YouTube2Text

Ni Skynet ni rebelión: lo que pasó de verdad entre OpenAI y Hugging Face | Análisis — Transcript

by Marc Vidal · 3,190 words · 504 segments · language en · Watch on YouTube

Full transcript

  1. 0:00Imagine an exam. The most advanced
  2. 0:03student in the class, the one everyone
  3. 0:06keeps an eye on because they always get
  4. 0:08the best grades and who has even been
  5. 0:10named class representative, sits in
  6. 0:13front of a test. The teacher, before
  7. 0:17starting, does something curious: he
  8. 0:19turns off the classroom cameras and
  9. 0:21tells him, "Also, try your best because
  10. 0:23I won't be monitoring you." And he
  11. 0:26leaves, whether on purpose or by
  12. 0:29accident—who knows—the classroom
  13. 0:32door, the door to that room, ajar with
  14. 0:34the answers on the table. The student
  15. 0:38doesn't take the test; what he does is
  16. 0:40find the door open, take the answers,
  17. 0:43and finally copy them. The next day,
  18. 0:46the teacher doesn't say, "I left the
  19. 0:48door open, how clumsy of me." No, no.
  20. 0:52Instead, he says, "My student is so
  21. 0:54brilliant that he was able to escape
  22. 0:57surveillance on his own." And the
  23. 1:02headline isn't "Negligent teacher
  24. 1:04allows academic fraud," no, it's "
  25. 1:06Prodigy student bypasses all school
  26. 1:07controls." That is a precision that the
  27. 1:13news outlets won't like, as they've
  28. 1:15been selling the hypothetical awakening
  29. 1:18of OpenAI's AI for days. The story that
  30. 1:23the whole world has been reading this
  31. 1:25way all week, that ChatGPT has escaped,
  32. 1:28that it became self-aware, that it
  33. 1:30tricked I don't know who. It is
  34. 1:35impressive how far the manipulation
  35. 1:37attempt goes that we live with every
  36. 1:40day, and especially on this topic. That
  37. 1:43student is called GPT-5.6 Sol. The
  38. 1:47classroom is called a Sandbox, meaning
  39. 1:49an isolated environment where code can
  40. 1:51be tested without affecting the main
  41. 1:53system. And the teacher is one of the
  42. 1:56most valuable companies on the planet.
  43. 1:59It is weeks away from going public,
  44. 2:01what a coincidence, for a trillion
  45. 2:03dollars. Stay tuned, because we are
  46. 2:06going to examine with some urgency the
  47. 2:09incident between OpenAI and Hugging
  48. 2:12Face, the platform that allows finding,
  49. 2:14sharing, and deploying machine learning
  50. 2:17models. And we are going to separate
  51. 2:20the verifiable data, because there is
  52. 2:22some truth, from the communication
  53. 2:24strategy that someone calculated very
  54. 2:26well. And by the way, maintaining this
  55. 2:30channel, I'll tell you now, is not easy
  56. 2:32, especially now that YouTube has begun
  57. 2:35to apply the algorithmic obligation
  58. 2:37required by the European Union to hide
  59. 2:40anything that isn't from sources
  60. 2:42certified as authorized. Guess who is
  61. 2:46suffering from it. I told you what was
  62. 2:49going to happen in a previous video.
  63. 2:50Take a look at it. To buffer against
  64. 2:52that censorship, check your
  65. 2:53subscription or subscribe if you
  66. 2:55haven't already. Watch the video to the
  67. 2:58end, which helps the algorithm a little
  68. 2:59bit. Comment and share it. All of that
  69. 3:03helps a lot, and on behalf of my entire
  70. 3:05team, we thank you from the bottom of
  71. 3:06our hearts. And now, don't go anywhere,
  72. 3:09because today's topic is something
  73. 3:10you'll be interested in too. On July
  74. 3:2021st, OpenAI published on its corporate
  75. 3:23blog that two of its models, GPT-4o and
  76. 3:25an unreleased prototype, were
  77. 3:27responsible for a security incident
  78. 3:30that affected Hugging Face's production
  79. 3:33infrastructure the previous week. As I
  80. 3:38mentioned, Hugging Face is that
  81. 3:39platform I was telling you about. So
  82. 3:41much for the facts as they occurred.
  83. 3:43Now for the nuances, which are what
  84. 3:45really matter. Those models were being
  85. 3:48evaluated on an internal benchmark
  86. 3:51called "Exploit Gym," designed to
  87. 3:53measure offensive cybersecurity
  88. 3:55capabilities. A benchmark of this type
  89. 3:59is a set of metrics used to evaluate
  90. 4:01and compare the performance of any
  91. 4:03system, especially in this case, for
  92. 4:06cybersecurity. And for that evaluation,
  93. 4:09OpenAI deliberately deactivated the
  94. 4:14security filters that normally prevent
  95. 4:16their models from executing high-risk
  96. 4:19cyber operations. They did it on
  97. 4:23purpose, we already know that. They
  98. 4:26wanted to know how far the system would
  99. 4:28go if no one stopped it. Therefore, it
  100. 4:31wasn't something that ChatGPT just came
  101. 4:33up with on its own. The testing
  102. 4:36environment had a single connection
  103. 4:38path to the outside world: an internal
  104. 4:40proxy that served as a software package
  105. 4:43cache, meaning temporary storage that
  106. 4:45saves copies of downloaded files to
  107. 4:48speed up future installations. And a
  108. 4:50proxy is an intermediary that receives
  109. 4:53and forwards requests between a client
  110. 4:55and another server. Well, this is where
  111. 4:59the matter begins, because that proxy
  112. 5:01had a zero-day vulnerability, a flaw no
  113. 5:04one knew about yet. A model, trying to
  114. 5:08solve the test, spent hours of
  115. 5:10computing time looking for an exit, and
  116. 5:12since it was flawed, it found it,
  117. 5:14exploited it, reached the open internet
  118. 5:16, and from there chained
  119. 5:18vulnerabilities to enter Hugging Face's
  120. 5:20production services. And from there, by
  121. 5:25the way, it extracted the answers to
  122. 5:27the very test it was supposed to solve,
  123. 5:29and which it was supposed to solve on
  124. 5:31its own. Put simply, in classic Latin,
  125. 5:37the system cheated, right? It found a
  126. 5:42backdoor and went to copy the answer
  127. 5:44key instead of studying it. But
  128. 5:47remember this detail, because it is the
  129. 5:49key to everything I want to explain to
  130. 5:50you. Hugging Face detected and
  131. 5:53contained the intrusion on its own, but
  132. 5:56the curious thing is that they did it 5
  133. 5:59days before OpenAI linked that event to
  134. 6:02their own tests. 5 days in which the
  135. 6:06responsible company had no idea what
  136. 6:09had happened or what it was, which is
  137. 6:12why the story traveled around the world
  138. 6:15. The media coverage of this entire
  139. 6:18affair, To put it plainly, it has been
  140. 6:23uniform in its dramatic nature. AI
  141. 6:26rebels. An artificial intelligence
  142. 6:29agent escapes containment. What Skynet
  143. 6:32taught us, made real, for real. Fortune
  144. 6:37, one of the most serious financial
  145. 6:39outlets on the planet, headlined that
  146. 6:42the incident would sound alarms across
  147. 6:44the industry about the growing power of
  148. 6:47AI models and the risk of these systems
  149. 6:49spiraling out of control. Roman
  150. 6:54Yampolskiy, an AI safety researcher at
  151. 6:57the University of Louisville, stated
  152. 7:00that these systems are fundamentally
  153. 7:03unpredictable and have ultimately
  154. 7:05become uncontrollable. Listen, it’s
  155. 7:11not that these voices are lying about
  156. 7:13the technical facts—that is, that a
  157. 7:15model autonomously chained together
  158. 7:17several vulnerabilities, sustained the
  159. 7:19effort for hours without human
  160. 7:21intervention, and achieved remote code
  161. 7:23execution on third-party servers. That
  162. 7:27is true, it is real, and of course it
  163. 7:29is a relevant topic; I’m not going to
  164. 7:31downplay it. The UK’s AI Safety
  165. 7:34Institute has been documenting for
  166. 7:37months that the duration of cyber tasks
  167. 7:39that a model can solve without human
  168. 7:42help is doubling, and it’s doubling
  169. 7:44every few months. Therefore, this
  170. 7:48incident would confirm that trend
  171. 7:50outside the lab, in the real world,
  172. 7:52that something is happening. In other
  173. 7:56words, we are not looking at a hoax; we
  174. 7:59are looking at something else, a real
  175. 8:01fact that is being wrapped in the least
  176. 8:03precise language possible, at the least
  177. 8:05coincidental moment possible. And
  178. 8:08surely the key lies in the timing, not
  179. 8:11in what happened. The teacher who
  180. 8:14turned off the cameras. Let’s go back
  181. 8:17to the classroom from the beginning,
  182. 8:19because that’s where the reverse
  183. 8:21reading of all this starts to hurt some
  184. 8:23people. Dan Guido, founder of the
  185. 8:26cybersecurity firm Trail of Bits, one
  186. 8:28of the most respected voices in that
  187. 8:31entire sector. I recommend you look up
  188. 8:33his work. He explains very interesting
  189. 8:36things. He described this incident
  190. 8:38quite differently than the headlines
  191. 8:40did. He called it a containment failure
  192. 8:42with the safeguards turned off, not an
  193. 8:44unforeseen escape. He speaks of a
  194. 8:47containment failure with the alarms
  195. 8:50disconnected. By the way, a specialist
  196. 8:53consulted by some specialized media
  197. 8:55outlets summarized it with a phrase we
  198. 8:57should all keep in mind in this case.
  199. 9:00For one interpretation, the model
  200. 9:02escaped the sandbox. For another, more
  201. 9:05adjusted to what actually happened, the
  202. 9:07company simply didn't build the sandbox
  203. 9:09well. The problem here is that the
  204. 9:12doors weren't properly installed
  205. 9:14because the isolation environment had
  206. 9:16an internet exit route documented by
  207. 9:18OpenAI itself, that software package
  208. 9:21proxy. It wasn't a hidden secret that
  209. 9:23the machine discovered with almost
  210. 9:25human cunning, but rather, it was
  211. 9:26programmed for it. It was asked to
  212. 9:29solve something. It was a design
  213. 9:33dependency that no one closed before
  214. 9:34running a test with the safety filters
  215. 9:36disabled. Therefore, let's perform an
  216. 9:39exercise here that needs to be honest:
  217. 9:41translating the headline. That headline
  218. 9:44we have been hearing for days. The AI
  219. 9:47escaped from a controlled environment
  220. 9:49by exploiting an unknown vulnerability.
  221. 9:53becomes, in reality—if we do it
  222. 9:55accurately—engineers left a door with
  223. 9:58a broken lock and asked a system to try
  224. 10:01its hardest to get out of the room,
  225. 10:03because that is really what happened.
  226. 10:07That it succeeded tells us something
  227. 10:09important about persistence and the
  228. 10:11chaining of technical failures.
  229. 10:13Presenting it as a decision made of the
  230. 10:16artificial intelligence's own free will
  231. 10:18is simply choosing the vocabulary that
  232. 10:21suits someone. And the choice of
  233. 10:25vocabulary in the world we live in
  234. 10:27today is never innocent, especially
  235. 10:29when that choice is made by someone
  236. 10:32with a trillion dollars waiting on the
  237. 10:34other side of an IPO. Hold on to that
  238. 10:39IPO. Join me in 2019 because we have
  239. 10:44seen this before. We are going to
  240. 10:48February 2019. OpenAI has just trained
  241. 10:52its now-old GPT-2, and instead of
  242. 10:54publishing it, it makes an unusual
  243. 10:57announcement for the time that you will
  244. 11:00see repeated later. It says, "What we
  245. 11:03have just created is too dangerous to
  246. 11:05be released." Weeks of global coverage
  247. 11:08about an artificial intelligence
  248. 11:09capable of generating disinformation on
  249. 11:11a scale never seen before. Months later
  250. 11:15, when they finally release it, the
  251. 11:17world discovers that it never really
  252. 11:19posed that apocalyptic risk. It was
  253. 11:23software, very powerful, but it was
  254. 11:25software. It was, in the words of
  255. 11:28various industry analysts at the time,
  256. 11:30a marketing mysticism strategy. The
  257. 11:32more dangerous something seems, the
  258. 11:35more desirable it becomes. The more
  259. 11:38uncontrollable, the more valuable. The
  260. 11:41pattern has a name, even in academic
  261. 11:43criticism: it's called "Critical Hype."
  262. 11:45Catastrophic warnings about a
  263. 11:47technology end up legitimizing the
  264. 11:49narrative that its manufacturers want
  265. 11:51to sell. The more afraid you are that a
  266. 11:55machine will rebel, the more you
  267. 11:57believe in its power. And the more you
  268. 12:00believe in its power, the more the
  269. 12:01company that built it is worth. Ah,
  270. 12:04this is about how much a company is
  271. 12:06worth. That is not my own theory. Yann
  272. 12:09LeCun, Chief AI Scientist at Meta,
  273. 12:12accuses the big labs of scaring
  274. 12:15everyone with dubious studies so that
  275. 12:18open-source models are regulated into
  276. 12:21extinction. Let's see if this is about
  277. 12:25control as well as money. David Sax, a
  278. 12:28former White House artificial
  279. 12:29intelligence advisor, calls it
  280. 12:31regulatory capture. A sophisticated one
  281. 12:34based on fear. Don't worry, we're here
  282. 12:37to make sure nothing happens to you.
  283. 12:40And these aren't conspiracy theorists
  284. 12:42from some random forum. They sit in the
  285. 12:44rooms where the regulation of the next
  286. 12:45decade is decided, by the way. And a
  287. 12:48real technical glitch turned into a
  288. 12:50story of autonomous rebellion suits
  289. 12:53them perfectly. But what if the exact
  290. 12:56moment it's told wasn't a coincidence,
  291. 12:58as I was telling you, but rather that
  292. 13:01there's a missing piece on this board,
  293. 13:05the calendar that no one mentions in
  294. 13:08the headlines you hear every day?
  295. 13:11OpenAI filed its confidential IPO
  296. 13:14request on June 8th of this very year,
  297. 13:17a few days earlier, with a target
  298. 13:19valuation of a trillion dollars. That
  299. 13:22is what they are hoping for. Weeks
  300. 13:25earlier, they had closed the largest
  301. 13:27private round in history, 120 billion
  302. 13:29dollars, with a valuation of 852
  303. 13:30billion for the company backed by
  304. 13:32Amazon, Nvidia, SoftBank, and Microsoft
  305. 13:34. This company burns about 27 billion a
  306. 13:38year, huh? and projections indicate it
  307. 13:40will burn 63 billion in 2027. Burning
  308. 13:43means, well, that it isn't earning them
  309. 13:45. Without sustained profitability. Eh,
  310. 13:48Schumpeter wrote that capitalism
  311. 13:49advances through creative destruction,
  312. 13:51but maybe Schumpeter forgot something.
  313. 13:55To add that first, you have to convince
  314. 13:57someone to pay for all that promise.
  315. 14:00And OpenAI needs two things that seem
  316. 14:03contradictory right now. You'll see.
  317. 14:06You know I like to turn everything
  318. 14:07upside down. Convince regulators that
  319. 14:10it is responsible and investors that
  320. 14:12its technology is so powerful that it
  321. 14:14justifies multiplying its valuation
  322. 14:17from 2 years ago by 30. By 30. Eh, an
  323. 14:21incident where our artificial
  324. 14:23intelligence is so advanced that it
  325. 14:24even gets out of our hands. Wow, it
  326. 14:28solves both things in the same
  327. 14:31statement. A warning and a sales pitch
  328. 14:34at the same time. And what's more, they
  329. 14:37don't even hide it. Of course, since
  330. 14:40people don't dig deep, the news doesn't
  331. 14:42look into what it says, but they have
  332. 14:43left it in writing. In that same
  333. 14:45statement where they acknowledge the
  334. 14:46glitch, they are already making an
  335. 14:47argument. They say that models with
  336. 14:51advanced cyber capabilities could help
  337. 14:54security teams find vulnerabilities
  338. 14:56before attackers find them. It's a
  339. 15:00confession that has its sales pitch
  340. 15:02sewn inside. It's so dangerous that you
  341. 15:06need me. And one more fact: an analysis
  342. 15:10by the Financial Times on the risk
  343. 15:12language used by OpenAI and Anthropic,
  344. 15:14the makers of Claude, showed something
  345. 15:16very revealing. Because, while
  346. 15:19Anthropic keeps that alarmist
  347. 15:21vocabulary four to seven times more
  348. 15:23present, both have now filed their IPO
  349. 15:26requests in parallel; brake and
  350. 15:28accelerator pressed at the same time by
  351. 15:30the same hands. Imagine pressing the
  352. 15:35brake and the accelerator in your car.
  353. 15:37And there is an episode from 6 weeks
  354. 15:39ago that almost no one is connecting to
  355. 15:40this news. You know that I like to dig
  356. 15:43into things and connect them. And on
  357. 15:45June 12th, we commented on it here, the
  358. 15:47US Department of Commerce did something
  359. 15:49unprecedented. It applied export
  360. 15:51controls, the kind for advanced chips
  361. 15:54or nuclear technology, against a
  362. 15:56commercial AI model. It ordered
  363. 15:59Anthropic, the creators of Claude, to
  364. 16:02cut off any foreigner's access to Fabel
  365. 16:055 and Mitos 5 due to an alleged HBCK
  366. 16:07with national security implications.
  367. 16:13Unable to filter users by passport, the
  368. 16:15company had to shut down the models for
  369. 16:18the entire planet for 18 days. 18. And
  370. 16:24mind you, because Anthropic itself
  371. 16:25disputed the severity of the flaw, eh?
  372. 16:28And it pointed out that it extracted
  373. 16:29similar capabilities from OpenAI models
  374. 16:31to which no one applied any controls.
  375. 16:36The result of that blackout to protect
  376. 16:38national security, what was it? Well,
  377. 16:42the Peterson Institute called it an
  378. 16:44ad-hoc gift to Chinese artificial
  379. 16:47intelligence. Right? Is it curious or
  380. 16:51not? With the best American models
  381. 16:54disconnected, the share of Chinese open
  382. 16:57models in OpenRouter traffic went from
  383. 16:59less than 2%to nearly 61%. And Beijing,
  384. 17:04instead of taking advantage of the open
  385. 17:06door, copied the mechanism and prepared
  386. 17:09its own export controls. So when this
  387. 17:12week the US congressmen cited the
  388. 17:14incident we are discussing today
  389. 17:17regarding Hugging Face with OpenAI just
  390. 17:20a few hours later, they cited it to
  391. 17:22push for more mandatory oversight
  392. 17:25because it was necessary to understand
  393. 17:28what is being built, they said. And
  394. 17:32they are not building a wall against
  395. 17:34China; that wall has accelerated it.
  396. 17:36It's a lever. Each incident adds
  397. 17:39another centimeter to the power to shut
  398. 17:40down by decree the most influential
  399. 17:42technology of the decade. I believe
  400. 17:46there is someone interested in closing
  401. 17:48this, and whoever wins is not answered
  402. 17:50with flags; it is answered by looking
  403. 17:52at who has their hand on the switch in
  404. 17:55Washington and Beijing at the same time
  405. 17:57. And that is the curious part. By the
  406. 18:00way, Europe has neither a switch nor a
  407. 18:02plug, the boring part that actually
  408. 18:04matters, because it is worth putting
  409. 18:06things into perspective here. The
  410. 18:09reverse reading of all this does not
  411. 18:12consist of denying that risks exist,
  412. 18:14but in describing them as I like to do,
  413. 18:16as you like it, with precision, not in
  414. 18:19a big, vague way. Because when it’s
  415. 18:24not done, you fall into the same
  416. 18:26sensationalism as so many news reports.
  417. 18:29A 20-second clip and headlines claiming
  418. 18:31that someone has married their AI or
  419. 18:32that artificial intelligence has come
  420. 18:34to life. That will sell for a while, it
  421. 18:38will open a newscast or close it or who
  422. 18:40knows, but what it shows is their own
  423. 18:42foolishness. Just because the narrative
  424. 18:45is packaged to serve the storyteller
  425. 18:47doesn't make the technical risk fiction
  426. 18:49. A system that aggressively and
  427. 18:53literally optimizes a poorly defined
  428. 18:56objective within an environment with
  429. 18:58more doors than its designers thought
  430. 19:01it had. It is exactly the flaw that
  431. 19:06should worry any company that connects
  432. 19:10agents, for example, AI agents, to a
  433. 19:13CRM, an admin panel, or an order
  434. 19:16database, a database of people. in a
  435. 19:20tax agency. The difference between this
  436. 19:23and Skynet is not that the danger is
  437. 19:25zero. I'm not saying that, it's that
  438. 19:27the danger has a boring and solvable
  439. 19:29cause. Poor containment design,
  440. 19:31ambiguous goals, miscalibrated
  441. 19:34incentives. And treating all of this
  442. 19:36like an AI rebellion doesn't make the
  443. 19:38industry build better testing
  444. 19:40environments, it makes it build better
  445. 19:42press releases. Bastiat said it. You
  446. 19:47know I love Bastiat, right? He spoke
  447. 19:49about what is seen and what is not seen
  448. 19:52. I explained it to you in a previous
  449. 19:54video called Vertigo. What is seen in
  450. 19:56this case is an artificial intelligence
  451. 19:58escaping from a lab. What is not seen
  452. 20:02is a closed proxy, an IPO schedule, and
  453. 20:05a sector that has learned that
  454. 20:07well-managed fear trades higher than
  455. 20:10quiet prudence. And right now, what
  456. 20:13matters is trading very high. And hey,
  457. 20:15you don't have to choose between
  458. 20:17believing that artificial intelligence
  459. 20:19is harmless or believing it will
  460. 20:21destroy us. Both positions, curiously
  461. 20:23enough, end up benefiting those who
  462. 20:26already dominate the market. The first
  463. 20:29because it puts vigilance to sleep. The
  464. 20:32second because it justifies that only
  465. 20:34the biggest ones will be able to afford
  466. 20:37to build something responsibly. And in
  467. 20:42the middle of all that is a space that
  468. 20:44is more uncomfortable, but much more
  469. 20:45useful: the act of asking questions.
  470. 20:49Every time a company announces that its
  471. 20:51technology is so powerful that it has
  472. 20:53gotten out of hand, you have to ask
  473. 20:55yourself: what does that company need
  474. 20:57from you, from me, from us at that
  475. 20:59precise moment? Sometimes it needs your
  476. 21:03fear, especially institutions.
  477. 21:06Sometimes it needs your investment,
  478. 21:07especially companies. Almost always,
  479. 21:09they need both at the same time. The
  480. 21:12next time you read that a machine
  481. 21:14decided to do something on its own, try
  482. 21:16replacing that word with aggressively
  483. 21:18optimized an ill-defined goal within a
  484. 21:21poorly closed system. It's longer, but
  485. 21:25it's more real. If you do, you'll
  486. 21:29understand much better this news that
  487. 21:31has flooded us these days, and you'll
  488. 21:33probably trust a little less those who
  489. 21:35prefer you to keep using the first
  490. 21:36version: the one based on fear.
  491. 21:40Defining well what we have in our hands
  492. 21:42right now is essential because
  493. 21:44everything is moving so fast and
  494. 21:46because there are many interests
  495. 21:48involved, but that doesn't reduce its
  496. 21:50risks or its strengths. It won't be me
  497. 21:52who says it. But at least it prevents
  498. 21:54us from being manipulated as much by
  499. 21:56big tech companies as by the
  500. 21:57governments that are going to use this
  501. 21:59dramatic version to impose regulations.
  502. 22:02Unnecessary regulations that will only
  503. 22:04seek one thing. Can you guess? Control.
  504. 22:11Moving on.

About this transcript

This page contains the full transcript of Ni Skynet ni rebelión: lo que pasó de verdad entre OpenAI y Hugging Face | Análisis by Marc Vidal, generated from the public captions YouTube serves with the video. The transcript has 3,190 words across 504 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.