00:42:37,369
of misalignment error. Of course. So they introduced

909
00:42:37,369 --> 00:42:39,829
a brilliant engineering trick. During the text

910
00:42:39,829 --> 00:42:41,929
-only training of the decoder, they deliberately

911
00:42:41,929 --> 00:42:44,230
injected a specific amount of Gaussian noise

912
00:42:44,230 --> 00:42:46,630
into the text representations. They made the

913
00:42:46,630 --> 00:42:50,030
text coordinates a little fuzzy. Why? They mathematically

914
00:42:50,030 --> 00:42:52,530
calculated the expected alignment error between

915
00:42:52,530 --> 00:42:55,570
the vision and text spaces and matched the injected

916
00:42:55,570 --> 00:42:58,489
noise to that exact error rate, which roughly

917
00:42:58,489 --> 00:43:01,550
came out to a standard deviation, or sigma, of

918
00:43:01,550 --> 00:43:05,530
3 .3. Okay, what does that actually do? By doing

919
00:43:05,530 --> 00:43:07,590
this, they were essentially training the text

920
00:43:07,590 --> 00:43:09,989
decoder to say, the coordinates you are going

921
00:43:09,989 --> 00:43:11,949
to receive from the vision model might be a little

922
00:43:11,949 --> 00:43:14,769
blurry or slightly off -center, so learn how

923
00:43:14,769 --> 00:43:16,909
to generate accurate text even when the data

924
00:43:16,909 --> 00:43:20,630
is fuzzy. Oh, that is incredibly clever. Right.

925
00:43:21,030 --> 00:43:23,329
It made the system incredibly robust to the slight

926
00:43:23,329 --> 00:43:25,769
imperfections in the geometric rotation. It's

927
00:43:25,769 --> 00:43:27,849
like teaching a sniper to shoot in a crosswind

928
00:43:27,849 --> 00:43:29,769
before you even put them in the field. That's

929
00:43:29,769 --> 00:43:32,300
a great analogy. So they stitched this Frankenstein

930
00:43:32,300 --> 00:43:34,460
model together with nothing but a rotation matrix

931
00:43:34,460 --> 00:43:38,219
and some calculated noise. How well did it actually

932
00:43:38,219 --> 00:43:41,420
perform at captioning unseen images? They tested

933
00:43:41,420 --> 00:43:43,920
it on the widely recognized COCO dataset benchmark

934
00:43:43,920 --> 00:43:48,139
and it achieved a CIDR score of 0 .838. Explain

935
00:43:48,139 --> 00:43:51,059
CIDR for us. What does that score actually represent

936
00:43:51,059 --> 00:43:53,760
in human terms? CIDR is a metric designed to

937
00:43:53,760 --> 00:43:56,619
evaluate how human -like an AI's image description

938
00:43:56,619 --> 00:43:59,320
is by comparing it to multiple human reference

939
00:43:59,320 --> 00:44:03,519
captions. An 8 .38 is a highly competitive score.

940
00:44:03,619 --> 00:44:05,840
It means this AI system, which was never actually

941
00:44:05,840 --> 00:44:07,880
taught how words connect to pictures and has

942
00:44:07,880 --> 00:44:10,139
never seen an image text paired during its core

943
00:44:10,139 --> 00:44:12,860
training, is looking at novel images and generating

944
00:44:12,860 --> 00:44:15,039
descriptive, accurate captions that are nearly

945
00:44:15,039 --> 00:44:17,219
indistinguishable from what a real person would

946
00:44:17,219 --> 00:44:19,880
write. That alone is a massive leap in how we

947
00:44:19,880 --> 00:44:22,960
combine AI modalities. But the paper pushes it

948
00:44:22,960 --> 00:44:25,579
even further into what feels like magic. It does.

949
00:44:25,840 --> 00:44:28,500
The second application is zero -pair alignment.

950
00:44:28,840 --> 00:44:31,860
Yes. And this is perhaps the most mind -bending

951
00:44:31,860 --> 00:44:34,579
proof in the entire paper. It means perfectly

952
00:44:34,579 --> 00:44:36,639
aligning the vision model's understanding of

953
00:44:36,639 --> 00:44:39,300
reality with the text model's understanding of

954
00:44:39,300 --> 00:44:42,019
reality without using a single -paired image

955
00:44:42,019 --> 00:44:45,880
-text example. None. None. Zero. None. I have

956
00:44:45,880 --> 00:44:48,510
to pause. How is that mathematically possible?

957
00:44:48,650 --> 00:44:50,610
If you don't show the models at least one example

958
00:44:50,610 --> 00:44:52,590
of how they match up, like here is a picture

959
00:44:52,590 --> 00:44:55,170
of a cat and here is the word cat, now anchor

960
00:44:55,170 --> 00:44:57,869
yourselves to this, how can you possibly figure

961
00:44:57,869 --> 00:44:59,949
out how to rotate their spaces to align them?

962
00:45:00,250 --> 00:45:03,030
It traces back to proposition two in the paper.

963
00:45:03,250 --> 00:45:05,989
It proves that you don't need semantic anchors.

964
00:45:06,030 --> 00:45:08,849
You can find the perfect rotation matrix purely

965
00:45:08,849 --> 00:45:11,230
by analyzing the unique geometry of the data

966
00:45:11,230 --> 00:45:13,730
clouds themselves in the mathematical space.

967
00:45:13,909 --> 00:45:16,150
I need an analogy for this because visualizing

968
00:45:16,150 --> 00:45:19,110
multi -dimensional data clouds is tough. Let's

969
00:45:19,110 --> 00:45:22,489
use the analogy of maps. Okay. Imagine I give

970
00:45:22,489 --> 00:45:24,510
you two different maps of the exact same city.

971
00:45:25,070 --> 00:45:28,250
Map A is written entirely in English. Map B is

972
00:45:28,250 --> 00:45:31,170
written entirely in French. All right. And crucially,

973
00:45:31,170 --> 00:45:33,630
I have removed all the labels. There are no street

974
00:45:33,630 --> 00:45:36,210
names, no building names, no text whatsoever.

975
00:45:36,409 --> 00:45:39,550
It is just the pure geographic shapes of the

976
00:45:39,550 --> 00:45:42,030
city. Okay, so I can't just look for Main Street

977
00:45:42,030 --> 00:45:45,389
on the English map and Rue Principal on the French

978
00:45:45,389 --> 00:45:48,349
map to figure out how they line up. I have zero

979
00:45:48,349 --> 00:45:52,070
paired examples. Precisely. But if both maps

980
00:45:52,070 --> 00:45:54,610
are accurate representations of the exact same

981
00:45:54,610 --> 00:45:57,869
underlying reality, the physical city, you don't

982
00:45:57,869 --> 00:46:00,590
need labels. What do you look at then? You can

983
00:46:00,590 --> 00:46:02,769
just look at the geometric shapes. You notice

984
00:46:02,769 --> 00:46:05,690
the distinct winding curve of the river cutting

985
00:46:05,690 --> 00:46:08,059
through the center. You notice the specific grid

986
00:46:08,059 --> 00:46:10,539
pattern of the downtown streets. You see the

987
00:46:10,539 --> 00:46:12,880
unique asymmetrical shape of the large park in

988
00:46:12,880 --> 00:46:15,579
the east. Oh, I see. Because the geography of

989
00:46:15,579 --> 00:46:17,900
that city is a unique geometric fingerprint,

990
00:46:18,199 --> 00:46:20,739
you can simply rotate the English map until the

991
00:46:20,739 --> 00:46:23,460
river, the grid, and the park perfectly overlay

992
00:46:23,460 --> 00:46:26,679
onto the French map. You align them purely by

993
00:46:26,679 --> 00:46:29,099
matching the geometry of the lines on the page

994
00:46:29,099 --> 00:46:32,300
without ever needing a translator. Yes. Oh, wow.

995
00:46:32,920 --> 00:46:35,079
Because the structure of reality is unique. The

996
00:46:35,079 --> 00:46:37,420
way the AI clusters data about reality forms

997
00:46:37,420 --> 00:46:39,920
a unique geometric fingerprint in its embedding

998
00:46:39,920 --> 00:46:42,840
space. Exactly. The cluster for dogs and the

999
00:46:42,840 --> 00:46:45,360
cluster for cats and the cluster for cars form

1000
00:46:45,360 --> 00:46:47,820
a constellation. And the vision constellation

1001
00:46:47,820 --> 00:46:50,219
has the exact same shape as the text constellation.

1002
00:46:50,440 --> 00:46:52,400
Because they are observing the same reality.

1003
00:46:53,139 --> 00:46:55,679
So you just mathematically spin one constellation

1004
00:46:55,679 --> 00:46:58,440
until it clicks perfectly into the other? You've

1005
00:46:58,440 --> 00:47:01,440
got it exactly. And they put this theory to the

1006
00:47:01,440 --> 00:47:04,340
ultimate test using the schnauz benchmark. The

1007
00:47:04,340 --> 00:47:06,280
Schnell's benchmark. Yes. What does that do?

1008
00:47:06,880 --> 00:47:09,519
This benchmark takes the massive COCO data set

1009
00:47:09,519 --> 00:47:11,960
and splits it completely in half. The vision

1010
00:47:11,960 --> 00:47:14,059
model is allowed to look at the images from one

1011
00:47:14,059 --> 00:47:16,119
half. The text model is only allowed to look

1012
00:47:16,119 --> 00:47:18,380
at the captions from the other half. So there

1013
00:47:18,380 --> 00:47:22,480
was absolutely no overlap? No shared pairs. Exactly.

1014
00:47:22,840 --> 00:47:25,199
They are completely blind to each other, looking

1015
00:47:25,199 --> 00:47:27,539
at entirely different sets of data. They just

1016
00:47:27,539 --> 00:47:29,519
have to try and match the geometric shape of

1017
00:47:29,519 --> 00:47:33,130
the constellations. And what happened? The 109

1018
00:47:33,130 --> 00:47:36,230
million parameter good view text JEPA found the

1019
00:47:36,230 --> 00:47:38,969
alignment. Completely blind. And it didn't just

1020
00:47:38,969 --> 00:47:41,929
work as a proof of concept. It set new industry

1021
00:47:41,929 --> 00:47:45,190
records. What were the numbers? They evaluated

1022
00:47:45,190 --> 00:47:48,230
it on the ImageNet 100 zero -shot classification

1023
00:47:48,230 --> 00:47:51,050
test. This tests the model's ability to accurately

1024
00:47:51,050 --> 00:47:53,630
classify images it has never seen before purely

1025
00:47:53,630 --> 00:47:56,269
based on this blind geometric alignment. Okay.

1026
00:47:56,690 --> 00:48:00,409
The tiny JEPA achieved a 64 .4 % top -five accuracy.

1027
00:48:00,780 --> 00:48:04,380
Is that good for a zero -shot blind test? To

1028
00:48:04,380 --> 00:48:06,679
put it in perspective, the previous best score

1029
00:48:06,679 --> 00:48:10,199
ever reported on that exact benchmark using massive

1030
00:48:10,199 --> 00:48:12,920
language models scaling up to 8 billion parameters

1031
00:48:12,920 --> 00:48:17,420
was only 56%. Wait, 56 % from an 8 billion parameter

1032
00:48:17,420 --> 00:48:22,320
model? Yes. This tiny 109 million parameter model

1033
00:48:22,320 --> 00:48:24,719
armed with nothing but the pure geometry of good

1034
00:48:24,719 --> 00:48:27,280
views absolutely crushed the previous record

1035
00:48:27,280 --> 00:48:29,679
held by models nearly a hundred times its size.

1036
00:48:29,909 --> 00:48:32,250
A model you could run on a phone blindfolded

1037
00:48:32,250 --> 00:48:35,489
beat a massive data center model. It did. This

1038
00:48:35,489 --> 00:48:38,820
entire deep dive raises such an uncomfortable

1039
00:48:38,820 --> 00:48:41,300
question for the tech industry. I mean, if the

1040
00:48:41,300 --> 00:48:43,880
geometry of reality naturally aligns so perfectly

1041
00:48:43,880 --> 00:48:46,820
when using good views, how much of the current

1042
00:48:46,820 --> 00:48:49,480
AI boom is just an exercise in wasting billions

1043
00:48:49,480 --> 00:48:53,730
of dollars? A very fair question. We have massive

1044
00:48:53,730 --> 00:48:56,690
corporations spending unimaginable sums of money

1045
00:48:56,690 --> 00:48:59,670
paying humans to painstakingly pair and label

1046
00:48:59,670 --> 00:49:02,409
data sets or burning gigawatts of electricity

1047
00:49:02,409 --> 00:49:06,269
brute -forcing next token prediction across trillions

1048
00:49:06,269 --> 00:49:09,449
of words all in a desperate attempt to build

1049
00:49:09,449 --> 00:49:12,369
these clumsy massive bridges between vision and

1050
00:49:12,369 --> 00:49:16,949
text. Right. But Blestiero and Lacun just mathematically

1051
00:49:16,949 --> 00:49:19,090
prove that if you simply train the models on

1052
00:49:19,090 --> 00:49:22,170
good views from the start The bridge built itself.

1053
00:49:22,969 --> 00:49:25,610
It forces a profound and necessary reevaluation

1054
00:49:25,610 --> 00:49:28,530
of how we value scale versus architecture in

1055
00:49:28,530 --> 00:49:30,889
artificial intelligence. It really does. For

1056
00:49:30,889 --> 00:49:33,730
the past five years, we have been entirely captivated

1057
00:49:33,730 --> 00:49:37,090
by scale. We observed that a 70 billion parameter

1058
00:49:37,090 --> 00:49:40,349
model is vastly superior to a 7 billion parameter

1059
00:49:40,349 --> 00:49:43,269
model. And we mistakenly concluded that the path

1060
00:49:43,269 --> 00:49:46,010
to artificial general intelligence was just a

1061
00:49:46,010 --> 00:49:48,610
straight infinite line of adding more parameters

1062
00:49:48,610 --> 00:49:51,360
and scraping more data. But this paper is essentially

1063
00:49:51,360 --> 00:49:53,739
throwing a massive red flag on the play. It's

1064
00:49:53,739 --> 00:49:56,300
telling the industry stop. You are running as

1065
00:49:56,300 --> 00:49:58,840
fast as you can in the wrong direction. Yes,

1066
00:49:58,900 --> 00:50:01,239
you are covering ground, but you are on the wrong

1067
00:50:01,239 --> 00:50:03,880
path. Exactly. It proves that if next -token

1068
00:50:03,880 --> 00:50:06,420
prediction is fundamentally a bad view that forces

1069
00:50:06,420 --> 00:50:09,139
a model to spend its cognitive capacity memorizing

1070
00:50:09,139 --> 00:50:11,780
superficial nuisance rather than conceptual reality,

1071
00:50:12,380 --> 00:50:14,400
then it is mathematically doomed to be inefficient.

1072
00:50:14,559 --> 00:50:17,599
it will perpetually require exponentially more

1073
00:50:17,599 --> 00:50:21,780
data and exponentially more power to yield increasingly

1074
00:50:21,780 --> 00:50:25,360
diminishing returns in actual grounded understanding.

1075
00:50:27,119 --> 00:50:29,719
So let's bring all of this together and recap

1076
00:50:29,719 --> 00:50:31,639
the massive ground we've covered in this deep

1077
00:50:31,639 --> 00:50:34,940
dive. Sure. We started with Plato's cave and

1078
00:50:34,940 --> 00:50:37,980
a philosophical hypothesis to different AI models

1079
00:50:37,980 --> 00:50:40,800
trapped in different sensory modalities converge

1080
00:50:40,800 --> 00:50:43,750
on the same understanding of reality. And thanks

1081
00:50:43,750 --> 00:50:46,090
to this research, that is no longer a hypothesis.

1082
00:50:46,289 --> 00:50:48,869
It is the Platonic Representation Theorem. We

1083
00:50:48,869 --> 00:50:51,429
learned that JEPAs, when trained on good views,

1084
00:50:51,730 --> 00:50:53,650
which are completely different observations of

1085
00:50:53,650 --> 00:50:56,809
the exact same underlying scene, provably learned

1086
00:50:56,809 --> 00:50:59,190
the true mathematical shape of the world state,

1087
00:50:59,710 --> 00:51:02,090
completely ignoring the irrelevant nuisance of

1088
00:51:02,090 --> 00:51:05,030
their specific medium. And in stark contrast,

1089
00:51:05,150 --> 00:51:07,210
we learned that next -token prediction, which

1090
00:51:07,210 --> 00:51:09,269
is the foundational architecture of the entire

1091
00:51:09,269 --> 00:51:12,150
generative AI boom, acts inherently as a bad

1092
00:51:12,150 --> 00:51:14,909
view. Right. Because it only compares a prefix

1093
00:51:14,909 --> 00:51:18,329
to a continuation within a single isolated document.

1094
00:51:19,030 --> 00:51:21,469
It is doomed to learn the specific nuisance of

1095
00:51:21,469 --> 00:51:24,590
that document, the author's style, the specific

1096
00:51:24,590 --> 00:51:27,610
vocabulary, the formatting, rather than the actual

1097
00:51:27,610 --> 00:51:30,690
world state. It only manages to achieve a semblance

1098
00:51:30,690 --> 00:51:33,409
of alignment through massive, highly inefficient

1099
00:51:33,409 --> 00:51:36,110
brute force scale. We looked at the empirical

1100
00:51:36,110 --> 00:51:40,250
receipts that back up this math. A tiny 109 million

1101
00:51:40,250 --> 00:51:43,590
parameter model using good views decisively beat

1102
00:51:43,590 --> 00:51:47,869
a 70 billion parameter behemoth trained on 15

1103
00:51:47,869 --> 00:51:51,309
trillion tokens. It proved to be 42 times more

1104
00:51:51,309 --> 00:51:54,909
sample efficient. And it opens the door to sci

1105
00:51:54,909 --> 00:51:58,030
-fi level applications like aligning visual and

1106
00:51:58,030 --> 00:52:00,849
textual AI perfectly without a single piece of

1107
00:52:00,849 --> 00:52:03,670
paired training data simply by matching the unique

1108
00:52:03,670 --> 00:52:07,119
geometric fingerprints of reality itself. The

1109
00:52:07,119 --> 00:52:09,300
value of these insights for anyone following

1110
00:52:09,300 --> 00:52:12,579
the trajectory of AI really cannot be overstated.

1111
00:52:13,059 --> 00:52:14,960
Absolutely. We exist in a tech landscape that

1112
00:52:14,960 --> 00:52:17,739
is utterly obsessed with making models physically

1113
00:52:17,739 --> 00:52:20,340
bigger, building larger and more power -hungry

1114
00:52:20,340 --> 00:52:22,860
data centers and treating data like a raw material

1115
00:52:22,860 --> 00:52:25,719
to be strip mined. But this research definitively

1116
00:52:25,719 --> 00:52:28,199
proves that true intelligence and true understanding

1117
00:52:28,199 --> 00:52:30,940
of the world will likely come from changing how

1118
00:52:30,940 --> 00:52:33,579
we look at the data, not just continuously gorging

1119
00:52:33,579 --> 00:52:37,349
on more of it. Which leaves us with a final slightly

1120
00:52:37,349 --> 00:52:39,889
provocative thought to chew on as we wrap up

1121
00:52:39,889 --> 00:52:43,369
go for it Next token prediction is what birthed

1122
00:52:43,369 --> 00:52:46,269
chat GPT It is the core engine that launched

1123
00:52:46,269 --> 00:52:48,849
the current AI boom, which has completely consumed

1124
00:52:48,849 --> 00:52:51,309
the tech industry redirected billions in venture

1125
00:52:51,309 --> 00:52:54,590
capital and reshaped the stock market has But

1126
00:52:54,590 --> 00:52:57,469
if next token prediction is now proven fundamentally

1127
00:52:57,469 --> 00:53:00,150
and mathematically to be an incredibly inefficient

1128
00:53:00,150 --> 00:53:02,900
flawed way to map reality Are the hundreds of

1129
00:53:02,900 --> 00:53:04,940
billions of dollars currently being poured into

1130
00:53:04,940 --> 00:53:07,780
scaling these specific architectures going to

1131
00:53:07,780 --> 00:53:10,460
look like a massive evolutionary dead end in

1132
00:53:10,460 --> 00:53:13,400
just a few years? What happens to the titans

1133
00:53:13,400 --> 00:53:15,559
of the industry when everyone finally realizes

1134
00:53:15,559 --> 00:53:18,340
we've been spending billions trying to violently

1135
00:53:18,340 --> 00:53:21,300
brute force a lock when the mathematical key

1136
00:53:21,300 --> 00:53:23,300
was already sitting right there on the table?

1137
00:53:23,739 --> 00:53:26,599
It's a paradigm shifting question. I mean, if

1138
00:53:26,599 --> 00:53:29,380
you are investing billions of dollars and gigawatts

1139
00:53:29,380 --> 00:53:32,300
of power into studying shadow puppets, You are

1140
00:53:32,300 --> 00:53:34,699
going to be very surprised when someone finally

1141
00:53:34,699 --> 00:53:36,820
figures out how to turn around and just look

1142
00:53:36,820 --> 00:53:39,519
at the fire. We are still just staring at the

1143
00:53:39,519 --> 00:53:42,340
shadows on the wall of Plato's cave. But maybe,

1144
00:53:42,699 --> 00:53:45,860
just maybe, this math is the first step toward

1145
00:53:45,860 --> 00:53:46,920
finally turning around.