1
00:00:00,000 --> 00:00:03,553
Karpathy says the app you built should not even exist

2
00:00:03,904 --> 00:00:04,974
he is not scolding anyone else

3
00:00:05,264 --> 00:00:06,153
he is scolding himself

4
00:00:06,694 --> 00:00:07,623
in 2025

5
00:00:07,654 --> 00:00:09,443
OpenAI founding member Andrej

6
00:00:09,453 --> 00:00:11,342
Karpathy spent several days

7
00:00:11,373 --> 00:00:15,023
using AI assisted coding to build an app called MenuGen

8
00:00:15,213 --> 00:00:15,723
what does it do

9
00:00:16,214 --> 00:00:16,923
you walk into a restaurant

10
00:00:17,333 --> 00:00:18,783
the waiter hands you a menu

11
00:00:18,973 --> 00:00:20,702
most of the dishes have no pictures

12
00:00:20,894 --> 00:00:22,862
you have no idea what they look like

13
00:00:23,054 --> 00:00:24,322
a lot of people know this feeling

14
00:00:24,693 --> 00:00:25,943
especially abroad

15
00:00:25,973 --> 00:00:26,882
you cannot read the dish names

16
00:00:27,293 --> 00:00:29,142
and you have no clue what will arrive

17
00:00:29,414 --> 00:00:33,143
so Karpathy built a tool take a photo of the menu

18
00:00:33,293 --> 00:00:36,143
and it generates an image of every dish for you

19
00:00:36,333 --> 00:00:37,962
the app runs in the cloud

20
00:00:38,373 --> 00:00:39,742
front end and back end complete

21
00:00:40,574 --> 00:00:43,212
it used a multimodal large model for text recognition

22
00:00:43,214 --> 00:00:44,263
to extract the dish names

23
00:00:44,373 --> 00:00:47,062
and an image generation model to render the dishes

24
00:00:47,093 --> 00:00:49,373
plus user signup and online payment

25
00:00:49,373 --> 00:00:50,742
deployed on Vercel

26
00:00:50,854 --> 00:00:52,782
a full web application

27
00:00:53,359 --> 00:00:53,707
and then

28
00:00:54,078 --> 00:00:55,438
in April 2026

29
00:00:55,438 --> 00:00:57,347
Karpathy at the Sequoia Capital AI

30
00:00:57,347 --> 00:00:59,487
Ascent summit said one sentence

31
00:00:59,639 --> 00:01:01,847
that sentenced his own creation to death

32
00:01:01,942 --> 00:01:02,743
he said

33
00:01:02,758 --> 00:01:05,437
you just throw the menu photo at a large model

34
00:01:05,438 --> 00:01:09,288
with one line help me draw these dishes right onto the menu

35
00:01:09,598 --> 00:01:10,268
a few seconds later

36
00:01:10,678 --> 00:01:12,118
the model returns an image

37
00:01:12,119 --> 00:01:14,238
almost identical to the original menu

38
00:01:14,238 --> 00:01:17,167
except every dish now has a lifelike rendering next to it

39
00:01:17,318 --> 00:01:18,027
at the pixel level

40
00:01:18,279 --> 00:01:19,728
the rendering is simply done

41
00:01:19,839 --> 00:01:20,628
the entire MenuGen

42
00:01:20,798 --> 00:01:21,958
days of work

43
00:01:21,958 --> 00:01:22,788
thousands of lines of code

44
00:01:22,958 --> 00:01:24,368
did not need to exist at all

45
00:01:24,809 --> 00:01:25,157
hello

46
00:01:25,169 --> 00:01:25,637
hi everyone

47
00:01:25,768 --> 00:01:26,777
I am Leo Wang

48
00:01:27,008 --> 00:01:29,777
this episode is extremely dense with information

49
00:01:30,048 --> 00:01:32,377
the interview Karpathy gave at that summit

50
00:01:32,449 --> 00:01:33,678
I watched it over and over again

51
00:01:34,089 --> 00:01:35,167
and the more I watched the more I felt

52
00:01:35,169 --> 00:01:38,738
he is describing something most of us have not noticed yet

53
00:01:39,672 --> 00:01:41,472
he is not saying AI got better again

54
00:01:41,472 --> 00:01:43,242
not saying writing code got faster

55
00:01:43,353 --> 00:01:44,311
what he is saying is

56
00:01:44,312 --> 00:01:46,642
a brand new kind of computer is being born

57
00:01:47,073 --> 00:01:48,552
not a faster computer

58
00:01:48,552 --> 00:01:50,162
not smarter software

59
00:01:50,552 --> 00:01:53,002
but compared with what we have used for seventy years

60
00:01:53,112 --> 00:01:56,041
a completely different computing paradigm

61
00:01:56,392 --> 00:01:58,841
he calls it Software 3 0

62
00:01:59,713 --> 00:02:00,822
to understand 3 0

63
00:02:01,114 --> 00:02:03,602
you have to look back at 1 0 and 2 0

64
00:02:04,034 --> 00:02:05,283
Software 1 0

65
00:02:05,314 --> 00:02:07,683
is the traditional programming we know best

66
00:02:08,073 --> 00:02:09,063
a programmer writes code

67
00:02:09,474 --> 00:02:10,262
telling the computer

68
00:02:10,673 --> 00:02:11,512
take input A

69
00:02:11,513 --> 00:02:11,903
do B

70
00:02:12,233 --> 00:02:13,123
output C

71
00:02:13,434 --> 00:02:15,632
every rule and every branch

72
00:02:15,634 --> 00:02:18,162
is precisely defined by a human in code

73
00:02:18,393 --> 00:02:19,142
whatever you write

74
00:02:19,353 --> 00:02:20,482
is exactly what it does

75
00:02:20,553 --> 00:02:21,102
no more

76
00:02:21,393 --> 00:02:22,403
and no less

77
00:02:23,175 --> 00:02:24,544
Software 2 0

78
00:02:24,654 --> 00:02:27,463
is a concept Karpathy proposed back in 2017

79
00:02:27,735 --> 00:02:30,503
he was leading autopilot at Tesla then

80
00:02:30,935 --> 00:02:31,484
and he realized

81
00:02:31,854 --> 00:02:34,424
some problems cannot be enumerated with rules

82
00:02:34,814 --> 00:02:37,503
like getting a car to recognize a pedestrian

83
00:02:37,895 --> 00:02:38,843
how would you write that rule

84
00:02:39,559 --> 00:02:40,148
two legs

85
00:02:40,918 --> 00:02:41,507
wearing clothes

86
00:02:42,222 --> 00:02:42,667
has a head

87
00:02:43,679 --> 00:02:45,347
then what about someone riding a bike

88
00:02:45,999 --> 00:02:46,787
or holding an umbrella

89
00:02:47,358 --> 00:02:48,907
or crouching down to tie their shoes

90
00:02:49,679 --> 00:02:52,007
you could write rules all night and never finish

91
00:02:52,478 --> 00:02:53,548
the 2 0 approach is

92
00:02:53,879 --> 00:02:55,287
you stop writing rules

93
00:02:55,638 --> 00:02:57,967
you give the machine a mountain of labeled data

94
00:02:58,078 --> 00:03:01,407
and let the neural network learn the rules from the data itself

95
00:03:01,758 --> 00:03:03,048
the essence of programming

96
00:03:03,119 --> 00:03:07,127
becomes preparing datasets and designing training objectives

97
00:03:07,680 --> 00:03:09,709
so what about Software 3 0

98
00:03:10,321 --> 00:03:13,329
in 3 0 programming becomes talking

99
00:03:13,761 --> 00:03:14,790
you no longer write code

100
00:03:14,960 --> 00:03:16,730
you no longer prepare datasets

101
00:03:17,081 --> 00:03:18,869
you just tell a large language model

102
00:03:19,441 --> 00:03:21,850
draw the dishes on this menu for me

103
00:03:22,240 --> 00:03:23,549
the model understands what you mean

104
00:03:23,960 --> 00:03:25,010
and then executes

105
00:03:25,520 --> 00:03:29,290
the information and instructions you stuff into the context window

106
00:03:29,361 --> 00:03:32,010
are the lever you use to move this interpreter

107
00:03:32,481 --> 00:03:35,369
and the interpreter is the large language model itself

108
00:03:35,840 --> 00:03:37,290
it understands context

109
00:03:37,361 --> 00:03:39,570
runs computation in information space

110
00:03:39,641 --> 00:03:41,089
and gives you a result

111
00:03:41,474 --> 00:03:43,364
Karpathy says this is not a metaphor

112
00:03:43,514 --> 00:03:45,283
this is what is already happening

113
00:03:45,474 --> 00:03:47,964
he gave another example much closer to daily life

114
00:03:48,155 --> 00:03:50,443
there is an open source tool called OpenClaw

115
00:03:50,514 --> 00:03:52,270
the project I covered in an earlier episode

116
00:03:52,270 --> 00:03:54,244
by that Austrian programmer

117
00:03:54,435 --> 00:03:55,383
by traditional thinking

118
00:03:55,674 --> 00:03:56,913
to install software

119
00:03:56,914 --> 00:03:58,503
there should be an install script right

120
00:03:58,995 --> 00:03:59,873
but the problem is

121
00:03:59,875 --> 00:04:03,883
different operating systems different hardware different software environments

122
00:04:03,914 --> 00:04:06,193
that install script gets more and more bloated

123
00:04:06,194 --> 00:04:07,563
and more and more fragile

124
00:04:07,714 --> 00:04:08,474
fundamentally

125
00:04:08,474 --> 00:04:11,274
this is because you are still thinking in Software 1 0

126
00:04:11,275 --> 00:04:14,764
trying to write every install step precisely into code

127
00:04:15,314 --> 00:04:16,644
OpenClaw does not work that way

128
00:04:16,914 --> 00:04:18,884
what it gives you is just a block of text

129
00:04:19,274 --> 00:04:22,244
you copy and paste that text to your AI assistant

130
00:04:22,435 --> 00:04:23,514
the assistant reads it

131
00:04:23,514 --> 00:04:25,843
then checks your machine environment itself

132
00:04:25,995 --> 00:04:27,684
figures out which dependencies are needed

133
00:04:27,754 --> 00:04:30,084
and installs the software step by step

134
00:04:30,355 --> 00:04:31,304
if something breaks along the way

135
00:04:31,714 --> 00:04:32,754
it debugs itself

136
00:04:32,754 --> 00:04:33,723
and fixes itself

137
00:04:34,194 --> 00:04:35,783
did you notice the paradigm shift

138
00:04:36,314 --> 00:04:37,663
the old programming question was

139
00:04:37,995 --> 00:04:39,764
what code should I write

140
00:04:40,074 --> 00:04:41,673
the new question has become

141
00:04:41,675 --> 00:04:45,204
which block of text should I paste to my AI

142
00:04:45,795 --> 00:04:47,723
that sounds like a joke

143
00:04:47,874 --> 00:04:50,483
but it is daily life in Software 3 0

144
00:04:51,419 --> 00:04:52,779
once you understand these three stages

145
00:04:52,779 --> 00:04:54,698
look back at the MenuGen story

146
00:04:54,700 --> 00:04:56,068
and it reads completely differently

147
00:04:56,260 --> 00:04:58,068
when Karpathy was writing that app

148
00:04:58,180 --> 00:05:00,909
his head was still stuck in 1 0 and 2 0

149
00:05:01,099 --> 00:05:01,529
he was thinking

150
00:05:01,859 --> 00:05:04,229
I need a front end to upload images

151
00:05:04,260 --> 00:05:06,068
a back end to call the API

152
00:05:06,099 --> 00:05:08,258
a database to store user info

153
00:05:08,260 --> 00:05:10,029
and a payment system to charge money

154
00:05:10,339 --> 00:05:12,669
he was using the old paradigm on a new problem

155
00:05:13,020 --> 00:05:14,568
so what is the 3 0 solution

156
00:05:15,060 --> 00:05:15,909
there is no app

157
00:05:16,299 --> 00:05:19,378
just one natural language instruction and one image going in

158
00:05:19,380 --> 00:05:20,789
and one image coming out

159
00:05:21,020 --> 00:05:22,948
with no intermediate layer needed at all

160
00:05:24,032 --> 00:05:25,881
this is why Karpathy says

161
00:05:25,912 --> 00:05:28,361
everyone needs to readjust how they think

162
00:05:28,631 --> 00:05:30,760
do not keep reasoning along the existing paradigm

163
00:05:30,871 --> 00:05:34,441
and do not treat AI as just doing the old things faster

164
00:05:34,751 --> 00:05:35,910
the real change is

165
00:05:35,912 --> 00:05:38,481
a lot of new things have become possible

166
00:05:38,551 --> 00:05:41,160
things that simply did not exist in the old paradigm

167
00:05:41,511 --> 00:05:41,700
alright

168
00:05:42,071 --> 00:05:43,871
at this point you might think

169
00:05:43,871 --> 00:05:44,460
that sounds great

170
00:05:44,751 --> 00:05:46,080
the future looks bright

171
00:05:46,352 --> 00:05:47,950
but then Karpathy said

172
00:05:47,951 --> 00:05:50,041
something that makes everyone a little uncomfortable

173
00:05:50,444 --> 00:05:50,793
he said

174
00:05:51,124 --> 00:05:52,283
these AI models

175
00:05:52,285 --> 00:05:52,913
are not animals

176
00:05:53,244 --> 00:05:54,174
they are ghosts

177
00:05:54,525 --> 00:05:56,913
he once wrote an essay called Animals

178
00:05:57,004 --> 00:05:58,093
vs Ghosts

179
00:05:58,444 --> 00:05:59,814
animals versus ghosts

180
00:06:00,564 --> 00:06:02,314
where does animal intelligence come from

181
00:06:03,009 --> 00:06:03,809
evolution

182
00:06:04,124 --> 00:06:05,733
hundreds of millions of years of natural selection

183
00:06:05,805 --> 00:06:07,733
from single cell to multicell

184
00:06:07,764 --> 00:06:10,603
from fish to reptiles to mammals

185
00:06:10,605 --> 00:06:12,523
DNA encodes the learning algorithm

186
00:06:12,525 --> 00:06:16,293
then a long evolutionary outer loop polishes it bit by bit

187
00:06:16,644 --> 00:06:18,033
animals have intrinsic motivation

188
00:06:18,244 --> 00:06:18,954
they have curiosity

189
00:06:19,165 --> 00:06:20,083
they have fear

190
00:06:20,084 --> 00:06:20,674
they have hunger

191
00:06:20,925 --> 00:06:22,134
they have a drive to reproduce

192
00:06:22,444 --> 00:06:24,853
and all of it is baked deep into the genes

193
00:06:25,524 --> 00:06:26,974
AI did not come from any of that

194
00:06:27,284 --> 00:06:29,534
AI is a statistical simulation circuit

195
00:06:29,604 --> 00:06:31,733
trained on the text of the entire internet

196
00:06:32,084 --> 00:06:33,774
its base layer is pretraining

197
00:06:33,805 --> 00:06:35,733
which is large scale pattern matching

198
00:06:36,125 --> 00:06:38,293
with reinforcement learning stacked on top

199
00:06:38,365 --> 00:06:40,774
giving it a few directionally boosted abilities

200
00:06:41,204 --> 00:06:42,413
so Karpathy says

201
00:06:42,445 --> 00:06:44,243
we are not building animals

202
00:06:44,245 --> 00:06:45,974
we are summoning ghosts

203
00:06:46,284 --> 00:06:47,274
what are ghosts like

204
00:06:47,964 --> 00:06:49,053
they are uneven

205
00:06:49,284 --> 00:06:50,353
in English the word is jagged

206
00:06:50,604 --> 00:06:51,654
jagged like a saw blade

207
00:06:51,964 --> 00:06:53,053
the very same model

208
00:06:53,204 --> 00:06:56,173
ask it to refactor a hundred thousand line codebase

209
00:06:56,284 --> 00:06:57,694
and it does a beautiful job

210
00:06:57,964 --> 00:06:59,673
ask it to find a zero day security hole

211
00:06:59,964 --> 00:07:01,053
and it can find one

212
00:07:01,815 --> 00:07:05,783
but ask it this I want to wash my car fifty meters away

213
00:07:05,935 --> 00:07:07,644
should I drive or walk

214
00:07:08,214 --> 00:07:09,723
and it will tell you with a straight face

215
00:07:10,055 --> 00:07:10,853
just walk

216
00:07:10,855 --> 00:07:12,144
because the distance is short

217
00:07:12,254 --> 00:07:13,014
fifty meters

218
00:07:13,014 --> 00:07:14,144
so you walk over to the car wash

219
00:07:14,454 --> 00:07:16,464
and your car is still parked where it was

220
00:07:16,534 --> 00:07:17,243
you walk back

221
00:07:17,454 --> 00:07:18,704
and the car is still dirty

222
00:07:18,894 --> 00:07:22,904
a frontier model that handles human level engineering problems

223
00:07:22,935 --> 00:07:26,303
lacks the common sense that washing a car means driving it there

224
00:07:26,646 --> 00:07:28,696
the example everyone used to love was

225
00:07:28,807 --> 00:07:32,215
how many r are there in the English word strawberry

226
00:07:32,487 --> 00:07:34,016
models often got that wrong

227
00:07:34,326 --> 00:07:37,295
but that one has already been patched

228
00:07:37,607 --> 00:07:39,126
the car wash example shows

229
00:07:39,126 --> 00:07:41,525
that this jaggedness is structural

230
00:07:41,526 --> 00:07:42,735
not accidental

231
00:07:42,938 --> 00:07:43,489
why is that

232
00:07:43,700 --> 00:07:46,188
Karpathy gave a very fundamental explanation

233
00:07:46,539 --> 00:07:50,308
two variables decide where a model suddenly gets strong

234
00:07:50,780 --> 00:07:52,909
the first variable is verifiability

235
00:07:53,219 --> 00:07:56,349
the core mechanism of model training is reinforcement learning

236
00:07:56,619 --> 00:07:57,888
get it right and you get a reward

237
00:07:58,219 --> 00:07:59,469
get it wrong and you lose points

238
00:07:59,859 --> 00:08:02,659
but only tasks whose results can be verified automatically

239
00:08:02,659 --> 00:08:05,229
can go into that training loop effectively

240
00:08:05,780 --> 00:08:07,308
math can be verified

241
00:08:07,659 --> 00:08:08,768
code can be verified

242
00:08:09,179 --> 00:08:10,989
just run it and you know if it works

243
00:08:11,340 --> 00:08:13,909
so models made huge leaps in those areas

244
00:08:14,451 --> 00:08:17,420
but common sense like driving to a car wash

245
00:08:17,492 --> 00:08:18,441
how do you verify that automatically

246
00:08:18,932 --> 00:08:20,001
with no verification environment

247
00:08:20,211 --> 00:08:21,741
there is no training signal

248
00:08:21,932 --> 00:08:24,741
the second variable is what AI companies care about

249
00:08:25,052 --> 00:08:26,741
Karpathy gave an example

250
00:08:27,011 --> 00:08:29,460
from GPT 3 5 to GPT 4

251
00:08:29,492 --> 00:08:33,020
many people felt the model got much better at chess

252
00:08:33,211 --> 00:08:35,941
and assumed it came from general intelligence rising

253
00:08:36,052 --> 00:08:37,340
he says not necessarily

254
00:08:37,532 --> 00:08:38,240
from what he knows

255
00:08:38,571 --> 00:08:40,651
someone at OpenAI made a decision

256
00:08:40,652 --> 00:08:44,340
to add a huge amount of chess game data into the pretraining set

257
00:08:44,611 --> 00:08:45,860
and that single decision

258
00:08:45,892 --> 00:08:47,931
made the model performance at chess

259
00:08:47,932 --> 00:08:50,140
spike into an enormous capability peak

260
00:08:50,772 --> 00:08:53,620
completely out of proportion with gains elsewhere

261
00:08:54,752 --> 00:08:56,600
so the capability landscape of a model

262
00:08:56,632 --> 00:08:58,681
is not a plane rising evenly

263
00:08:58,752 --> 00:09:02,120
it is rugged terrain full of peaks and deep valleys

264
00:09:02,752 --> 00:09:06,681
the peaks are areas reinforcement learning and datasets covered heavily

265
00:09:07,032 --> 00:09:10,360
the valleys are blind spots the data distribution never touched

266
00:09:10,912 --> 00:09:13,831
if your use case happens to stand on a peak

267
00:09:13,831 --> 00:09:15,401
you will think AI is a god

268
00:09:15,872 --> 00:09:17,431
if you land in a valley

269
00:09:17,432 --> 00:09:19,801
you will think AI is dumber than a schoolkid

270
00:09:20,351 --> 00:09:22,480
for everyone who uses AI

271
00:09:22,512 --> 00:09:24,561
this is an extremely important realization

272
00:09:25,032 --> 00:09:27,760
you cannot just say AI works or does not work

273
00:09:28,071 --> 00:09:29,100
you have to figure out

274
00:09:29,351 --> 00:09:30,870
in your own specific scenario

275
00:09:30,872 --> 00:09:33,041
where exactly you stand on this terrain

276
00:09:33,792 --> 00:09:36,521
if you are not inside the circuits reinforcement learning covered

277
00:09:36,552 --> 00:09:38,480
you may need to fine tune the model yourself

278
00:09:38,632 --> 00:09:41,401
or do extra engineering work to make up for it

279
00:09:42,269 --> 00:09:44,697
so what is left for humans in this new world

280
00:09:45,188 --> 00:09:48,077
Karpathy quoted a line he cannot stop thinking about

281
00:09:48,188 --> 00:09:50,118
and it stuck with me too

282
00:09:50,389 --> 00:09:52,118
you can outsource your thinking

283
00:09:52,149 --> 00:09:54,118
but you cannot outsource your understanding

284
00:09:54,428 --> 00:09:56,758
he says he is becoming a bottleneck himself

285
00:09:57,029 --> 00:09:59,707
even just understanding what we actually want to build

286
00:09:59,708 --> 00:10:01,237
why this thing is worth doing

287
00:10:01,389 --> 00:10:03,638
how I should direct my AI assistant

288
00:10:03,828 --> 00:10:06,638
these things alone are becoming the bottleneck

289
00:10:07,068 --> 00:10:10,118
information has to get into his brain somehow

290
00:10:10,188 --> 00:10:12,357
and the brain has limited bandwidth

291
00:10:12,969 --> 00:10:14,567
this is also why Karpathy

292
00:10:14,569 --> 00:10:17,297
is so excited about something called a knowledge base

293
00:10:17,489 --> 00:10:20,338
he built a project called LLM Wiki

294
00:10:20,368 --> 00:10:22,858
what he calls a large model knowledge base

295
00:10:23,008 --> 00:10:25,338
you let the model read a pile of documents

296
00:10:25,408 --> 00:10:27,738
and it automatically builds a structured knowledge base

297
00:10:27,929 --> 00:10:29,618
every time he reads an article

298
00:10:29,689 --> 00:10:32,098
the knowledge base updates and reorganizes itself

299
00:10:32,288 --> 00:10:36,138
then he can ask questions follow up and dig deeper around it

300
00:10:36,408 --> 00:10:37,958
what is the essence of a tool like this

301
00:10:38,408 --> 00:10:40,218
it is not amplifying your output

302
00:10:40,329 --> 00:10:41,978
it is amplifying your understanding

303
00:10:42,168 --> 00:10:43,927
because if you do not understand

304
00:10:43,929 --> 00:10:46,378
you cannot be a good commander

305
00:10:46,689 --> 00:10:49,937
large language models are not good at real understanding

306
00:10:50,809 --> 00:10:52,057
and you as a human

307
00:10:52,089 --> 00:10:54,608
are still the one unique role in the system

308
00:10:54,609 --> 00:10:56,138
the one responsible for understanding

309
00:10:56,413 --> 00:10:58,092
in this interview Karpathy

310
00:10:58,092 --> 00:11:00,021
also made a prediction about the future

311
00:11:00,092 --> 00:11:01,501
that I find very bold

312
00:11:01,656 --> 00:11:02,413
he said

313
00:11:02,413 --> 00:11:06,261
he believes computers built entirely from neural networks are coming

314
00:11:06,653 --> 00:11:09,621
the device takes raw video and audio input directly

315
00:11:09,653 --> 00:11:11,141
feeds it into a neural network

316
00:11:11,212 --> 00:11:14,702
and a diffusion model renders a user interface in real time

317
00:11:14,852 --> 00:11:18,422
an interface generated on the spot for this exact moment

318
00:11:18,895 --> 00:11:21,623
computers today are still fundamentally calculators

319
00:11:21,895 --> 00:11:22,764
back in the fifties

320
00:11:23,135 --> 00:11:24,863
people were actually not sure

321
00:11:24,934 --> 00:11:28,463
whether a computer should look like a calculator or like a brain

322
00:11:28,895 --> 00:11:31,103
later we chose the calculator path

323
00:11:31,574 --> 00:11:32,823
and ran with it for seventy years

324
00:11:32,855 --> 00:11:36,264
which gave us chips operating systems and app stores

325
00:11:36,775 --> 00:11:39,144
but the future may flip this around

326
00:11:39,615 --> 00:11:41,664
the neural network becomes the main process

327
00:11:41,775 --> 00:11:44,144
the traditional CPU becomes a coprocessor

328
00:11:44,574 --> 00:11:47,424
the neural network carries the vast majority of the core load

329
00:11:47,735 --> 00:11:51,333
and the deterministic computing where one plus one must equal two

330
00:11:51,334 --> 00:11:52,823
becomes an attached tool

331
00:11:53,535 --> 00:11:57,823
just like GPUs already matter more than CPUs in many scenarios today

332
00:11:57,855 --> 00:12:01,264
except this trend will push into every kind of computing

333
00:12:01,374 --> 00:12:04,622
Karpathy also raised an impact most people never considered

334
00:12:04,622 --> 00:12:05,398
hiring

335
00:12:05,774 --> 00:12:09,622
he says most companies now want to hire strong AI engineers

336
00:12:09,813 --> 00:12:12,543
but the interview format is stuck in the old paradigm

337
00:12:12,854 --> 00:12:14,222
hand out a little puzzle

338
00:12:14,293 --> 00:12:16,942
and have the candidate write an algorithm on a whiteboard

339
00:12:17,293 --> 00:12:18,362
what is that actually testing

340
00:12:18,894 --> 00:12:19,643
it tests memory

341
00:12:20,014 --> 00:12:21,982
it tests fluency at writing code by hand

342
00:12:22,414 --> 00:12:25,372
but in the Software 3 0 era those skills

343
00:12:25,374 --> 00:12:26,742
are worth almost nothing

344
00:12:27,791 --> 00:12:29,879
he suggests interviews should look like this

345
00:12:30,031 --> 00:12:31,879
give the candidate a big project

346
00:12:31,911 --> 00:12:34,670
say use AI to build a complete social platform

347
00:12:34,671 --> 00:12:35,999
and deploy it live

348
00:12:36,511 --> 00:12:39,639
then the company sends ten AI agents to attack that platform

349
00:12:39,710 --> 00:12:40,960
and try to break it

350
00:12:41,271 --> 00:12:42,899
watch how the candidate uses tools

351
00:12:43,191 --> 00:12:44,479
how they organize the architecture

352
00:12:44,511 --> 00:12:48,759
how they build a system that is both powerful and secure with AI helping

353
00:12:49,151 --> 00:12:51,519
the test is no longer whether you can write code

354
00:12:51,590 --> 00:12:55,119
it is whether you can direct AI to build something genuinely solid

355
00:12:55,911 --> 00:12:57,700
what is the implicit signal here

356
00:12:58,271 --> 00:13:01,440
even the standard for judging talent is being redefined

357
00:13:02,317 --> 00:13:03,806
hearing this you might want to ask

358
00:13:03,917 --> 00:13:05,625
so what exactly should humans do now

359
00:13:06,197 --> 00:13:07,846
the answer Karpathy gives is very practical

360
00:13:07,981 --> 00:13:08,780
he says

361
00:13:08,797 --> 00:13:11,885
right now AI is still an intern with very jagged skills

362
00:13:11,957 --> 00:13:13,556
sometimes the work is stunning

363
00:13:13,557 --> 00:13:15,446
sometimes it makes incredibly dumb mistakes

364
00:13:15,688 --> 00:13:17,978
he used a real bug in MenuGen as an example

365
00:13:18,089 --> 00:13:19,978
users signed up with a Google account

366
00:13:20,048 --> 00:13:22,538
but bought credits through Stripe

367
00:13:22,688 --> 00:13:24,697
both systems had an email address

368
00:13:24,729 --> 00:13:26,618
and when the AI wrote the code

369
00:13:26,688 --> 00:13:29,207
it simply matched the two email addresses

370
00:13:29,209 --> 00:13:31,697
to link payments back to users

371
00:13:31,808 --> 00:13:32,398
the problem is

372
00:13:32,648 --> 00:13:35,087
you could easily use one email for Google

373
00:13:35,089 --> 00:13:36,618
and another email for payment

374
00:13:36,688 --> 00:13:37,478
and the moment they differ

375
00:13:37,688 --> 00:13:39,898
the system loses your money

376
00:13:40,009 --> 00:13:41,888
because the AI never realized

377
00:13:41,889 --> 00:13:45,728
you should bind all systems with one unified user ID

378
00:13:45,729 --> 00:13:48,898
not with something as arbitrary as an email

379
00:13:49,528 --> 00:13:50,528
a mistake like this

380
00:13:50,528 --> 00:13:53,538
any experienced engineer spots instantly

381
00:13:53,808 --> 00:13:55,297
but the AI cannot see it

382
00:13:55,648 --> 00:13:59,057
it is only predicting the next most likely snippet of code

383
00:13:59,369 --> 00:14:02,097
not understanding the underlying logic of the system design

384
00:14:02,527 --> 00:14:03,816
so Karpathy says

385
00:14:04,007 --> 00:14:06,326
humans must still own the spec

386
00:14:06,326 --> 00:14:07,365
own the plan

387
00:14:07,367 --> 00:14:08,235
own the taste

388
00:14:08,446 --> 00:14:09,536
own the judgment

389
00:14:09,806 --> 00:14:12,335
and AI fills in the blanks inside that frame

390
00:14:12,806 --> 00:14:13,796
the human is the architect

391
00:14:14,127 --> 00:14:15,576
the AI is the construction crew

392
00:14:16,007 --> 00:14:18,855
but then he said something that sent a chill down my spine

393
00:14:18,970 --> 00:14:19,771
he said

394
00:14:19,966 --> 00:14:23,936
there is nothing fundamental stopping AI from improving at these things too

395
00:14:24,326 --> 00:14:27,135
taste judgment aesthetics

396
00:14:27,287 --> 00:14:28,975
the reason they do not work yet

397
00:14:29,086 --> 00:14:31,326
may simply be that AI companies have not

398
00:14:31,326 --> 00:14:34,215
put them into the reinforcement learning training environment

399
00:14:34,486 --> 00:14:37,896
nobody has built a reward function that trains taste specifically

400
00:14:38,887 --> 00:14:40,335
and once they figure out how

401
00:14:40,367 --> 00:14:42,686
these abilities will follow coding ability

402
00:14:42,686 --> 00:14:43,696
and take a sudden leap

403
00:14:44,366 --> 00:14:45,735
he gave one more example

404
00:14:45,927 --> 00:14:48,495
he built a project called microGPT

405
00:14:48,567 --> 00:14:52,296
the goal was to simplify LLM training to the absolute minimum

406
00:14:52,687 --> 00:14:55,015
but the model really resisted it

407
00:14:55,167 --> 00:14:58,336
he kept prompting the AI to simplify and to refine

408
00:14:58,407 --> 00:14:59,735
and the AI just could not do it

409
00:14:59,886 --> 00:15:01,395
it felt like pulling teeth

410
00:15:01,687 --> 00:15:03,296
rather than moving at light speed

411
00:15:03,567 --> 00:15:06,096
because that sense of simplicity and elegance

412
00:15:06,167 --> 00:15:08,895
is simply not in the reinforcement learning circuits of the model

413
00:15:09,167 --> 00:15:11,296
but that does not mean it never will be

414
00:15:11,687 --> 00:15:15,015
it is just that nobody has trained that ability yet

415
00:15:15,967 --> 00:15:18,576
so here is the situation we are facing

416
00:15:18,886 --> 00:15:21,736
a brand new computing paradigm is being born

417
00:15:22,126 --> 00:15:23,635
programming is no longer writing code

418
00:15:23,886 --> 00:15:24,895
it is talking

419
00:15:25,246 --> 00:15:27,135
apps are no longer a necessity

420
00:15:27,327 --> 00:15:31,415
in many cases one natural language sentence replaces a whole application

421
00:15:31,967 --> 00:15:33,515
AI capability is jagged

422
00:15:33,927 --> 00:15:35,086
just like a ghost

423
00:15:35,087 --> 00:15:36,566
superhuman in some domains

424
00:15:36,567 --> 00:15:39,096
and without common sense in others

425
00:15:39,447 --> 00:15:40,876
the last moat humans have

426
00:15:41,207 --> 00:15:42,126
is not skill

427
00:15:42,126 --> 00:15:43,175
it is not knowledge

428
00:15:43,207 --> 00:15:44,555
it is not even the ability to think

429
00:15:44,727 --> 00:15:45,775
it is understanding

430
00:15:46,126 --> 00:15:47,956
but how long can that moat hold

431
00:15:48,687 --> 00:15:52,535
Karpathy says taste and judgment just have not been trained yet

432
00:15:53,286 --> 00:15:54,246
which means

433
00:15:54,246 --> 00:15:58,336
the window left for humans may be far shorter than we assume

434
00:15:58,882 --> 00:16:00,172
as an investor

435
00:16:00,283 --> 00:16:02,291
here is the picture I see

436
00:16:02,602 --> 00:16:04,491
in the past when we invested in a company

437
00:16:04,642 --> 00:16:06,452
we mainly looked at its technical moat

438
00:16:06,803 --> 00:16:08,052
the problem now is

439
00:16:08,122 --> 00:16:11,852
technical moats are being eroded faster than ever before

440
00:16:12,242 --> 00:16:16,651
one person with a few sentences can have AI build a full product prototype

441
00:16:17,083 --> 00:16:18,432
so where is the real moat

442
00:16:19,002 --> 00:16:21,531
maybe it sits inside that one word Karpathy used

443
00:16:21,666 --> 00:16:22,467
understanding

444
00:16:23,175 --> 00:16:24,903
deep understanding of an industry

445
00:16:25,214 --> 00:16:26,863
deep understanding of users

446
00:16:27,214 --> 00:16:30,224
deep understanding of what is worth doing and what is not

447
00:16:30,575 --> 00:16:33,263
those are things AI still cannot hand you

448
00:16:33,494 --> 00:16:34,243
but here is the question

449
00:16:34,614 --> 00:16:36,724
are you actually doing the understanding

450
00:16:37,094 --> 00:16:39,724
or are you outsourcing understanding to AI as well

451
00:16:40,374 --> 00:16:40,844
that question

452
00:16:41,134 --> 00:16:42,103
I leave to you

453
00:16:42,214 --> 00:16:42,883
I am Leo Wang

454
00:16:43,175 --> 00:16:44,353
see you next episode