YouTube2Text

El fin del mundo según OpenAI y Anthropic: ¿qué hay detrás de sus advertencias? — Transcript

by Marc Vidal · 6,552 words · 1,030 segments · language en · Watch on YouTube

Full transcript

  1. 0:00A researcher leaves Anthropic, warns
  2. 0:02that artificial intelligence could
  3. 0:04destroy us before the end of this
  4. 0:07decade, and his message has been viewed
  5. 0:09more than 100 million times in a very
  6. 0:11short space of time. Other researchers
  7. 0:14confirm that they are also very worried
  8. 0:17. The media, of course, fills their
  9. 0:20headlines with the possibility of
  10. 0:21losing control of machines and that the
  11. 0:23end of the world is getting closer. And
  12. 0:27a few days later, Dario Amodei, the CEO
  13. 0:30of Anthropic, publishes a proposal to
  14. 0:32slow down the development of the most
  15. 0:34advanced systems, and immediately after
  16. 0:37, to bring external evaluators into
  17. 0:39large laboratories and allow companies
  18. 0:41to coordinate to set certain safety
  19. 0:43limits. Then, Samalman, the head of
  20. 0:47Openi, the one from Chat GPT, is in
  21. 0:49favor. Elon Musk too, and almost at the
  22. 0:53same time, even Openi announces that
  23. 0:552026 does not seem like a good time to
  24. 0:57go public, and they push their IPO back
  25. 0:59to next year. When an entire industry
  26. 1:03asks to be regulated, it would be
  27. 1:05advisable to be suspicious by default,
  28. 1:07and not because they are lying, I am
  29. 1:09not saying that they are, but because
  30. 1:10there is almost always another reason,
  31. 1:12at least two reasons, and neither of
  32. 1:14them is the one that appears in the
  33. 1:16headline. The first is that the numbers
  34. 1:18do not add up for them, which is why
  35. 1:20they want to delay the launch. The
  36. 1:23second is that the State, at least the
  37. 1:25State, has never been bothered by
  38. 1:27keeping the key, the key to regulation.
  39. 1:30That is what I want to present to you
  40. 1:31today. Not that they have reached an
  41. 1:33agreement in a secret room, right?
  42. 1:35Because I cannot prove that and it
  43. 1:37probably has not happened. What I see
  44. 1:40is an industry that has entered into a
  45. 1:42race that demands hundreds of billions
  46. 1:44of dollars, electricity measured in
  47. 1:45gigawatts, chips that age at a brutal
  48. 1:47speed and an infrastructure that does
  49. 1:49not grow at the same rate as demand. A
  50. 1:53race in which no one has yet
  51. 1:54demonstrated, by the way, how much
  52. 1:56money each dollar invested returns. And
  53. 2:00at that precise moment, those ahead
  54. 2:02have discovered that slowing down is
  55. 2:04the most prudent thing to do, and
  56. 2:05governments discover that they finally
  57. 2:07have an impeccable excuse to decide who
  58. 2:10can build the technology that will move
  59. 2:12the century. Hm. The first existential
  60. 2:16problem of artificial intelligence may
  61. 2:19not be existential for us; it may be
  62. 2:21more financial for them, more political
  63. 2:24for those who sign the licenses. Stay
  64. 2:28tuned, because we are going to turn the
  65. 2:29official message around, and today it
  66. 2:31is going to be difficult, because today
  67. 2:33we are going to see how or why the
  68. 2:35accounts of this industry still do not
  69. 2:37add up and what exactly is in it for
  70. 2:39those who ask to slow down; why safety
  71. 2:41regulation is the best possible deal
  72. 2:42for those who are already in the lead
  73. 2:44and the worst for those who come after,
  74. 2:46and what part of the spoils a state, an
  75. 2:48administration, a union is going to
  76. 2:50take in this distribution and, above
  77. 2:52all, how much is all this going to cost
  78. 2:54you? And we are going to do so by
  79. 2:57trying to analyze this hypothesis
  80. 2:59before accepting it as it is being done
  81. 3:02. Anyway, remember that you have three
  82. 3:03ways to support this channel by
  83. 3:05commenting and watching the video until
  84. 3:06the end, which is what pushes the
  85. 3:07algorithm, because in these videos they
  86. 3:09usually push it very little.
  87. 3:10Subscribing, which is free, or becoming
  88. 3:12a member for very little. You know that
  89. 3:14my team will be especially happy to do
  90. 3:16so, although any of the three options
  91. 3:17will seem fine to us. As I said, don't
  92. 3:19go away because today's news will also
  93. 3:21interest you. Let's go with what we
  94. 3:29know. A certain Jacob Coxson, who first
  95. 3:32worked at Openi and then at Anthropic,
  96. 3:34left the latter company a few days ago
  97. 3:36and issued a very serious warning. He
  98. 3:41maintains that a good part of those who
  99. 3:43are building the most advanced models
  100. 3:45sincerely believe that there is a
  101. 3:46significant probability that future
  102. 3:48systems will escape the control of
  103. 3:50human beings. And there is a detail
  104. 3:53that makes it difficult to dismiss him
  105. 3:56as an angry ex-employee, because Coxson
  106. 3:58left before consolidating part of his
  107. 4:00economic participation, that is, he had
  108. 4:02money on the table to stay and still
  109. 4:05this man decided to leave. Therefore,
  110. 4:07man, we have to take that into account.
  111. 4:10His diagnosis may even be wrong, but
  112. 4:12despite everything, his concern can be
  113. 4:14taken seriously, I'm not saying it
  114. 4:16isn't. And he's not alone. Other
  115. 4:19security researchers have come out of
  116. 4:21frontier labs or have publicly
  117. 4:22expressed similar concerns. Anthropic,
  118. 4:26Openei and the rest are already making
  119. 4:28specific assessments on cybersecurity,
  120. 4:30biology, chemical weapons, deception
  121. 4:33capacity, loss of control and autonomy
  122. 4:35of agents. To say that there are real
  123. 4:39risks there would be as absurd as those
  124. 4:41who claim that we already know that a
  125. 4:43superintelligence is going to end
  126. 4:44humanity by 2030, as they are saying,
  127. 4:46by the way. And there have been serious
  128. 4:49incidents that must be taken into
  129. 4:50account. The Haging Face incident is
  130. 4:53the most striking, we explained it here
  131. 4:54a few weeks ago in a previous video.
  132. 4:57During internal cybersecurity
  133. 4:58assessments, several Openi agents found
  134. 5:01vulnerabilities in the infrastructure
  135. 5:03that was supposed to keep them isolated
  136. 5:05. They set up communication channels
  137. 5:07that no one had anticipated and, in
  138. 5:09addition, these algorithms gained
  139. 5:11access to the internet and ended up
  140. 5:12compromising third-party platform
  141. 5:14systems. It is a serious matter, of
  142. 5:18course, very serious, but it is worth
  143. 5:20looking at what it demonstrates and
  144. 5:22what it does not demonstrate, because
  145. 5:23it does not demonstrate that an
  146. 5:25artificial intelligence woke up or
  147. 5:27became aware and decided to escape from
  148. 5:29something. Openi's own technical report
  149. 5:33, which many of those who speak about
  150. 5:35it seem not to have read in detail,
  151. 5:37describes models subjected to computer
  152. 5:39exploitation tasks with safeguards
  153. 5:41purposely reduced, optimizing
  154. 5:43objectives within a flawed environment
  155. 5:45and finding paths that the researchers
  156. 5:47had not anticipated. Mathematics
  157. 5:51optimizing an objective function on a
  158. 5:53poorly closed infrastructure. That is
  159. 5:56exactly why the problem is serious. Yes
  160. 5:59. Yes. Although it is not what they say
  161. 6:02, it is a serious issue. You do not
  162. 6:04need to attribute a soul to a matrix
  163. 6:06for a highly competent machine to cause
  164. 6:08a disaster. An algorithm does not need
  165. 6:11to hate you to hack your server. It is
  166. 6:14enough for it to do so in the most
  167. 6:15effective way to achieve the objective
  168. 6:17that you have given it. The problem
  169. 6:20begins when we move from the risk to
  170. 6:22what is now filling all the news in the
  171. 6:25world. The story. A technically
  172. 6:27plausible possibility first becomes a
  173. 6:30journalistic probability, then a
  174. 6:32headline, then a prophecy, and finally
  175. 6:35a political demand. Be careful with
  176. 6:36that. The media steals the attention.
  177. 6:39The disaster reporter steals the
  178. 6:41relevance, the security expert steals
  179. 6:43the public spotlight, and organizations
  180. 6:45dedicated to risk take the funding.
  181. 6:48Even governments get a frightened
  182. 6:50public opinion that begs them to do
  183. 6:52something, to stop this at all costs,
  184. 6:54for our own good. And the big labs get
  185. 6:58the most valuable thing of all, which
  186. 6:59is the position from which to explain
  187. 7:01to that government exactly what should
  188. 7:02be regulated. It's their tool, it's
  189. 7:05their key. None of that shows bad faith
  190. 7:07. I'm not saying it doesn't. I'm not
  191. 7:08going to suspect anyone now, and that's
  192. 7:10why it works too. Economic incentives
  193. 7:12don't need a secret room where someone
  194. 7:14decides on a plan. While we're
  195. 7:17discussing whether an AI had become
  196. 7:19Skynet, behind the screen something
  197. 7:21else was happening that was much less
  198. 7:23cinematic and actually much more urgent
  199. 7:26for them. Because this, my friends,
  200. 7:29those of you who follow this channel
  201. 7:32know that all of this has a price. The
  202. 7:35monster also has a bill. On September
  203. 7:3910, AENA, Openei did something very
  204. 7:42strange. It temporarily suspended new
  205. 7:45registrations and upgrades to the Chat
  206. 7:47GPT Pro plan for its $ 200 subscription
  207. 7:50. And it didn't do it because of a lack
  208. 7:52of customers. On the contrary, the
  209. 7:55demand for Astra, which is its most
  210. 7:57advanced model, was putting such
  211. 7:58pressure on its computing capacity that
  212. 8:00the company decided to stop accepting
  213. 8:02some of the customers who were willing
  214. 8:04to pay it the most for a while. Think
  215. 8:10about how strange this is. You have a
  216. 8:13product that people want to buy. You
  217. 8:16have customers who offer you $ 200 each
  218. 8:18month, which is much more than what
  219. 8:19they were paying. And you have to close
  220. 8:22half the door because you can't serve
  221. 8:25them. Openi has too many customers and
  222. 8:27not enough manufacturing. And that
  223. 8:29manufacturing is no longer a room with
  224. 8:32servers, it's land, data centers, power
  225. 8:34plants, substations, cooling, fiber,
  226. 8:37transformers and hundreds of thousands
  227. 8:40of accelerators working at the same
  228. 8:42time. OpenAI has increased its cloud
  229. 8:47spending plans to, hold on, seven
  230. 8:50hundred and fifty billion dollars by
  231. 8:532030. Anthropic is not playing around
  232. 8:59with pocket change either. It has
  233. 9:01signed up to 5 GW of capacity with
  234. 9:03Amazon and technology commitments for
  235. 9:05more than $ 100 billion over the next
  236. 9:07decade. It has expanded its agreement
  237. 9:10with Google, with Broadcom for several
  238. 9:12additional gigawatts. It maintains a
  239. 9:15partnership with Microsoft and Nvidia
  240. 9:17that includes tens of billions of
  241. 9:19aureate capacity and has added more
  242. 9:21than 220,000 GPUs through a deal with
  243. 9:23SpaceX. They are all in the same boat.
  244. 9:26If you think about it, software
  245. 9:28companies are financing infrastructure
  246. 9:30on the scale of the energy industry. In
  247. 9:33traditional software, the dream was
  248. 9:35operating leverage, that is, building
  249. 9:38the product was expensive, but selling
  250. 9:40the millionth copy was almost nothing.
  251. 9:44At the cutting edge of artificial
  252. 9:45intelligence, accounting works, well,
  253. 9:47the other way around. Maybe it's
  254. 9:50related, you will see. Revenues can
  255. 9:53skyrocket as the capital needed to
  256. 9:54produce the next dollar of revenue
  257. 9:56skyrockets. More customers means more
  258. 9:59inference. More sophisticated models
  259. 10:02can consume more compute. Staying at
  260. 10:05the cutting edge, at the cutting edge,
  261. 10:08requires training new generations, and
  262. 10:10each leap requires securing chips,
  263. 10:12electricity, and more capacity before
  264. 10:14you know how long you're going to keep
  265. 10:16the advantage you just bought. Maybe
  266. 10:21that's also why they want to ease off
  267. 10:23the gas. Could be, could be. Anthropic
  268. 10:26reported in May that its revenue run
  269. 10:28rate had surpassed $ 47 billion, which
  270. 10:31is impressive. Open is also growing at
  271. 10:33a rate that we've never seen before,
  272. 10:35but growing at a high rate right now
  273. 10:37doesn't answer the financial question
  274. 10:38that matters. The question is, how much
  275. 10:41capital do you have to invest and
  276. 10:43reinvest to keep growing at that rate?
  277. 10:45And there's no public answer to that
  278. 10:47right now. Nobody has yet demonstrated
  279. 10:49how much return each dollar invested in
  280. 10:50that computation is. Nobody. We have to
  281. 10:54be precise about this because this is
  282. 10:56not going to happen, this is very
  283. 10:58easily exaggerated. These companies
  284. 11:00aren’t broke, they’re trapped,
  285. 11:02which is different. They have plenty of
  286. 11:04demand, revenue growing at a historic
  287. 11:06rate, that’s true, and a business
  288. 11:07model that has yet to prove that it
  289. 11:09pays for itself. With each passing
  290. 11:12quarter in this race, the investment
  291. 11:14gap needed to stay ahead grows faster
  292. 11:16than revenue itself. That’s a
  293. 11:20terrible position for the guy in charge
  294. 11:22, because he can’t stop without
  295. 11:23handing the lead to the guy next to him
  296. 11:25, unless they all stop at once, unless
  297. 11:27they all ease off the gas, and unless
  298. 11:29there’s a reason for stopping that no
  299. 11:31one can argue with. Why would they all
  300. 11:35ease off? Open raised $ 122 billion in
  301. 11:39committed capital in March at a
  302. 11:41post-money valuation of $ 852 billion.
  303. 11:44It’s already approaching a trillion.
  304. 11:46Anthropic raised another $ 65 billion
  305. 11:48in May, reaching a valuation of $ 965
  306. 11:50billion, almost a trillion. These are
  307. 11:54amounts that would have seemed
  308. 11:55impossible in the private market three
  309. 11:56years ago, and that breaks one of the
  310. 11:58historical rules of technological
  311. 11:59capitalism. For decades, a company
  312. 12:02would end up going public because it
  313. 12:04needed to access a much larger pool of
  314. 12:07capital than it had available in
  315. 12:09private equity. But what happens when
  316. 12:12you can raise $ 122 billion without
  317. 12:14going public? Well, the stock market
  318. 12:17ceases to be an immediate necessity,
  319. 12:19and when it ceases to be a necessity,
  320. 12:21you can also postpone something that in
  321. 12:23that situation is frankly very annoying
  322. 12:26. Watch out, millions of investors are
  323. 12:30deciding every day how much your
  324. 12:32business is really worth. And that is
  325. 12:35what they have just postponed, what an
  326. 12:38IPO forces us to show. Let's leave
  327. 12:41Silicon Valley for a moment, away from
  328. 12:43all this mess, and go to Wall Street,
  329. 12:45where interesting things are happening
  330. 12:47there, because right now there are
  331. 12:49three things that are constantly mixed
  332. 12:51together and they are not the same.
  333. 12:54Need for capital, need for liquidity,
  334. 12:56and need for price discovery. Let me
  335. 12:58explain. A company needs capital to
  336. 13:01finance its growth. Your shareholders,
  337. 13:04founders or employees may need
  338. 13:05liquidity to convert part of their
  339. 13:07stakes into cash. But the market needs
  340. 13:11something very different, which is to
  341. 13:12find out how much a company is really
  342. 13:14worth, right? Openei and Anthropic have
  343. 13:18shown that they can raise absolutely
  344. 13:20gigantic amounts of capital without
  345. 13:22going public, and there are also
  346. 13:23secondary sales that give liquidity to
  347. 13:25certain shareholders without having to
  348. 13:27go through an IPO. What is very
  349. 13:31difficult to replicate here off the
  350. 13:33stock exchange is permanent price
  351. 13:35discovery. That is, thousands of
  352. 13:38managers, analysts and investors
  353. 13:40looking at your accounts, comparing
  354. 13:42expectations with results and repricing
  355. 13:44your company every second. Open
  356. 13:48confidentially filed its draft S1 form,
  357. 13:50which is the one that has to be filed
  358. 13:52with the SEC, the regulators there. It
  359. 13:56did so on June 8. Anthropic had done
  360. 13:58the same a week earlier and an S1 is
  361. 14:00less spectacular than a discussion
  362. 14:02about the end of humanity, right? But
  363. 14:05it can end up telling us a lot more
  364. 14:07about this industry because it requires
  365. 14:10a detailed description of the business,
  366. 14:12the financial situation, the results
  367. 14:14and the risks, and also the audited
  368. 14:16financial statements. And be careful,
  369. 14:19being private, you can't fool your
  370. 14:21investors either. Obviously, SoftBank,
  371. 14:23Amazon, Nvidia or Seollya do their due
  372. 14:25diligence before signing a check.
  373. 14:26Obviously, there's a difference. It's
  374. 14:29one thing to open your numbers to a few
  375. 14:31dozen sophisticated investors with whom
  376. 14:32you're negotiating a round. It's quite
  377. 14:35another to put them in front of the
  378. 14:37entire market and let Wall Street ask
  379. 14:40you every quarter if you really justify
  380. 14:42or are really worth $ 800 billion, $ 1
  381. 14:45trillion or $ 2 trillion. And the
  382. 14:47questions are going to be very tough
  383. 14:49from that point on. How much does it
  384. 14:51really cost to train a generation of
  385. 14:53artificial intelligence models? How
  386. 14:55much does inference cost once you
  387. 14:57account for all that infrastructure?
  388. 15:00What gross margin do you have left?
  389. 15:01What future cloud commitments have
  390. 15:03already been signed? How much Capex
  391. 15:06does it require to stay at the cutting
  392. 15:08edge? What part of the growth requires
  393. 15:10raising capital again? These are
  394. 15:12questions, are they going to be asked?
  395. 15:14When will sustainable free cash flow
  396. 15:17appear? And the most important one of
  397. 15:19all, which they're also going to ask,
  398. 15:20what return does each additional dollar
  399. 15:22I invest in computing produce? That's
  400. 15:26the real economy, the real economy of
  401. 15:27what's called frontier artificial
  402. 15:29intelligence. And we still can't see it
  403. 15:32as clearly as we've been able to study
  404. 15:34Microsoft, Alphabet, Meta. And
  405. 15:37Anthropic may soon give us that window.
  406. 15:39It's true, it's still preparing a
  407. 15:41possible IPO that, according to what
  408. 15:44has been published, could raise up to $
  409. 15:47100 billion with a valuation close to
  410. 15:49...Yes, yes, $ 2 trillion. It's worth
  411. 15:53clarifying this because these figures
  412. 15:54are being thrown around too happily.
  413. 15:56Anthropic doesn't need to raise $ 2
  414. 15:58trillion, that would be its valuation.
  415. 16:02What it would raise would be $ 100
  416. 16:03billion, and a company valued at $ 2
  417. 16:05trillion would have to answer something
  418. 16:06quite simple. What future benefits
  419. 16:11justify paying $ 2 trillion for it
  420. 16:13today? Well, Fortune magazine did the
  421. 16:15calculation using the usual multiples
  422. 16:17of the big tech companies on the NASDAQ
  423. 16:19itself, where they are going to be
  424. 16:21listed or would like to be listed or
  425. 16:23say they are going to be listed. And
  426. 16:26the figure that comes out is that
  427. 16:27Anthropic would have to end up
  428. 16:28generating tens of billions of dollars
  429. 16:30of annual profit to sustain that
  430. 16:31capitalization and today it barely
  431. 16:33makes any money. In that case, showing
  432. 16:36growth is not enough. The market wants
  433. 16:39to know how much additional capital
  434. 16:40needs to be invested to achieve that
  435. 16:42growth. Openai is in a different
  436. 16:44position. When it announced its
  437. 16:47confidential S1 document, that document
  438. 16:49, it already said that it had not
  439. 16:51decided exactly when to go public and
  440. 16:53that there were things that would
  441. 16:55probably be easier for it as a private
  442. 16:57company than being in the public domain
  443. 17:00at the expense of all opinion. And in
  444. 17:03August, CFO Sara Frear said internally
  445. 17:05that Open would list in 2027. That's
  446. 17:07where we'd go, although she had left
  447. 17:10open an earlier exit if the business
  448. 17:12picked up speed, and now, in the midst
  449. 17:14of the explosion of security concerns,
  450. 17:16Samman has already said that 2026 is a
  451. 17:18very bad time to go public, because
  452. 17:20things are so bad, arguing that a
  453. 17:22public company would have a harder time
  454. 17:24right now making security decisions
  455. 17:26that would temporarily harm its
  456. 17:28shareholders. In other words, he says
  457. 17:31that if they have to take their foot
  458. 17:33off the gas, it wouldn't exactly be a
  459. 17:35good time to go public. That makes
  460. 17:37sense. I would say yes. It shows that
  461. 17:40security, uh, well, is an excuse to
  462. 17:42dodge Wall Street in this case, right?
  463. 17:45But going private is financially very
  464. 17:47comfortable when you've just raised $
  465. 17:49122 billion and your capital needs are
  466. 17:52colossal. Not because you have anything
  467. 17:55sinister to hide, but because the
  468. 17:57private market lets you negotiate
  469. 17:59valuation and narrative with a small
  470. 18:01group of investors, while in public you
  471. 18:03would be forced to live with a daily
  472. 18:05price and quarterly results capable of
  473. 18:07dismantling your story in an afternoon.
  474. 18:12And I'm going to correct my own thesis
  475. 18:14precisely so that all this is more
  476. 18:16bearable. It is not true that Open and
  477. 18:19Anthropic have jointly decided to delay
  478. 18:21their IPOs, their stock market launches
  479. 18:23, until 2027. Openi is targeting 2027,
  480. 18:25yes, but Anthropic is still preparing a
  481. 18:27possible launch this fall. Therefore,
  482. 18:32in principle there is no coordination,
  483. 18:34but I don't need that to demonstrate
  484. 18:35what I want to demonstrate. What is
  485. 18:38interesting is another coincidence.
  486. 18:41Just when we are about to begin to see
  487. 18:43the real economics of these
  488. 18:44laboratories in public, their top
  489. 18:46executives are beginning to say that
  490. 18:47the race may be moving too fast. And if
  491. 18:52the race is regulated by law, the
  492. 18:54question ceases to be who runs the
  493. 18:56fastest and becomes who can pay for the
  494. 18:59permit to run, who can pay for
  495. 19:01compliance, and here comes the best
  496. 19:04argument against my hypothesis. I know,
  497. 19:07I have to launch it and it has to be
  498. 19:09put on the table. Anthropic is not
  499. 19:12proposing that any programmer who sets
  500. 19:14up an AI startup tomorrow has to hire
  501. 19:16500 lawyers and 20 auditors. It doesn't
  502. 19:18say that. The border regulations that
  503. 19:20the company defends use high thresholds
  504. 19:22. It's true. California Law SB53
  505. 19:25considers a major frontier developer to
  506. 19:27be someone who works with models
  507. 19:29trained above 10 to the power of 26
  508. 19:31computational operations, which is
  509. 19:33absolutely outrageous. and belongs to a
  510. 19:35group with more than $ 500 million in
  511. 19:37annual revenue. That's the official
  512. 19:40version. New York has ended up
  513. 19:42converging towards similar thresholds
  514. 19:44in its race race act, which is in
  515. 19:47theory the obligations that fall on the
  516. 19:49giants. Well, that design makes sense
  517. 19:52and dismantles the caricature according
  518. 19:54to which these companies want to
  519. 19:56prevent anyone from setting up a
  520. 19:58startup. I have defended it, not like
  521. 20:00that, but I have said that these people
  522. 20:03are basically trying to keep them alone
  523. 20:05. But the question does not disappear
  524. 20:07with this evidence, it is displaced
  525. 20:09because a barrier to entry does not
  526. 20:11have to prevent you from entering. They
  527. 20:13say they won't stop anyone from getting
  528. 20:14in. It can stop you from growing. You
  529. 20:17can start your company, raise $ 20
  530. 20:19million, then $ 100 million. It's the
  531. 20:21same, train better and better models
  532. 20:23and compete for years. The problem
  533. 20:24comes just when you start to be a
  534. 20:26serious threat. That's when you cross
  535. 20:29the threshold and start having to deal
  536. 20:31with auditing, reporting, evaluation,
  537. 20:33security protocols, external
  538. 20:35supervision, documentation, internal
  539. 20:37systems, regulatory relationships, and
  540. 20:39they may all be necessary. That's not
  541. 20:42the issue. The issue is who's going to
  542. 20:44be able to pay for them. Let's travel
  543. 20:47back to 2002 for a moment, as it will
  544. 20:49help us understand many things. You'll
  545. 20:51see. After Enron and Walcom, the United
  546. 20:54States passed the Sabarness Hoxley Act
  547. 20:56to strengthen controls, audits and the
  548. 20:58quality of financial information. Well,
  549. 21:01there were plenty of reasons to do it
  550. 21:03and it had an economic effect that has
  551. 21:05since been studied to death. A good
  552. 21:08part of the regulatory costs were fixed
  553. 21:12. The Government Accountability Office
  554. 21:15found that these costs weigh
  555. 21:16proportionally much more heavily on
  556. 21:18small companies than on large ones.
  557. 21:21These guys have studied that, you'll
  558. 21:22see. That's what happens in banking.
  559. 21:26When you need a certain compliance
  560. 21:27structure, the giant spreads it across
  561. 21:29tens of billions of revenue. But the
  562. 21:33little guy has to support almost the
  563. 21:35same infrastructure on a business that
  564. 21:37is actually minuscule compared to the
  565. 21:39big guys, the ones that were already
  566. 21:41there. I'm not saying that all
  567. 21:43regulation is a corporate conspiracy.
  568. 21:46Their work shaped the economic theory
  569. 21:48of regulatory capture. Regulations
  570. 21:50generate costs and benefits, and
  571. 21:52whoever has more money at stake also
  572. 21:54has more incentive to influence how
  573. 21:56they're designed. There's even a common
  574. 21:59expression in business strategy called
  575. 22:01regulatory mod. It's the regulatory
  576. 22:04moat. A regulation can be perfectly
  577. 22:06justified for safety reasons while also
  578. 22:08reinforcing the competitive advantage
  579. 22:10of whoever already has the resources to
  580. 22:12comply with it. And we're starting from
  581. 22:15a brutally concentrated market.
  582. 22:17According to Peachbook data compiled by
  583. 22:21Axios, Open AI, and Anthropic, they
  584. 22:23together took more than 60%of all
  585. 22:26venture capital committed to US
  586. 22:29startups during the first half of 2026
  587. 22:32from all startups. Anthropic has also
  588. 22:35significantly increased its political
  589. 22:37activity. Just look, over $ 3.5 million
  590. 22:40in federal funding in the first half of
  591. 22:43the year, on top of everything it spent
  592. 22:46in 2025, and now $ 40 million
  593. 22:48contributed to the public for public
  594. 22:50education and transparency policies and
  595. 22:53AI safeguards. And Anthropic itself
  596. 22:56specifies that this money cannot be
  597. 22:58used to support candidates. Perfect.
  598. 23:01We're not talking about corruption, I'm
  599. 23:03not. We're talking about a company that
  600. 23:06is legitimately trying to influence the
  601. 23:09institutional architecture of an
  602. 23:11industry where it has hundreds of
  603. 23:13billions at stake. And none of that is
  604. 23:17a conspiracy. This is a company
  605. 23:22defending what it invested in, the
  606. 23:24incentives. When this becomes a battle
  607. 23:27between doomsday believers and doomsday
  608. 23:30deniers, the only thing that really
  609. 23:32matters here is lost. Anthropic may be
  610. 23:34genuinely concerned regarding the
  611. 23:36potential risks posed by artificial
  612. 23:37intelligence. I'm not saying it isn't,
  613. 23:39but at the same time it could benefit
  614. 23:41from regulation that is very expensive
  615. 23:44for new entrants. It is possible for
  616. 23:46both things to be true at the same time
  617. 23:47, right? An economic incentive does not
  618. 23:51override a moral concern, and a moral
  619. 23:53concern does not make the economic
  620. 23:55incentive disappear. Not when slowing
  621. 23:58down is a bailout. We come to September
  622. 24:0212th. Dario Amodei publishes Frontier's
  623. 24:04Wimo Space. The central idea is that
  624. 24:07capacity building is moving faster than
  625. 24:09safety, and that the industry needs to
  626. 24:10keep up with the frontier. Anthropic
  627. 24:13commits to bringing in external
  628. 24:14evaluators with very deep access to its
  629. 24:16internal processes. And he also asks,
  630. 24:19hey, that the other big labs do
  631. 24:20something similar, they're not going to
  632. 24:22be the ones to do it alone. Amodei then
  633. 24:25proposes that the frontier companies,
  634. 24:27which are right there on the edge,
  635. 24:29coordinate standards and some speed
  636. 24:31limits, preferably through public
  637. 24:33regulation. Public regulation, and he
  638. 24:38acknowledges that some types of
  639. 24:40coordination could conflict with
  640. 24:42competition law, so he proposes some
  641. 24:44limited mechanism to allow for those
  642. 24:46conversations. And the guy ends with an
  643. 24:49international element. He says that any
  644. 24:51major slowdown in the United States is
  645. 24:53meaningless if China can continue to
  646. 24:55advance at its own pace, that the
  647. 24:56Chinese would have to be convinced.
  648. 25:00That the Chinese would have to be
  649. 25:01convinced, he says.
  650. 25:01And then Samman comes out to support
  651. 25:03the direction of all this. Elon was
  652. 25:05more in favor there, one after the
  653. 25:08other. We have to proceed calmly
  654. 25:09because it is very easy to jump from a
  655. 25:11coincidence to a conspiracy and there
  656. 25:13is no need to imagine any secret
  657. 25:14meeting, as he said when Modei himself
  658. 25:16is proposing in public that certain
  659. 25:17competitors coordinate, nothing happens
  660. 25:19. There explaining why he needs legal
  661. 25:21cover to do so. So far so good, there
  662. 25:23may be a totally legitimate reason. I
  663. 25:25am not saying there is not. Mancur
  664. 25:26Olson explained 60 years ago why a
  665. 25:29group with common interests often fails
  666. 25:31to act in its own interest. He said
  667. 25:35that what is good for the whole
  668. 25:37sometimes does not suit any of the
  669. 25:38members individually. And that is
  670. 25:41exactly what is happening here. If each
  671. 25:45laboratory initially thinks that
  672. 25:47slowing down on its own is dangerous
  673. 25:49because the rival will take advantage
  674. 25:51of the pause to overtake it, then it is
  675. 25:54normal that everyone may prefer a
  676. 25:56slower race and none can afford to slow
  677. 25:58down first, right? A common standard
  678. 26:02would eliminate part of that problem
  679. 26:04and could truly improve everyone's
  680. 26:07safety, and in that way these CEOs will
  681. 26:10save the world. Always helped,
  682. 26:13obviously, by appropriate regulation.
  683. 26:15But now let's read it in reverse. What
  684. 26:17happens financially when everyone slows
  685. 26:18down at the same time and it's also
  686. 26:20written in an official bulletin? Well,
  687. 26:22the burn rate goes down. An unlimited
  688. 26:25race forces you to continually train
  689. 26:27the next model because you don't know
  690. 26:28when your competitor will introduce
  691. 26:30something that will make yours obsolete
  692. 26:31. Well, if you slow down, you lengthen
  693. 26:35the training cycles and make the same
  694. 26:37capital last longer. The same capital
  695. 26:40that you've been given lasts longer.
  696. 26:42The marginal pressure on Capex and what
  697. 26:44it computes also goes down. Investments
  698. 26:47don't disappear, but the logic is no
  699. 26:49longer to buy all the capacity possible
  700. 26:51before your rival takes it away. The
  701. 26:53path they travel is also lengthened
  702. 26:55because the same 50,000 or 100,000
  703. 26:57million finances more months of
  704. 26:59operation and allows each generation of
  705. 27:01models to be amortized for a longer
  706. 27:03time. The mad race forces them to slow
  707. 27:07down. If you've just invested a fortune
  708. 27:09in training and deploying a system and
  709. 27:11six months later you have to replace it
  710. 27:13with an even more expensive one, the
  711. 27:15payback period for all that is
  712. 27:16ridiculous. A slower pace lets you
  713. 27:19squeeze out what you've already built
  714. 27:20for longer. And here's something even
  715. 27:23more subtle, uh, which is where I want
  716. 27:25to focus. A company that slows down
  717. 27:27because the numbers don't add up is
  718. 27:29admitting weakness. But a company that
  719. 27:32slows down because the safety of
  720. 27:34humanity demands it is demonstrating
  721. 27:37responsibility. It's the same
  722. 27:39accounting move, but with two very
  723. 27:41different headlines. Regulation turns a
  724. 27:44cash flow problem into a public virtue
  725. 27:46and turns a decision that would
  726. 27:48penalize your valuation into one that
  727. 27:50protects it. They're good guys. And I
  728. 27:54didn't make this up, uh? And it's not
  729. 27:56new either. There's a historian,
  730. 27:59Gabriel Colco, who documented something
  731. 28:01similar in the American railroads of
  732. 28:03the late 19th century. These guys have
  733. 28:05probably read it. An industry with
  734. 28:08enormous fixed costs, price wars that
  735. 28:09broke everyone, and competition that no
  736. 28:11one could keep up with. That industry,
  737. 28:15that railroad, where the big companies
  738. 28:17failed to stabilize the market due to
  739. 28:19private agreements that were
  740. 28:20continually broken. They ended up
  741. 28:24supporting federal intervention. The
  742. 28:28regulation that the official story
  743. 28:30tells as a victory of the public
  744. 28:32interest against monopolies was
  745. 28:34actually largely the mechanism by which
  746. 28:37those companies incapable of sustaining
  747. 28:39competition achieved by law the
  748. 28:41stability that the market was denying
  749. 28:43them. Now replace railroads with data
  750. 28:48centers, and replace price wars with an
  751. 28:50intense race for training. And hey, you
  752. 28:53have an identical map. History rhymes.
  753. 28:59If the slowdown is also accompanied by
  754. 29:01audits, verifiers, certifications,
  755. 29:03compliance departments, security
  756. 29:05controls, mandatory reports,
  757. 29:07sophisticated relationships with
  758. 29:09governments, then every additional mile
  759. 29:11is more expensive for the next guy. The
  760. 29:15giant already has lawyers, it already
  761. 29:16has hundreds of security researchers,
  762. 29:18it already has teams to fix all that,
  763. 29:20it already has a direct line to
  764. 29:22Washington, it already has gig workers
  765. 29:24hired, and it already has investors
  766. 29:26capable of writing checks for tens of
  767. 29:28billions. But the challenger, the one
  768. 29:31who could become a good competitor, has
  769. 29:34to build all that while trying to catch
  770. 29:36up. And that's where security and
  771. 29:39finances stop being incompatible
  772. 29:40explanations. A general slowdown can be
  773. 29:43justified by genuine security reasons.
  774. 29:46Yes. Yes. And at the same time be the
  775. 29:49best possible rescue for the financial
  776. 29:51structure of those leading that race.
  777. 29:54Sure. And therefore you don't need
  778. 29:57Amodei, Alman, or Mask to sit in a room
  779. 30:00to decide that. You just need their
  780. 30:03incentives to converge somewhere. Now
  781. 30:06look at the calendar again. It's very
  782. 30:08important. Openi presents its S1 in
  783. 30:12June. In August it talks about 2027. In
  784. 30:14September, in the midst of a wave of
  785. 30:16concerns about security, Alman explains
  786. 30:18that this would not be a good time to
  787. 30:20list. That doesn't prove that the
  788. 30:23security argument is false, but it
  789. 30:25produces an alignment of interests that
  790. 30:26, man, would at least deserve some
  791. 30:28attention from those media outlets that
  792. 30:30are talking about how the end of the
  793. 30:31world is coming because of AI and that
  794. 30:33it must be stopped, regulated, and
  795. 30:34something done. Staying private longer
  796. 30:39allows you to face a massive investment
  797. 30:41phase without having Wall Street asking
  798. 30:43you every quarter how much you’re
  799. 30:45burning and whether an $ 852 billion
  800. 30:47valuation still makes sense. If
  801. 30:50Anthropic does go public with its S1,
  802. 30:52we’ll finally get a wide window into
  803. 30:54the plumbing of a cutting-edge lab.
  804. 30:58Margins, costs, compute commitments,
  805. 31:00capital requirements, cash flow—and
  806. 31:03then we can start to answer what’s
  807. 31:05still very much hidden. How much does
  808. 31:10it really cost to manufacture the
  809. 31:11intelligence that promises to transform
  810. 31:13the world? Because we don’t know yet.
  811. 31:18So yes, so far this explains what the
  812. 31:20company is in for. We’re missing the
  813. 31:22other partner, the other partner in the
  814. 31:25deal, the one who hasn’t put up a
  815. 31:27dollar of his own and is taking the
  816. 31:29lion’s share. Who gets to hold the
  817. 31:31key? Let’s start with the argument
  818. 31:34that justifies it all. China. Amodei
  819. 31:36acknowledges as much himself. Besides,
  820. 31:37a Western slowdown only works as long
  821. 31:39as it doesn’t hand a geopolitical
  822. 31:41competitor a decisive advantage. That's
  823. 31:44why Anthropic advocates both curbing
  824. 31:47certain capabilities in the United
  825. 31:49States and tightening controls on
  826. 31:51advanced chips to limit Chinese access
  827. 31:53to that border infrastructure and
  828. 31:55combat what they call unauthorized
  829. 31:57distillation by smaller labs that can
  830. 31:59extract capabilities from Western
  831. 32:01models. Let me clarify. The logic in
  832. 32:06Washington is obvious. If you're going
  833. 32:09to hit the brakes, make sure the guy
  834. 32:11behind you doesn't seize the chance to
  835. 32:14pass. And they think that guy behind
  836. 32:16you is China. But look at what's being
  837. 32:19built with this justification. Export
  838. 32:22controls on chips, restrictions on data
  839. 32:24centers, audit systems, independent
  840. 32:27evaluators, computation thresholds,
  841. 32:29model registries, certifications,
  842. 32:31controls on the weights of all this,
  843. 32:33verification of certain training,
  844. 32:36agreements between governments. The
  845. 32:39thing is that each piece, each piece
  846. 32:41separately, can have a reasonable
  847. 32:43justification. I'm not saying it
  848. 32:45doesn't, but all together they are a
  849. 32:47licensing regime for computing capacity
  850. 32:49. And that's where the second
  851. 32:51beneficiary of this story comes in, who
  852. 32:54by the way has been waiting for decades
  853. 32:56, because there is a technology that
  854. 32:58states have not been able to fully
  855. 33:00control since it existed and that is
  856. 33:03the one that allows people to produce,
  857. 33:05copy and distribute information without
  858. 33:07asking permission. Remember, look, they
  859. 33:11arrived late with the internet, they
  860. 33:13arrived late with social networks, they
  861. 33:15have been fighting and losing with
  862. 33:17encryption for 30 years, with
  863. 33:19artificial intelligence they have
  864. 33:21arrived on time for the first time and
  865. 33:23they have arrived, moreover, invited by
  866. 33:25the industry itself. What's happening?
  867. 33:28Nobody is telling you this. Charles
  868. 33:30Tily put it in a way that is very
  869. 33:32unpleasant for any government. State
  870. 33:35building is similar to the sale of
  871. 33:37protection with the particularity that
  872. 33:39whoever offers to protect you is
  873. 33:41usually also the one who defines what
  874. 33:42you have to protect yourself from.
  875. 33:45There is no need to believe that
  876. 33:46governments are evil, no. Just look at
  877. 33:49what each one takes. If the risk of AI
  878. 33:53is existential, then whoever grants
  879. 33:55licenses decides who can train a model,
  880. 33:57with what data to do it, up to what
  881. 33:59computational threshold it can be
  882. 34:00carried and with what content it can be
  883. 34:03blocked. We're no longer talking about
  884. 34:06computer security; here we're talking
  885. 34:08about who approves the capabilities of
  886. 34:09a country's cognitive infrastructure.
  887. 34:12And we don't have to go to China to see
  888. 34:14where all this is heading. In Europe,
  889. 34:17we've been seeing for years how
  890. 34:19perfectly presentable security
  891. 34:21arguments end up being converted into
  892. 34:24control infrastructures. Everything is
  893. 34:27always for our own good. Whenever
  894. 34:30there's a danger to humanity, a
  895. 34:33regulation appears.
  896. 34:37Against CO2, bottle caps. Protecting
  897. 34:39minors ended in proposals to scan
  898. 34:41everyone's private communications.
  899. 34:43Fighting fraud ended in mandatory
  900. 34:45digital identity. Combating
  901. 34:46disinformation ended in mechanisms that
  902. 34:48decide which content deserves
  903. 34:49visibility. In no case was an evil plan
  904. 34:52necessary. A legitimate reason, a
  905. 34:54frightened public opinion, and an
  906. 34:56administrative opportunity were enough.
  907. 34:59Now add it all up, everything,
  908. 35:00everything, everything. Four or five
  909. 35:02companies with enough capital,
  910. 35:04computing, and lawyers to comply. A
  911. 35:06handful of governments, all of them,
  912. 35:08all of them, with the power to decide
  913. 35:10who comes in and who goes out. and a
  914. 35:12technology that will cut across
  915. 35:14productivity, research, software,
  916. 35:16business automation, and a huge chunk
  917. 35:18of the work that you and I do. We will
  918. 35:23have reduced the likelihood that anyone
  919. 35:25can build an intelligence that is
  920. 35:27extremely dangerous to everyone, and we
  921. 35:29will have concentrated the ability to
  922. 35:31build the most powerful machine in the
  923. 35:33smallest group we’ve ever had at the
  924. 35:35helm of a general-purpose technology.
  925. 35:39And with the key to those who claim to
  926. 35:42be watching over us. And my criticism,
  927. 35:46mind you, is not to say that Dario
  928. 35:48Amodei is lying. I’m not going to say
  929. 35:50that because it’s not true. He
  930. 35:51probably genuinely believes what he’s
  931. 35:52saying. And Jacob Coxson, whoever he is
  932. 35:55, and so do the researchers who are
  933. 35:57concerned about alignment. And the
  934. 36:00official who writes the regulation
  935. 36:01probably believes that he’s
  936. 36:02protecting the public. All of that may
  937. 36:05be true; I’m not saying it isn’t,
  938. 36:06it may be true. But the mistake, the
  939. 36:10mistake, my friends, is to believe that
  940. 36:12a sincere conviction removes a
  941. 36:14financial incentive or that an honest
  942. 36:16official removes the natural tendency
  943. 36:18of any administration to keep the damn
  944. 36:20key to what it regulates and what it
  945. 36:22doesn’t. That's why we should also
  946. 36:27look at our part, because we live in an
  947. 36:29information ecosystem where alarm is
  948. 36:31immediately rewarded. This video is
  949. 36:35going to be watched much less than if
  950. 36:37it had been titled AI is going to end
  951. 36:38humanity. A scientist who explains that
  952. 36:41we still can't accurately estimate the
  953. 36:43probability of that catastrophe is
  954. 36:45actually competing against a headline
  955. 36:47that claims that AI is going to kill us
  956. 36:48all before the end of this decade. You
  957. 36:52have to share this. You don't need to
  958. 36:54know about algorithms to guess which of
  959. 36:56the two is shared more. Then comes a
  960. 36:59popularizer who turns uncertainty into
  961. 37:01certainty because certainty gets clicks
  962. 37:02. Fear is normalized. The politician
  963. 37:06takes advantage of terrified public
  964. 37:08opinion and offers protection and
  965. 37:09everyone is reassured. And he also ends
  966. 37:11up asking the companies that build the
  967. 37:14technology what rules they think are
  968. 37:15reasonable to impose. Don't you see the
  969. 37:19circle? That circle closes there. I'm
  970. 37:22not crazy, it's crazy, it's different.
  971. 37:27So that's the verdict. We may need
  972. 37:29external verifiers, it is possible that
  973. 37:31certain capabilities require limits,
  974. 37:32and some future models may be too
  975. 37:34dangerous to deploy without
  976. 37:35implementing strict controls. I'm not
  977. 37:37saying they won't. At the same time,
  978. 37:39this regulation solves two problems
  979. 37:41that have nothing to do with your
  980. 37:43security. It allows an industry that
  981. 37:46hasn't yet proven that its accounts add
  982. 37:48up to stop burning money without
  983. 37:50admitting that the numbers didn't add
  984. 37:52up. And it also gives governments
  985. 37:57control—Mrs. Ursula will be delighted
  986. 37:59with this—control over the most
  987. 38:01important technologies of the century,
  988. 38:02just when they still had time to get
  989. 38:04their hands on them. Hey, fear up top,
  990. 38:06invoice below, and key in middle. And
  991. 38:10this isn't going to come free. If the
  992. 38:13intelligence that's going to drive the
  993. 38:15software, the accounting, the research,
  994. 38:17and a good part of your work is
  995. 38:19manufactured by four companies behind a
  996. 38:21wall that they helped design, and a
  997. 38:23government decides who crosses it, you
  998. 38:25know that you're not going to choose
  999. 38:26the provider. You pay the rent they
  1000. 38:30charge you and you'll use the tool they
  1001. 38:32authorize. That they authorize you. The
  1002. 38:36price of a mandatory input is not set
  1003. 38:38by the market, it is set by whoever
  1004. 38:40controls the gate. And the only thing
  1005. 38:43we can do is not buy the whole story,
  1006. 38:45man. When in the coming months you hear
  1007. 38:48that we need new audits, agencies,
  1008. 38:50licenses, and evaluators to protect us
  1009. 38:52from the terrible artificial
  1010. 38:54intelligence, ask yourself also what
  1011. 38:56cash problem each of these solutions
  1012. 38:57solves and how much power transfers to
  1013. 39:00whoever signs them. And when Anthropic
  1014. 39:04finally publishes its S1, let's read it
  1015. 39:07, we'll do it here. There will be more
  1016. 39:10truth about this industry there than in
  1017. 39:12100 prophecies about the end of the
  1018. 39:13world. You'll see, we'll find that
  1019. 39:17things were not going well. Because the
  1020. 39:21decisive question of this revolution is
  1021. 39:23not who will be able to create the most
  1022. 39:25powerful intelligence, it's who will
  1023. 39:27have permission to do so and who
  1024. 39:29distributes the permissions too. That's
  1025. 39:32why, before handing over the keys to
  1026. 39:34the wall, we have to look at who's
  1027. 39:36already inside that wall. Let's take a
  1028. 39:39look. Don't be impressed by some of the
  1029. 39:43headlines, at least not yet. Let's
  1030. 39:47continue. no.

About this transcript

This page contains the full transcript of El fin del mundo según OpenAI y Anthropic: ¿qué hay detrás de sus advertencias? by Marc Vidal, generated from the public captions YouTube serves with the video. The transcript has 6,552 words across 1,030 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to choose a caption track available on the video.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, caption language selection, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.