AI Is About to Crash. Here’s Why. — Transcript

Full transcript

1. 0:00Hello world. I'm an unemployed ex-big
2. 0:02tech software engineer with 25 years of
3. 0:05experience in the tech industry. So, is
4. 0:07it just me or is the AI bubble looking
5. 0:10extra bubbly these days? By extra
6. 0:13bubbly, I mean these AI companies,
7. 0:15they're still burning through truckloads
8. 0:17of cash every month. But all of a
9. 0:19sudden, the tone of their leadership
10. 0:22seems to have changed. Now they're
11. 0:24pushing for some of these insanely
12. 0:26valued IPOs that's higher than anything
13. 0:29that's ever been seen. While at the same
14. 0:31time, guys like Sam Altman, they're now
15. 0:34asking the government to help them
16. 0:36financially prop up their companies.
17. 0:39This is like peak bubble behavior right
18. 0:42there. It's something that you would see
19. 0:44right before a financial bubble bursts.
20. 0:47Everybody is running around trying to
21. 0:49find the next sucker. Oh, I mean
22. 0:51investor to take the bag before the
23. 0:53music stops. But why is this all
24. 0:56happening now?
25. 0:58Well, there has been a number of recent
26. 1:00changes that exposes the fundamental
27. 1:03unsustainability of the current AI
28. 1:05bubble. So, let's take a step back and
29. 1:08walk through what's actually going on
30. 1:10here. Now, I know I'm going to be
31. 1:12upsetting some true believers who think
32. 1:15that we're at or near artificial general
33. 1:17intelligence or AGI. But the truth is,
34. 1:20we don't have AGI. Sure, there's been
35. 1:23some fancy agentic loops and nice tool
36. 1:27integrations introduced, but the current
37. 1:29state of AI is still just a
38. 1:31probabilistic parrot predicting the most
39. 1:35likely next word based on its past
40. 1:38training data. Such a model is just not
41. 1:40capable of true reasoning and true
42. 1:43logic. And because of this derivative
43. 1:46nature, AI is simply not capable of left
44. 1:51field type innovations that
45. 1:53fundamentally increases economic
46. 1:55productivity. Like Claude, it's not just
47. 1:58going to go and invent the warp drive
48. 2:00anytime soon. However, what current AI
49. 2:04technology can conceivably do is to
50. 2:07replace repetitive cognitive human
51. 2:10labor. And it is my belief that the AI
52. 2:14boom is a gigantic leveraged bet on AI
53. 2:18profitably replacing this human labor.
54. 2:21Why is the replacement of human labor
55. 2:24the most logical use case for AI you
56. 2:26say? Well, if we look at the hard
57. 2:28numbers, right? Somewhere between 3 to 4
58. 2:32trillion dollars have been invested in
59. 2:34the American AI industry so far. Now,
60. 2:37some of that money is in the form of
61. 2:39investor cash, right? But the vast
62. 2:41majority of that money is in the form of
63. 2:43debt. Corporate [snorts]
64. 2:45bond debt.
65. 2:47And that debt has to be serviced.
66. 2:49Suppose we use a normal interest rate
67. 2:52for corporate bonds like 3 to 4%.
68. 2:55Well, for 2 to 3 trillion dollars of
69. 2:58debt, that works out to be about 100
70. 3:01billion dollars in interest that has to
71. 3:03be paid every year. That means the AI
72. 3:06industry has to make at least that much
73. 3:09profit every year just to break even,
74. 3:13just to service their debt. Now, suppose
75. 3:16they got a good profit margin going, say
76. 3:1810%, which in reality they don't, but
77. 3:21suppose they did. To make that kind of
78. 3:23profit, the AI industry would need to
79. 3:26replace a slice of the American economy
80. 3:29that's equivalent to around a trillion
81. 3:31dollars every year. And guess what? The
82. 3:34only slice of the American economy
83. 3:37that's big enough to sustain this kind
84. 3:39of replacement is the 10 trillion dollar
85. 3:42white-collar jobs economy. And that's
86. 3:45why AI must profitably replace
87. 3:48white-collar jobs to keep this bubble
88. 3:51going. Now, profitably replacing a
89. 3:54trillion dollars worth of white-collar
90. 3:56jobs,
91. 3:57that's like saying we have to profitably
92. 4:00replace 10 million American white-collar
93. 4:03workers a year. And this plan, I think,
94. 4:06is why all your AI leaders like Sam
95. 4:09Altman and Dario Amodei, they've been
96. 4:12going around for years now prophesizing
97. 4:14that huge amounts of jobs will simply
98. 4:17disappear in a kind of job apocalypse.
99. 4:20But things are not going according to
100. 4:22plan. In fact, the plan is actually
101. 4:25turning into a kind of dumpster fire
102. 4:27right now due to a couple of key
103. 4:29reasons.
104. 4:30Let's get into these reasons. Now,
105. 4:33frontier American AI models like
106. 4:36OpenAI's ChatGPT or Anthropic's Claude,
107. 4:39for example, these are all closed
108. 4:42models, meaning the tech companies
109. 4:45behind these models, they control
110. 4:48everything around the models. The
111. 4:50algorithms, the data, the compute,
112. 4:52everything. And they can sell their AI
113. 4:55models to consumers for money in the
114. 4:57form of subscriptions. But the problem
115. 4:59is that these closed AI frontier models,
116. 5:02they are insanely expensive to train and
117. 5:05to operate. The entire operation is
118. 5:08grossly unprofitable. Companies like
119. 5:11OpenAI and Anthropic, they're literally
120. 5:13losing money on every single API call
121. 5:16being made to their models. But here's
122. 5:19the thing.
123. 5:20Tech companies have long used a strategy
124. 5:22where they would burn tons and tons of
125. 5:25money to subsidize a service below
126. 5:28operational cost. Then they would try to
127. 5:30gain market share and become a monopoly.
128. 5:33Once all the competitors are dead, they
129. 5:36can then jack up the prices and profit.
130. 5:39Now, I have spoken at length by another
131. 5:42vlog on all of the algorithmic,
132. 5:45software, and hardware innovations that
133. 5:47Chinese AI companies have been making in
134. 5:50this space. But, long story short, with
135. 5:53just a fraction of America's compute
136. 5:56resources, these Chinese tech companies
137. 5:59have managed to create competitive
138. 6:01open-source models. Models that, by most
139. 6:05measures, are either slightly behind, on
140. 6:08par, or even slightly ahead the best
141. 6:12American frontier models. These Chinese
142. 6:15models, being open-source, it means that
143. 6:18they can be downloaded for free and then
144. 6:20run on a customer's own compute
145. 6:22infrastructure. And this could be done
146. 6:25for a tiny fraction of the cost of using
147. 6:28American closed AI models. A concrete
148. 6:31example of this is Moonshot's Kimi 3
149. 6:34model, right? I've been using this model
150. 6:37for a couple of days now, and to me,
151. 6:40this model's performance is comparable
152. 6:43to the nerfed version of Claude 5 Fable,
153. 6:46at least for my use cases. And I'm not
154. 6:49the only person recognizing this, right?
155. 6:52According to OpenRouter, Chinese
156. 6:55open-source models now account for more
157. 6:57than 60% of all tokens used by American
158. 7:01firms. So, the idea that American AI
159. 7:04companies can somehow create a monopoly,
160. 7:08jack up the prices, and rake in the
161. 7:10profits, this idea is now off the table.
162. 7:13Now, speaking of open-source models,
163. 7:16you can take a frontier open-source
164. 7:19model, and through techniques like
165. 7:21quantization and distillation, you can
166. 7:23compress this massive model down into a
167. 7:27much smaller local AI model. And instead
168. 7:30of being run on a big data center
169. 7:33somewhere, these local AI models can be
170. 7:36run on your home desktop, or a home
171. 7:38server, or even a good laptop. Now,
172. 7:41there are two benefits to this approach,
173. 7:44right? One is that because these models
174. 7:47are running locally on your own
175. 7:48computer, you don't have to pay any
176. 7:51subscription costs to the big tech AI
177. 7:53companies. The second benefit here is
178. 7:55that these local models can be run
179. 7:58entirely offline, disconnected from the
180. 8:01internet, giving people total privacy
181. 8:04over their own data. And in the last
182. 8:07couple of months, these local AI models
183. 8:10have suddenly become very capable. A
184. 8:13good local model like the Gwen 3.5, for
185. 8:17example, it's roughly comparable with
186. 8:20Claude Sonnet 4 on most tasks. And as
187. 8:24basic tasks are easily handled by these
188. 8:27local open-source models, people's
189. 8:29propensity to pay for premium AI from
190. 8:32these big tech companies, well, it
191. 8:34diminishes drastically. And all of this
192. 8:38leads to the most important reason,
193. 8:40which is that productivity gains from AI
194. 8:43is happening way slower than expected.
195. 8:46We ain't seeing anywhere near 10 million
196. 8:49white-collar workers being laid off by
197. 8:52AI this year. Now, I have spoken at
198. 8:55length about my own experiences with
199. 8:57agentic AI as a software engineer, and
200. 9:00all the challenges around getting good
201. 9:02quality output from AI. Fundamentally,
202. 9:05the challenges around managing context,
203. 9:08tiptoeing around training data gaps,
204. 9:11creating consistent workflows, and
205. 9:14avoiding hallucinations. These
206. 9:16challenges are real. It just takes a lot
207. 9:19of human effort and human intelligence
208. 9:22to use AI productively. And these
209. 9:25challenges are not just happening in
210. 9:27software engineering. It's happening in
211. 9:29other fields like customer service,
212. 9:32which was long considered to be this
213. 9:34low-hanging fruit for AI to automate.
214. 9:36There was this recent study that I read
215. 9:39where they surveyed thousands of
216. 9:41companies and it turned out that over
217. 9:4370% of the customer service agents that
218. 9:47went live, well, had to be either rolled
219. 9:49back or shut down because of various
220. 9:52mistakes and miscommunications that it
221. 9:54was making. The same study actually
222. 9:56showed a number of companies had
223. 9:58prematurely jumped the gun and laid off
224. 10:00their customer service reps only to then
225. 10:03frantically having to hire these people
226. 10:05back. So, the bottom line here is that
227. 10:08we're nowhere near being able to replace
228. 10:1010 million white-collar American workers
229. 10:13a year with AI. The most optimistic
230. 10:16estimates show something less than
231. 10:18100,000 workers being replaced a year by
232. 10:21AI, which I think is probably why both
233. 10:25Sam Altman and Dario Amodei are now kind
234. 10:28of walking back their AI job apocalypse
235. 10:31prophecies. So, net net, the American AI
236. 10:35industry has borrowed and spent huge
237. 10:37sums of money. They did it based on the
238. 10:39assumption that AI will replace human
239. 10:42cognitive labor at scale and for immense
240. 10:45profits. Now, don't get me wrong. I
241. 10:48strongly believe that AI will have a
242. 10:51transformative effect on the economy
243. 10:54over the long term in the same way that
244. 10:56railroads or the internet did. But AI,
245. 10:59just like these past technologies, is
246. 11:01simply not productive enough or reliable
247. 11:05enough to do this today. And these
248. 11:07American AI companies, to service their
249. 11:10mountain of debt and not go bankrupt,
250. 11:13they need to realize these profits
251. 11:16today. And the basic economics of it
252. 11:18all, it just doesn't work out. Thus, I
253. 11:21believe that the AI bubble, very much
254. 11:24like the railroad and internet bubbles,
255. 11:26is going to pop and likely very soon.
256. 11:29So, here's my one piece of advice to
257. 11:31you. These AI companies are going to be
258. 11:33desperate to raise money to keep this
259. 11:35whole show going and I expect that
260. 11:38they'll say or do just about anything to
261. 11:41keep the machine running. One way they
262. 11:43will try to do that is through wildly
263. 11:46overvalued IPOs. They're hoping that
264. 11:49retail investors are going to pile in
265. 11:52and buy into this dream. Don't be that
266. 11:55investor because when the music stops,
267. 11:58the people at the top they would have
268. 12:00all cashed out and someone will be left
269. 12:02holding the bag. Don't let that someone
270. 12:05be you and that's all I have to say
271. 12:08about that. Hope it helps. Anyways, if
272. 12:11you have a morbid curiosity to join me
273. 12:13on this life journey, please feel free
274. 12:15to subscribe to my channel and subscribe
275. 12:17to my Substack newsletter.
276. 12:19If you want to support me in my V log
277. 12:22creation efforts, please consider
278. 12:24becoming a paid member of this channel,
279. 12:26a paid member of my Substack, or just
280. 12:28buy me a coffee.
281. 12:30If you would like a one-on-one coaching
282. 12:32session with me, please feel free to
283. 12:33schedule it.
284. 12:35Anyways, thanks so much for watching.
285. 12:37Talk soon. Bye.