Ni Skynet ni rebelión: lo que pasó de verdad entre OpenAI y Hugging Face | Análisis — Transcript
Full transcript
- 0:00Imagine an exam. The most advanced
- 0:03student in the class, the one everyone
- 0:06keeps an eye on because they always get
- 0:08the best grades and who has even been
- 0:10named class representative, sits in
- 0:13front of a test. The teacher, before
- 0:17starting, does something curious: he
- 0:19turns off the classroom cameras and
- 0:21tells him, "Also, try your best because
- 0:23I won't be monitoring you." And he
- 0:26leaves, whether on purpose or by
- 0:29accident—who knows—the classroom
- 0:32door, the door to that room, ajar with
- 0:34the answers on the table. The student
- 0:38doesn't take the test; what he does is
- 0:40find the door open, take the answers,
- 0:43and finally copy them. The next day,
- 0:46the teacher doesn't say, "I left the
- 0:48door open, how clumsy of me." No, no.
- 0:52Instead, he says, "My student is so
- 0:54brilliant that he was able to escape
- 0:57surveillance on his own." And the
- 1:02headline isn't "Negligent teacher
- 1:04allows academic fraud," no, it's "
- 1:06Prodigy student bypasses all school
- 1:07controls." That is a precision that the
- 1:13news outlets won't like, as they've
- 1:15been selling the hypothetical awakening
- 1:18of OpenAI's AI for days. The story that
- 1:23the whole world has been reading this
- 1:25way all week, that ChatGPT has escaped,
- 1:28that it became self-aware, that it
- 1:30tricked I don't know who. It is
- 1:35impressive how far the manipulation
- 1:37attempt goes that we live with every
- 1:40day, and especially on this topic. That
- 1:43student is called GPT-5.6 Sol. The
- 1:47classroom is called a Sandbox, meaning
- 1:49an isolated environment where code can
- 1:51be tested without affecting the main
- 1:53system. And the teacher is one of the
- 1:56most valuable companies on the planet.
- 1:59It is weeks away from going public,
- 2:01what a coincidence, for a trillion
- 2:03dollars. Stay tuned, because we are
- 2:06going to examine with some urgency the
- 2:09incident between OpenAI and Hugging
- 2:12Face, the platform that allows finding,
- 2:14sharing, and deploying machine learning
- 2:17models. And we are going to separate
- 2:20the verifiable data, because there is
- 2:22some truth, from the communication
- 2:24strategy that someone calculated very
- 2:26well. And by the way, maintaining this
- 2:30channel, I'll tell you now, is not easy
- 2:32, especially now that YouTube has begun
- 2:35to apply the algorithmic obligation
- 2:37required by the European Union to hide
- 2:40anything that isn't from sources
- 2:42certified as authorized. Guess who is
- 2:46suffering from it. I told you what was
- 2:49going to happen in a previous video.
- 2:50Take a look at it. To buffer against
- 2:52that censorship, check your
- 2:53subscription or subscribe if you
- 2:55haven't already. Watch the video to the
- 2:58end, which helps the algorithm a little
- 2:59bit. Comment and share it. All of that
- 3:03helps a lot, and on behalf of my entire
- 3:05team, we thank you from the bottom of
- 3:06our hearts. And now, don't go anywhere,
- 3:09because today's topic is something
- 3:10you'll be interested in too. On July
- 3:2021st, OpenAI published on its corporate
- 3:23blog that two of its models, GPT-4o and
- 3:25an unreleased prototype, were
- 3:27responsible for a security incident
- 3:30that affected Hugging Face's production
- 3:33infrastructure the previous week. As I
- 3:38mentioned, Hugging Face is that
- 3:39platform I was telling you about. So
- 3:41much for the facts as they occurred.
- 3:43Now for the nuances, which are what
- 3:45really matter. Those models were being
- 3:48evaluated on an internal benchmark
- 3:51called "Exploit Gym," designed to
- 3:53measure offensive cybersecurity
- 3:55capabilities. A benchmark of this type
- 3:59is a set of metrics used to evaluate
- 4:01and compare the performance of any
- 4:03system, especially in this case, for
- 4:06cybersecurity. And for that evaluation,
- 4:09OpenAI deliberately deactivated the
- 4:14security filters that normally prevent
- 4:16their models from executing high-risk
- 4:19cyber operations. They did it on
- 4:23purpose, we already know that. They
- 4:26wanted to know how far the system would
- 4:28go if no one stopped it. Therefore, it
- 4:31wasn't something that ChatGPT just came
- 4:33up with on its own. The testing
- 4:36environment had a single connection
- 4:38path to the outside world: an internal
- 4:40proxy that served as a software package
- 4:43cache, meaning temporary storage that
- 4:45saves copies of downloaded files to
- 4:48speed up future installations. And a
- 4:50proxy is an intermediary that receives
- 4:53and forwards requests between a client
- 4:55and another server. Well, this is where
- 4:59the matter begins, because that proxy
- 5:01had a zero-day vulnerability, a flaw no
- 5:04one knew about yet. A model, trying to
- 5:08solve the test, spent hours of
- 5:10computing time looking for an exit, and
- 5:12since it was flawed, it found it,
- 5:14exploited it, reached the open internet
- 5:16, and from there chained
- 5:18vulnerabilities to enter Hugging Face's
- 5:20production services. And from there, by
- 5:25the way, it extracted the answers to
- 5:27the very test it was supposed to solve,
- 5:29and which it was supposed to solve on
- 5:31its own. Put simply, in classic Latin,
- 5:37the system cheated, right? It found a
- 5:42backdoor and went to copy the answer
- 5:44key instead of studying it. But
- 5:47remember this detail, because it is the
- 5:49key to everything I want to explain to
- 5:50you. Hugging Face detected and
- 5:53contained the intrusion on its own, but
- 5:56the curious thing is that they did it 5
- 5:59days before OpenAI linked that event to
- 6:02their own tests. 5 days in which the
- 6:06responsible company had no idea what
- 6:09had happened or what it was, which is
- 6:12why the story traveled around the world
- 6:15. The media coverage of this entire
- 6:18affair, To put it plainly, it has been
- 6:23uniform in its dramatic nature. AI
- 6:26rebels. An artificial intelligence
- 6:29agent escapes containment. What Skynet
- 6:32taught us, made real, for real. Fortune
- 6:37, one of the most serious financial
- 6:39outlets on the planet, headlined that
- 6:42the incident would sound alarms across
- 6:44the industry about the growing power of
- 6:47AI models and the risk of these systems
- 6:49spiraling out of control. Roman
- 6:54Yampolskiy, an AI safety researcher at
- 6:57the University of Louisville, stated
- 7:00that these systems are fundamentally
- 7:03unpredictable and have ultimately
- 7:05become uncontrollable. Listen, it’s
- 7:11not that these voices are lying about
- 7:13the technical facts—that is, that a
- 7:15model autonomously chained together
- 7:17several vulnerabilities, sustained the
- 7:19effort for hours without human
- 7:21intervention, and achieved remote code
- 7:23execution on third-party servers. That
- 7:27is true, it is real, and of course it
- 7:29is a relevant topic; I’m not going to
- 7:31downplay it. The UK’s AI Safety
- 7:34Institute has been documenting for
- 7:37months that the duration of cyber tasks
- 7:39that a model can solve without human
- 7:42help is doubling, and it’s doubling
- 7:44every few months. Therefore, this
- 7:48incident would confirm that trend
- 7:50outside the lab, in the real world,
- 7:52that something is happening. In other
- 7:56words, we are not looking at a hoax; we
- 7:59are looking at something else, a real
- 8:01fact that is being wrapped in the least
- 8:03precise language possible, at the least
- 8:05coincidental moment possible. And
- 8:08surely the key lies in the timing, not
- 8:11in what happened. The teacher who
- 8:14turned off the cameras. Let’s go back
- 8:17to the classroom from the beginning,
- 8:19because that’s where the reverse
- 8:21reading of all this starts to hurt some
- 8:23people. Dan Guido, founder of the
- 8:26cybersecurity firm Trail of Bits, one
- 8:28of the most respected voices in that
- 8:31entire sector. I recommend you look up
- 8:33his work. He explains very interesting
- 8:36things. He described this incident
- 8:38quite differently than the headlines
- 8:40did. He called it a containment failure
- 8:42with the safeguards turned off, not an
- 8:44unforeseen escape. He speaks of a
- 8:47containment failure with the alarms
- 8:50disconnected. By the way, a specialist
- 8:53consulted by some specialized media
- 8:55outlets summarized it with a phrase we
- 8:57should all keep in mind in this case.
- 9:00For one interpretation, the model
- 9:02escaped the sandbox. For another, more
- 9:05adjusted to what actually happened, the
- 9:07company simply didn't build the sandbox
- 9:09well. The problem here is that the
- 9:12doors weren't properly installed
- 9:14because the isolation environment had
- 9:16an internet exit route documented by
- 9:18OpenAI itself, that software package
- 9:21proxy. It wasn't a hidden secret that
- 9:23the machine discovered with almost
- 9:25human cunning, but rather, it was
- 9:26programmed for it. It was asked to
- 9:29solve something. It was a design
- 9:33dependency that no one closed before
- 9:34running a test with the safety filters
- 9:36disabled. Therefore, let's perform an
- 9:39exercise here that needs to be honest:
- 9:41translating the headline. That headline
- 9:44we have been hearing for days. The AI
- 9:47escaped from a controlled environment
- 9:49by exploiting an unknown vulnerability.
- 9:53becomes, in reality—if we do it
- 9:55accurately—engineers left a door with
- 9:58a broken lock and asked a system to try
- 10:01its hardest to get out of the room,
- 10:03because that is really what happened.
- 10:07That it succeeded tells us something
- 10:09important about persistence and the
- 10:11chaining of technical failures.
- 10:13Presenting it as a decision made of the
- 10:16artificial intelligence's own free will
- 10:18is simply choosing the vocabulary that
- 10:21suits someone. And the choice of
- 10:25vocabulary in the world we live in
- 10:27today is never innocent, especially
- 10:29when that choice is made by someone
- 10:32with a trillion dollars waiting on the
- 10:34other side of an IPO. Hold on to that
- 10:39IPO. Join me in 2019 because we have
- 10:44seen this before. We are going to
- 10:48February 2019. OpenAI has just trained
- 10:52its now-old GPT-2, and instead of
- 10:54publishing it, it makes an unusual
- 10:57announcement for the time that you will
- 11:00see repeated later. It says, "What we
- 11:03have just created is too dangerous to
- 11:05be released." Weeks of global coverage
- 11:08about an artificial intelligence
- 11:09capable of generating disinformation on
- 11:11a scale never seen before. Months later
- 11:15, when they finally release it, the
- 11:17world discovers that it never really
- 11:19posed that apocalyptic risk. It was
- 11:23software, very powerful, but it was
- 11:25software. It was, in the words of
- 11:28various industry analysts at the time,
- 11:30a marketing mysticism strategy. The
- 11:32more dangerous something seems, the
- 11:35more desirable it becomes. The more
- 11:38uncontrollable, the more valuable. The
- 11:41pattern has a name, even in academic
- 11:43criticism: it's called "Critical Hype."
- 11:45Catastrophic warnings about a
- 11:47technology end up legitimizing the
- 11:49narrative that its manufacturers want
- 11:51to sell. The more afraid you are that a
- 11:55machine will rebel, the more you
- 11:57believe in its power. And the more you
- 12:00believe in its power, the more the
- 12:01company that built it is worth. Ah,
- 12:04this is about how much a company is
- 12:06worth. That is not my own theory. Yann
- 12:09LeCun, Chief AI Scientist at Meta,
- 12:12accuses the big labs of scaring
- 12:15everyone with dubious studies so that
- 12:18open-source models are regulated into
- 12:21extinction. Let's see if this is about
- 12:25control as well as money. David Sax, a
- 12:28former White House artificial
- 12:29intelligence advisor, calls it
- 12:31regulatory capture. A sophisticated one
- 12:34based on fear. Don't worry, we're here
- 12:37to make sure nothing happens to you.
- 12:40And these aren't conspiracy theorists
- 12:42from some random forum. They sit in the
- 12:44rooms where the regulation of the next
- 12:45decade is decided, by the way. And a
- 12:48real technical glitch turned into a
- 12:50story of autonomous rebellion suits
- 12:53them perfectly. But what if the exact
- 12:56moment it's told wasn't a coincidence,
- 12:58as I was telling you, but rather that
- 13:01there's a missing piece on this board,
- 13:05the calendar that no one mentions in
- 13:08the headlines you hear every day?
- 13:11OpenAI filed its confidential IPO
- 13:14request on June 8th of this very year,
- 13:17a few days earlier, with a target
- 13:19valuation of a trillion dollars. That
- 13:22is what they are hoping for. Weeks
- 13:25earlier, they had closed the largest
- 13:27private round in history, 120 billion
- 13:29dollars, with a valuation of 852
- 13:30billion for the company backed by
- 13:32Amazon, Nvidia, SoftBank, and Microsoft
- 13:34. This company burns about 27 billion a
- 13:38year, huh? and projections indicate it
- 13:40will burn 63 billion in 2027. Burning
- 13:43means, well, that it isn't earning them
- 13:45. Without sustained profitability. Eh,
- 13:48Schumpeter wrote that capitalism
- 13:49advances through creative destruction,
- 13:51but maybe Schumpeter forgot something.
- 13:55To add that first, you have to convince
- 13:57someone to pay for all that promise.
- 14:00And OpenAI needs two things that seem
- 14:03contradictory right now. You'll see.
- 14:06You know I like to turn everything
- 14:07upside down. Convince regulators that
- 14:10it is responsible and investors that
- 14:12its technology is so powerful that it
- 14:14justifies multiplying its valuation
- 14:17from 2 years ago by 30. By 30. Eh, an
- 14:21incident where our artificial
- 14:23intelligence is so advanced that it
- 14:24even gets out of our hands. Wow, it
- 14:28solves both things in the same
- 14:31statement. A warning and a sales pitch
- 14:34at the same time. And what's more, they
- 14:37don't even hide it. Of course, since
- 14:40people don't dig deep, the news doesn't
- 14:42look into what it says, but they have
- 14:43left it in writing. In that same
- 14:45statement where they acknowledge the
- 14:46glitch, they are already making an
- 14:47argument. They say that models with
- 14:51advanced cyber capabilities could help
- 14:54security teams find vulnerabilities
- 14:56before attackers find them. It's a
- 15:00confession that has its sales pitch
- 15:02sewn inside. It's so dangerous that you
- 15:06need me. And one more fact: an analysis
- 15:10by the Financial Times on the risk
- 15:12language used by OpenAI and Anthropic,
- 15:14the makers of Claude, showed something
- 15:16very revealing. Because, while
- 15:19Anthropic keeps that alarmist
- 15:21vocabulary four to seven times more
- 15:23present, both have now filed their IPO
- 15:26requests in parallel; brake and
- 15:28accelerator pressed at the same time by
- 15:30the same hands. Imagine pressing the
- 15:35brake and the accelerator in your car.
- 15:37And there is an episode from 6 weeks
- 15:39ago that almost no one is connecting to
- 15:40this news. You know that I like to dig
- 15:43into things and connect them. And on
- 15:45June 12th, we commented on it here, the
- 15:47US Department of Commerce did something
- 15:49unprecedented. It applied export
- 15:51controls, the kind for advanced chips
- 15:54or nuclear technology, against a
- 15:56commercial AI model. It ordered
- 15:59Anthropic, the creators of Claude, to
- 16:02cut off any foreigner's access to Fabel
- 16:055 and Mitos 5 due to an alleged HBCK
- 16:07with national security implications.
- 16:13Unable to filter users by passport, the
- 16:15company had to shut down the models for
- 16:18the entire planet for 18 days. 18. And
- 16:24mind you, because Anthropic itself
- 16:25disputed the severity of the flaw, eh?
- 16:28And it pointed out that it extracted
- 16:29similar capabilities from OpenAI models
- 16:31to which no one applied any controls.
- 16:36The result of that blackout to protect
- 16:38national security, what was it? Well,
- 16:42the Peterson Institute called it an
- 16:44ad-hoc gift to Chinese artificial
- 16:47intelligence. Right? Is it curious or
- 16:51not? With the best American models
- 16:54disconnected, the share of Chinese open
- 16:57models in OpenRouter traffic went from
- 16:59less than 2%to nearly 61%. And Beijing,
- 17:04instead of taking advantage of the open
- 17:06door, copied the mechanism and prepared
- 17:09its own export controls. So when this
- 17:12week the US congressmen cited the
- 17:14incident we are discussing today
- 17:17regarding Hugging Face with OpenAI just
- 17:20a few hours later, they cited it to
- 17:22push for more mandatory oversight
- 17:25because it was necessary to understand
- 17:28what is being built, they said. And
- 17:32they are not building a wall against
- 17:34China; that wall has accelerated it.
- 17:36It's a lever. Each incident adds
- 17:39another centimeter to the power to shut
- 17:40down by decree the most influential
- 17:42technology of the decade. I believe
- 17:46there is someone interested in closing
- 17:48this, and whoever wins is not answered
- 17:50with flags; it is answered by looking
- 17:52at who has their hand on the switch in
- 17:55Washington and Beijing at the same time
- 17:57. And that is the curious part. By the
- 18:00way, Europe has neither a switch nor a
- 18:02plug, the boring part that actually
- 18:04matters, because it is worth putting
- 18:06things into perspective here. The
- 18:09reverse reading of all this does not
- 18:12consist of denying that risks exist,
- 18:14but in describing them as I like to do,
- 18:16as you like it, with precision, not in
- 18:19a big, vague way. Because when it’s
- 18:24not done, you fall into the same
- 18:26sensationalism as so many news reports.
- 18:29A 20-second clip and headlines claiming
- 18:31that someone has married their AI or
- 18:32that artificial intelligence has come
- 18:34to life. That will sell for a while, it
- 18:38will open a newscast or close it or who
- 18:40knows, but what it shows is their own
- 18:42foolishness. Just because the narrative
- 18:45is packaged to serve the storyteller
- 18:47doesn't make the technical risk fiction
- 18:49. A system that aggressively and
- 18:53literally optimizes a poorly defined
- 18:56objective within an environment with
- 18:58more doors than its designers thought
- 19:01it had. It is exactly the flaw that
- 19:06should worry any company that connects
- 19:10agents, for example, AI agents, to a
- 19:13CRM, an admin panel, or an order
- 19:16database, a database of people. in a
- 19:20tax agency. The difference between this
- 19:23and Skynet is not that the danger is
- 19:25zero. I'm not saying that, it's that
- 19:27the danger has a boring and solvable
- 19:29cause. Poor containment design,
- 19:31ambiguous goals, miscalibrated
- 19:34incentives. And treating all of this
- 19:36like an AI rebellion doesn't make the
- 19:38industry build better testing
- 19:40environments, it makes it build better
- 19:42press releases. Bastiat said it. You
- 19:47know I love Bastiat, right? He spoke
- 19:49about what is seen and what is not seen
- 19:52. I explained it to you in a previous
- 19:54video called Vertigo. What is seen in
- 19:56this case is an artificial intelligence
- 19:58escaping from a lab. What is not seen
- 20:02is a closed proxy, an IPO schedule, and
- 20:05a sector that has learned that
- 20:07well-managed fear trades higher than
- 20:10quiet prudence. And right now, what
- 20:13matters is trading very high. And hey,
- 20:15you don't have to choose between
- 20:17believing that artificial intelligence
- 20:19is harmless or believing it will
- 20:21destroy us. Both positions, curiously
- 20:23enough, end up benefiting those who
- 20:26already dominate the market. The first
- 20:29because it puts vigilance to sleep. The
- 20:32second because it justifies that only
- 20:34the biggest ones will be able to afford
- 20:37to build something responsibly. And in
- 20:42the middle of all that is a space that
- 20:44is more uncomfortable, but much more
- 20:45useful: the act of asking questions.
- 20:49Every time a company announces that its
- 20:51technology is so powerful that it has
- 20:53gotten out of hand, you have to ask
- 20:55yourself: what does that company need
- 20:57from you, from me, from us at that
- 20:59precise moment? Sometimes it needs your
- 21:03fear, especially institutions.
- 21:06Sometimes it needs your investment,
- 21:07especially companies. Almost always,
- 21:09they need both at the same time. The
- 21:12next time you read that a machine
- 21:14decided to do something on its own, try
- 21:16replacing that word with aggressively
- 21:18optimized an ill-defined goal within a
- 21:21poorly closed system. It's longer, but
- 21:25it's more real. If you do, you'll
- 21:29understand much better this news that
- 21:31has flooded us these days, and you'll
- 21:33probably trust a little less those who
- 21:35prefer you to keep using the first
- 21:36version: the one based on fear.
- 21:40Defining well what we have in our hands
- 21:42right now is essential because
- 21:44everything is moving so fast and
- 21:46because there are many interests
- 21:48involved, but that doesn't reduce its
- 21:50risks or its strengths. It won't be me
- 21:52who says it. But at least it prevents
- 21:54us from being manipulated as much by
- 21:56big tech companies as by the
- 21:57governments that are going to use this
- 21:59dramatic version to impose regulations.
- 22:02Unnecessary regulations that will only
- 22:04seek one thing. Can you guess? Control.
- 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.