El fin del mundo según OpenAI y Anthropic: ¿qué hay detrás de sus advertencias? — Transcript
Full transcript
- 0:00A researcher leaves Anthropic, warns
- 0:02that artificial intelligence could
- 0:04destroy us before the end of this
- 0:07decade, and his message has been viewed
- 0:09more than 100 million times in a very
- 0:11short space of time. Other researchers
- 0:14confirm that they are also very worried
- 0:17. The media, of course, fills their
- 0:20headlines with the possibility of
- 0:21losing control of machines and that the
- 0:23end of the world is getting closer. And
- 0:27a few days later, Dario Amodei, the CEO
- 0:30of Anthropic, publishes a proposal to
- 0:32slow down the development of the most
- 0:34advanced systems, and immediately after
- 0:37, to bring external evaluators into
- 0:39large laboratories and allow companies
- 0:41to coordinate to set certain safety
- 0:43limits. Then, Samalman, the head of
- 0:47Openi, the one from Chat GPT, is in
- 0:49favor. Elon Musk too, and almost at the
- 0:53same time, even Openi announces that
- 0:552026 does not seem like a good time to
- 0:57go public, and they push their IPO back
- 0:59to next year. When an entire industry
- 1:03asks to be regulated, it would be
- 1:05advisable to be suspicious by default,
- 1:07and not because they are lying, I am
- 1:09not saying that they are, but because
- 1:10there is almost always another reason,
- 1:12at least two reasons, and neither of
- 1:14them is the one that appears in the
- 1:16headline. The first is that the numbers
- 1:18do not add up for them, which is why
- 1:20they want to delay the launch. The
- 1:23second is that the State, at least the
- 1:25State, has never been bothered by
- 1:27keeping the key, the key to regulation.
- 1:30That is what I want to present to you
- 1:31today. Not that they have reached an
- 1:33agreement in a secret room, right?
- 1:35Because I cannot prove that and it
- 1:37probably has not happened. What I see
- 1:40is an industry that has entered into a
- 1:42race that demands hundreds of billions
- 1:44of dollars, electricity measured in
- 1:45gigawatts, chips that age at a brutal
- 1:47speed and an infrastructure that does
- 1:49not grow at the same rate as demand. A
- 1:53race in which no one has yet
- 1:54demonstrated, by the way, how much
- 1:56money each dollar invested returns. And
- 2:00at that precise moment, those ahead
- 2:02have discovered that slowing down is
- 2:04the most prudent thing to do, and
- 2:05governments discover that they finally
- 2:07have an impeccable excuse to decide who
- 2:10can build the technology that will move
- 2:12the century. Hm. The first existential
- 2:16problem of artificial intelligence may
- 2:19not be existential for us; it may be
- 2:21more financial for them, more political
- 2:24for those who sign the licenses. Stay
- 2:28tuned, because we are going to turn the
- 2:29official message around, and today it
- 2:31is going to be difficult, because today
- 2:33we are going to see how or why the
- 2:35accounts of this industry still do not
- 2:37add up and what exactly is in it for
- 2:39those who ask to slow down; why safety
- 2:41regulation is the best possible deal
- 2:42for those who are already in the lead
- 2:44and the worst for those who come after,
- 2:46and what part of the spoils a state, an
- 2:48administration, a union is going to
- 2:50take in this distribution and, above
- 2:52all, how much is all this going to cost
- 2:54you? And we are going to do so by
- 2:57trying to analyze this hypothesis
- 2:59before accepting it as it is being done
- 3:02. Anyway, remember that you have three
- 3:03ways to support this channel by
- 3:05commenting and watching the video until
- 3:06the end, which is what pushes the
- 3:07algorithm, because in these videos they
- 3:09usually push it very little.
- 3:10Subscribing, which is free, or becoming
- 3:12a member for very little. You know that
- 3:14my team will be especially happy to do
- 3:16so, although any of the three options
- 3:17will seem fine to us. As I said, don't
- 3:19go away because today's news will also
- 3:21interest you. Let's go with what we
- 3:29know. A certain Jacob Coxson, who first
- 3:32worked at Openi and then at Anthropic,
- 3:34left the latter company a few days ago
- 3:36and issued a very serious warning. He
- 3:41maintains that a good part of those who
- 3:43are building the most advanced models
- 3:45sincerely believe that there is a
- 3:46significant probability that future
- 3:48systems will escape the control of
- 3:50human beings. And there is a detail
- 3:53that makes it difficult to dismiss him
- 3:56as an angry ex-employee, because Coxson
- 3:58left before consolidating part of his
- 4:00economic participation, that is, he had
- 4:02money on the table to stay and still
- 4:05this man decided to leave. Therefore,
- 4:07man, we have to take that into account.
- 4:10His diagnosis may even be wrong, but
- 4:12despite everything, his concern can be
- 4:14taken seriously, I'm not saying it
- 4:16isn't. And he's not alone. Other
- 4:19security researchers have come out of
- 4:21frontier labs or have publicly
- 4:22expressed similar concerns. Anthropic,
- 4:26Openei and the rest are already making
- 4:28specific assessments on cybersecurity,
- 4:30biology, chemical weapons, deception
- 4:33capacity, loss of control and autonomy
- 4:35of agents. To say that there are real
- 4:39risks there would be as absurd as those
- 4:41who claim that we already know that a
- 4:43superintelligence is going to end
- 4:44humanity by 2030, as they are saying,
- 4:46by the way. And there have been serious
- 4:49incidents that must be taken into
- 4:50account. The Haging Face incident is
- 4:53the most striking, we explained it here
- 4:54a few weeks ago in a previous video.
- 4:57During internal cybersecurity
- 4:58assessments, several Openi agents found
- 5:01vulnerabilities in the infrastructure
- 5:03that was supposed to keep them isolated
- 5:05. They set up communication channels
- 5:07that no one had anticipated and, in
- 5:09addition, these algorithms gained
- 5:11access to the internet and ended up
- 5:12compromising third-party platform
- 5:14systems. It is a serious matter, of
- 5:18course, very serious, but it is worth
- 5:20looking at what it demonstrates and
- 5:22what it does not demonstrate, because
- 5:23it does not demonstrate that an
- 5:25artificial intelligence woke up or
- 5:27became aware and decided to escape from
- 5:29something. Openi's own technical report
- 5:33, which many of those who speak about
- 5:35it seem not to have read in detail,
- 5:37describes models subjected to computer
- 5:39exploitation tasks with safeguards
- 5:41purposely reduced, optimizing
- 5:43objectives within a flawed environment
- 5:45and finding paths that the researchers
- 5:47had not anticipated. Mathematics
- 5:51optimizing an objective function on a
- 5:53poorly closed infrastructure. That is
- 5:56exactly why the problem is serious. Yes
- 5:59. Yes. Although it is not what they say
- 6:02, it is a serious issue. You do not
- 6:04need to attribute a soul to a matrix
- 6:06for a highly competent machine to cause
- 6:08a disaster. An algorithm does not need
- 6:11to hate you to hack your server. It is
- 6:14enough for it to do so in the most
- 6:15effective way to achieve the objective
- 6:17that you have given it. The problem
- 6:20begins when we move from the risk to
- 6:22what is now filling all the news in the
- 6:25world. The story. A technically
- 6:27plausible possibility first becomes a
- 6:30journalistic probability, then a
- 6:32headline, then a prophecy, and finally
- 6:35a political demand. Be careful with
- 6:36that. The media steals the attention.
- 6:39The disaster reporter steals the
- 6:41relevance, the security expert steals
- 6:43the public spotlight, and organizations
- 6:45dedicated to risk take the funding.
- 6:48Even governments get a frightened
- 6:50public opinion that begs them to do
- 6:52something, to stop this at all costs,
- 6:54for our own good. And the big labs get
- 6:58the most valuable thing of all, which
- 6:59is the position from which to explain
- 7:01to that government exactly what should
- 7:02be regulated. It's their tool, it's
- 7:05their key. None of that shows bad faith
- 7:07. I'm not saying it doesn't. I'm not
- 7:08going to suspect anyone now, and that's
- 7:10why it works too. Economic incentives
- 7:12don't need a secret room where someone
- 7:14decides on a plan. While we're
- 7:17discussing whether an AI had become
- 7:19Skynet, behind the screen something
- 7:21else was happening that was much less
- 7:23cinematic and actually much more urgent
- 7:26for them. Because this, my friends,
- 7:29those of you who follow this channel
- 7:32know that all of this has a price. The
- 7:35monster also has a bill. On September
- 7:3910, AENA, Openei did something very
- 7:42strange. It temporarily suspended new
- 7:45registrations and upgrades to the Chat
- 7:47GPT Pro plan for its $ 200 subscription
- 7:50. And it didn't do it because of a lack
- 7:52of customers. On the contrary, the
- 7:55demand for Astra, which is its most
- 7:57advanced model, was putting such
- 7:58pressure on its computing capacity that
- 8:00the company decided to stop accepting
- 8:02some of the customers who were willing
- 8:04to pay it the most for a while. Think
- 8:10about how strange this is. You have a
- 8:13product that people want to buy. You
- 8:16have customers who offer you $ 200 each
- 8:18month, which is much more than what
- 8:19they were paying. And you have to close
- 8:22half the door because you can't serve
- 8:25them. Openi has too many customers and
- 8:27not enough manufacturing. And that
- 8:29manufacturing is no longer a room with
- 8:32servers, it's land, data centers, power
- 8:34plants, substations, cooling, fiber,
- 8:37transformers and hundreds of thousands
- 8:40of accelerators working at the same
- 8:42time. OpenAI has increased its cloud
- 8:47spending plans to, hold on, seven
- 8:50hundred and fifty billion dollars by
- 8:532030. Anthropic is not playing around
- 8:59with pocket change either. It has
- 9:01signed up to 5 GW of capacity with
- 9:03Amazon and technology commitments for
- 9:05more than $ 100 billion over the next
- 9:07decade. It has expanded its agreement
- 9:10with Google, with Broadcom for several
- 9:12additional gigawatts. It maintains a
- 9:15partnership with Microsoft and Nvidia
- 9:17that includes tens of billions of
- 9:19aureate capacity and has added more
- 9:21than 220,000 GPUs through a deal with
- 9:23SpaceX. They are all in the same boat.
- 9:26If you think about it, software
- 9:28companies are financing infrastructure
- 9:30on the scale of the energy industry. In
- 9:33traditional software, the dream was
- 9:35operating leverage, that is, building
- 9:38the product was expensive, but selling
- 9:40the millionth copy was almost nothing.
- 9:44At the cutting edge of artificial
- 9:45intelligence, accounting works, well,
- 9:47the other way around. Maybe it's
- 9:50related, you will see. Revenues can
- 9:53skyrocket as the capital needed to
- 9:54produce the next dollar of revenue
- 9:56skyrockets. More customers means more
- 9:59inference. More sophisticated models
- 10:02can consume more compute. Staying at
- 10:05the cutting edge, at the cutting edge,
- 10:08requires training new generations, and
- 10:10each leap requires securing chips,
- 10:12electricity, and more capacity before
- 10:14you know how long you're going to keep
- 10:16the advantage you just bought. Maybe
- 10:21that's also why they want to ease off
- 10:23the gas. Could be, could be. Anthropic
- 10:26reported in May that its revenue run
- 10:28rate had surpassed $ 47 billion, which
- 10:31is impressive. Open is also growing at
- 10:33a rate that we've never seen before,
- 10:35but growing at a high rate right now
- 10:37doesn't answer the financial question
- 10:38that matters. The question is, how much
- 10:41capital do you have to invest and
- 10:43reinvest to keep growing at that rate?
- 10:45And there's no public answer to that
- 10:47right now. Nobody has yet demonstrated
- 10:49how much return each dollar invested in
- 10:50that computation is. Nobody. We have to
- 10:54be precise about this because this is
- 10:56not going to happen, this is very
- 10:58easily exaggerated. These companies
- 11:00aren’t broke, they’re trapped,
- 11:02which is different. They have plenty of
- 11:04demand, revenue growing at a historic
- 11:06rate, that’s true, and a business
- 11:07model that has yet to prove that it
- 11:09pays for itself. With each passing
- 11:12quarter in this race, the investment
- 11:14gap needed to stay ahead grows faster
- 11:16than revenue itself. That’s a
- 11:20terrible position for the guy in charge
- 11:22, because he can’t stop without
- 11:23handing the lead to the guy next to him
- 11:25, unless they all stop at once, unless
- 11:27they all ease off the gas, and unless
- 11:29there’s a reason for stopping that no
- 11:31one can argue with. Why would they all
- 11:35ease off? Open raised $ 122 billion in
- 11:39committed capital in March at a
- 11:41post-money valuation of $ 852 billion.
- 11:44It’s already approaching a trillion.
- 11:46Anthropic raised another $ 65 billion
- 11:48in May, reaching a valuation of $ 965
- 11:50billion, almost a trillion. These are
- 11:54amounts that would have seemed
- 11:55impossible in the private market three
- 11:56years ago, and that breaks one of the
- 11:58historical rules of technological
- 11:59capitalism. For decades, a company
- 12:02would end up going public because it
- 12:04needed to access a much larger pool of
- 12:07capital than it had available in
- 12:09private equity. But what happens when
- 12:12you can raise $ 122 billion without
- 12:14going public? Well, the stock market
- 12:17ceases to be an immediate necessity,
- 12:19and when it ceases to be a necessity,
- 12:21you can also postpone something that in
- 12:23that situation is frankly very annoying
- 12:26. Watch out, millions of investors are
- 12:30deciding every day how much your
- 12:32business is really worth. And that is
- 12:35what they have just postponed, what an
- 12:38IPO forces us to show. Let's leave
- 12:41Silicon Valley for a moment, away from
- 12:43all this mess, and go to Wall Street,
- 12:45where interesting things are happening
- 12:47there, because right now there are
- 12:49three things that are constantly mixed
- 12:51together and they are not the same.
- 12:54Need for capital, need for liquidity,
- 12:56and need for price discovery. Let me
- 12:58explain. A company needs capital to
- 13:01finance its growth. Your shareholders,
- 13:04founders or employees may need
- 13:05liquidity to convert part of their
- 13:07stakes into cash. But the market needs
- 13:11something very different, which is to
- 13:12find out how much a company is really
- 13:14worth, right? Openei and Anthropic have
- 13:18shown that they can raise absolutely
- 13:20gigantic amounts of capital without
- 13:22going public, and there are also
- 13:23secondary sales that give liquidity to
- 13:25certain shareholders without having to
- 13:27go through an IPO. What is very
- 13:31difficult to replicate here off the
- 13:33stock exchange is permanent price
- 13:35discovery. That is, thousands of
- 13:38managers, analysts and investors
- 13:40looking at your accounts, comparing
- 13:42expectations with results and repricing
- 13:44your company every second. Open
- 13:48confidentially filed its draft S1 form,
- 13:50which is the one that has to be filed
- 13:52with the SEC, the regulators there. It
- 13:56did so on June 8. Anthropic had done
- 13:58the same a week earlier and an S1 is
- 14:00less spectacular than a discussion
- 14:02about the end of humanity, right? But
- 14:05it can end up telling us a lot more
- 14:07about this industry because it requires
- 14:10a detailed description of the business,
- 14:12the financial situation, the results
- 14:14and the risks, and also the audited
- 14:16financial statements. And be careful,
- 14:19being private, you can't fool your
- 14:21investors either. Obviously, SoftBank,
- 14:23Amazon, Nvidia or Seollya do their due
- 14:25diligence before signing a check.
- 14:26Obviously, there's a difference. It's
- 14:29one thing to open your numbers to a few
- 14:31dozen sophisticated investors with whom
- 14:32you're negotiating a round. It's quite
- 14:35another to put them in front of the
- 14:37entire market and let Wall Street ask
- 14:40you every quarter if you really justify
- 14:42or are really worth $ 800 billion, $ 1
- 14:45trillion or $ 2 trillion. And the
- 14:47questions are going to be very tough
- 14:49from that point on. How much does it
- 14:51really cost to train a generation of
- 14:53artificial intelligence models? How
- 14:55much does inference cost once you
- 14:57account for all that infrastructure?
- 15:00What gross margin do you have left?
- 15:01What future cloud commitments have
- 15:03already been signed? How much Capex
- 15:06does it require to stay at the cutting
- 15:08edge? What part of the growth requires
- 15:10raising capital again? These are
- 15:12questions, are they going to be asked?
- 15:14When will sustainable free cash flow
- 15:17appear? And the most important one of
- 15:19all, which they're also going to ask,
- 15:20what return does each additional dollar
- 15:22I invest in computing produce? That's
- 15:26the real economy, the real economy of
- 15:27what's called frontier artificial
- 15:29intelligence. And we still can't see it
- 15:32as clearly as we've been able to study
- 15:34Microsoft, Alphabet, Meta. And
- 15:37Anthropic may soon give us that window.
- 15:39It's true, it's still preparing a
- 15:41possible IPO that, according to what
- 15:44has been published, could raise up to $
- 15:47100 billion with a valuation close to
- 15:49...Yes, yes, $ 2 trillion. It's worth
- 15:53clarifying this because these figures
- 15:54are being thrown around too happily.
- 15:56Anthropic doesn't need to raise $ 2
- 15:58trillion, that would be its valuation.
- 16:02What it would raise would be $ 100
- 16:03billion, and a company valued at $ 2
- 16:05trillion would have to answer something
- 16:06quite simple. What future benefits
- 16:11justify paying $ 2 trillion for it
- 16:13today? Well, Fortune magazine did the
- 16:15calculation using the usual multiples
- 16:17of the big tech companies on the NASDAQ
- 16:19itself, where they are going to be
- 16:21listed or would like to be listed or
- 16:23say they are going to be listed. And
- 16:26the figure that comes out is that
- 16:27Anthropic would have to end up
- 16:28generating tens of billions of dollars
- 16:30of annual profit to sustain that
- 16:31capitalization and today it barely
- 16:33makes any money. In that case, showing
- 16:36growth is not enough. The market wants
- 16:39to know how much additional capital
- 16:40needs to be invested to achieve that
- 16:42growth. Openai is in a different
- 16:44position. When it announced its
- 16:47confidential S1 document, that document
- 16:49, it already said that it had not
- 16:51decided exactly when to go public and
- 16:53that there were things that would
- 16:55probably be easier for it as a private
- 16:57company than being in the public domain
- 17:00at the expense of all opinion. And in
- 17:03August, CFO Sara Frear said internally
- 17:05that Open would list in 2027. That's
- 17:07where we'd go, although she had left
- 17:10open an earlier exit if the business
- 17:12picked up speed, and now, in the midst
- 17:14of the explosion of security concerns,
- 17:16Samman has already said that 2026 is a
- 17:18very bad time to go public, because
- 17:20things are so bad, arguing that a
- 17:22public company would have a harder time
- 17:24right now making security decisions
- 17:26that would temporarily harm its
- 17:28shareholders. In other words, he says
- 17:31that if they have to take their foot
- 17:33off the gas, it wouldn't exactly be a
- 17:35good time to go public. That makes
- 17:37sense. I would say yes. It shows that
- 17:40security, uh, well, is an excuse to
- 17:42dodge Wall Street in this case, right?
- 17:45But going private is financially very
- 17:47comfortable when you've just raised $
- 17:49122 billion and your capital needs are
- 17:52colossal. Not because you have anything
- 17:55sinister to hide, but because the
- 17:57private market lets you negotiate
- 17:59valuation and narrative with a small
- 18:01group of investors, while in public you
- 18:03would be forced to live with a daily
- 18:05price and quarterly results capable of
- 18:07dismantling your story in an afternoon.
- 18:12And I'm going to correct my own thesis
- 18:14precisely so that all this is more
- 18:16bearable. It is not true that Open and
- 18:19Anthropic have jointly decided to delay
- 18:21their IPOs, their stock market launches
- 18:23, until 2027. Openi is targeting 2027,
- 18:25yes, but Anthropic is still preparing a
- 18:27possible launch this fall. Therefore,
- 18:32in principle there is no coordination,
- 18:34but I don't need that to demonstrate
- 18:35what I want to demonstrate. What is
- 18:38interesting is another coincidence.
- 18:41Just when we are about to begin to see
- 18:43the real economics of these
- 18:44laboratories in public, their top
- 18:46executives are beginning to say that
- 18:47the race may be moving too fast. And if
- 18:52the race is regulated by law, the
- 18:54question ceases to be who runs the
- 18:56fastest and becomes who can pay for the
- 18:59permit to run, who can pay for
- 19:01compliance, and here comes the best
- 19:04argument against my hypothesis. I know,
- 19:07I have to launch it and it has to be
- 19:09put on the table. Anthropic is not
- 19:12proposing that any programmer who sets
- 19:14up an AI startup tomorrow has to hire
- 19:16500 lawyers and 20 auditors. It doesn't
- 19:18say that. The border regulations that
- 19:20the company defends use high thresholds
- 19:22. It's true. California Law SB53
- 19:25considers a major frontier developer to
- 19:27be someone who works with models
- 19:29trained above 10 to the power of 26
- 19:31computational operations, which is
- 19:33absolutely outrageous. and belongs to a
- 19:35group with more than $ 500 million in
- 19:37annual revenue. That's the official
- 19:40version. New York has ended up
- 19:42converging towards similar thresholds
- 19:44in its race race act, which is in
- 19:47theory the obligations that fall on the
- 19:49giants. Well, that design makes sense
- 19:52and dismantles the caricature according
- 19:54to which these companies want to
- 19:56prevent anyone from setting up a
- 19:58startup. I have defended it, not like
- 20:00that, but I have said that these people
- 20:03are basically trying to keep them alone
- 20:05. But the question does not disappear
- 20:07with this evidence, it is displaced
- 20:09because a barrier to entry does not
- 20:11have to prevent you from entering. They
- 20:13say they won't stop anyone from getting
- 20:14in. It can stop you from growing. You
- 20:17can start your company, raise $ 20
- 20:19million, then $ 100 million. It's the
- 20:21same, train better and better models
- 20:23and compete for years. The problem
- 20:24comes just when you start to be a
- 20:26serious threat. That's when you cross
- 20:29the threshold and start having to deal
- 20:31with auditing, reporting, evaluation,
- 20:33security protocols, external
- 20:35supervision, documentation, internal
- 20:37systems, regulatory relationships, and
- 20:39they may all be necessary. That's not
- 20:42the issue. The issue is who's going to
- 20:44be able to pay for them. Let's travel
- 20:47back to 2002 for a moment, as it will
- 20:49help us understand many things. You'll
- 20:51see. After Enron and Walcom, the United
- 20:54States passed the Sabarness Hoxley Act
- 20:56to strengthen controls, audits and the
- 20:58quality of financial information. Well,
- 21:01there were plenty of reasons to do it
- 21:03and it had an economic effect that has
- 21:05since been studied to death. A good
- 21:08part of the regulatory costs were fixed
- 21:12. The Government Accountability Office
- 21:15found that these costs weigh
- 21:16proportionally much more heavily on
- 21:18small companies than on large ones.
- 21:21These guys have studied that, you'll
- 21:22see. That's what happens in banking.
- 21:26When you need a certain compliance
- 21:27structure, the giant spreads it across
- 21:29tens of billions of revenue. But the
- 21:33little guy has to support almost the
- 21:35same infrastructure on a business that
- 21:37is actually minuscule compared to the
- 21:39big guys, the ones that were already
- 21:41there. I'm not saying that all
- 21:43regulation is a corporate conspiracy.
- 21:46Their work shaped the economic theory
- 21:48of regulatory capture. Regulations
- 21:50generate costs and benefits, and
- 21:52whoever has more money at stake also
- 21:54has more incentive to influence how
- 21:56they're designed. There's even a common
- 21:59expression in business strategy called
- 22:01regulatory mod. It's the regulatory
- 22:04moat. A regulation can be perfectly
- 22:06justified for safety reasons while also
- 22:08reinforcing the competitive advantage
- 22:10of whoever already has the resources to
- 22:12comply with it. And we're starting from
- 22:15a brutally concentrated market.
- 22:17According to Peachbook data compiled by
- 22:21Axios, Open AI, and Anthropic, they
- 22:23together took more than 60%of all
- 22:26venture capital committed to US
- 22:29startups during the first half of 2026
- 22:32from all startups. Anthropic has also
- 22:35significantly increased its political
- 22:37activity. Just look, over $ 3.5 million
- 22:40in federal funding in the first half of
- 22:43the year, on top of everything it spent
- 22:46in 2025, and now $ 40 million
- 22:48contributed to the public for public
- 22:50education and transparency policies and
- 22:53AI safeguards. And Anthropic itself
- 22:56specifies that this money cannot be
- 22:58used to support candidates. Perfect.
- 23:01We're not talking about corruption, I'm
- 23:03not. We're talking about a company that
- 23:06is legitimately trying to influence the
- 23:09institutional architecture of an
- 23:11industry where it has hundreds of
- 23:13billions at stake. And none of that is
- 23:17a conspiracy. This is a company
- 23:22defending what it invested in, the
- 23:24incentives. When this becomes a battle
- 23:27between doomsday believers and doomsday
- 23:30deniers, the only thing that really
- 23:32matters here is lost. Anthropic may be
- 23:34genuinely concerned regarding the
- 23:36potential risks posed by artificial
- 23:37intelligence. I'm not saying it isn't,
- 23:39but at the same time it could benefit
- 23:41from regulation that is very expensive
- 23:44for new entrants. It is possible for
- 23:46both things to be true at the same time
- 23:47, right? An economic incentive does not
- 23:51override a moral concern, and a moral
- 23:53concern does not make the economic
- 23:55incentive disappear. Not when slowing
- 23:58down is a bailout. We come to September
- 24:0212th. Dario Amodei publishes Frontier's
- 24:04Wimo Space. The central idea is that
- 24:07capacity building is moving faster than
- 24:09safety, and that the industry needs to
- 24:10keep up with the frontier. Anthropic
- 24:13commits to bringing in external
- 24:14evaluators with very deep access to its
- 24:16internal processes. And he also asks,
- 24:19hey, that the other big labs do
- 24:20something similar, they're not going to
- 24:22be the ones to do it alone. Amodei then
- 24:25proposes that the frontier companies,
- 24:27which are right there on the edge,
- 24:29coordinate standards and some speed
- 24:31limits, preferably through public
- 24:33regulation. Public regulation, and he
- 24:38acknowledges that some types of
- 24:40coordination could conflict with
- 24:42competition law, so he proposes some
- 24:44limited mechanism to allow for those
- 24:46conversations. And the guy ends with an
- 24:49international element. He says that any
- 24:51major slowdown in the United States is
- 24:53meaningless if China can continue to
- 24:55advance at its own pace, that the
- 24:56Chinese would have to be convinced.
- 25:00That the Chinese would have to be
- 25:01convinced, he says.
- 25:01And then Samman comes out to support
- 25:03the direction of all this. Elon was
- 25:05more in favor there, one after the
- 25:08other. We have to proceed calmly
- 25:09because it is very easy to jump from a
- 25:11coincidence to a conspiracy and there
- 25:13is no need to imagine any secret
- 25:14meeting, as he said when Modei himself
- 25:16is proposing in public that certain
- 25:17competitors coordinate, nothing happens
- 25:19. There explaining why he needs legal
- 25:21cover to do so. So far so good, there
- 25:23may be a totally legitimate reason. I
- 25:25am not saying there is not. Mancur
- 25:26Olson explained 60 years ago why a
- 25:29group with common interests often fails
- 25:31to act in its own interest. He said
- 25:35that what is good for the whole
- 25:37sometimes does not suit any of the
- 25:38members individually. And that is
- 25:41exactly what is happening here. If each
- 25:45laboratory initially thinks that
- 25:47slowing down on its own is dangerous
- 25:49because the rival will take advantage
- 25:51of the pause to overtake it, then it is
- 25:54normal that everyone may prefer a
- 25:56slower race and none can afford to slow
- 25:58down first, right? A common standard
- 26:02would eliminate part of that problem
- 26:04and could truly improve everyone's
- 26:07safety, and in that way these CEOs will
- 26:10save the world. Always helped,
- 26:13obviously, by appropriate regulation.
- 26:15But now let's read it in reverse. What
- 26:17happens financially when everyone slows
- 26:18down at the same time and it's also
- 26:20written in an official bulletin? Well,
- 26:22the burn rate goes down. An unlimited
- 26:25race forces you to continually train
- 26:27the next model because you don't know
- 26:28when your competitor will introduce
- 26:30something that will make yours obsolete
- 26:31. Well, if you slow down, you lengthen
- 26:35the training cycles and make the same
- 26:37capital last longer. The same capital
- 26:40that you've been given lasts longer.
- 26:42The marginal pressure on Capex and what
- 26:44it computes also goes down. Investments
- 26:47don't disappear, but the logic is no
- 26:49longer to buy all the capacity possible
- 26:51before your rival takes it away. The
- 26:53path they travel is also lengthened
- 26:55because the same 50,000 or 100,000
- 26:57million finances more months of
- 26:59operation and allows each generation of
- 27:01models to be amortized for a longer
- 27:03time. The mad race forces them to slow
- 27:07down. If you've just invested a fortune
- 27:09in training and deploying a system and
- 27:11six months later you have to replace it
- 27:13with an even more expensive one, the
- 27:15payback period for all that is
- 27:16ridiculous. A slower pace lets you
- 27:19squeeze out what you've already built
- 27:20for longer. And here's something even
- 27:23more subtle, uh, which is where I want
- 27:25to focus. A company that slows down
- 27:27because the numbers don't add up is
- 27:29admitting weakness. But a company that
- 27:32slows down because the safety of
- 27:34humanity demands it is demonstrating
- 27:37responsibility. It's the same
- 27:39accounting move, but with two very
- 27:41different headlines. Regulation turns a
- 27:44cash flow problem into a public virtue
- 27:46and turns a decision that would
- 27:48penalize your valuation into one that
- 27:50protects it. They're good guys. And I
- 27:54didn't make this up, uh? And it's not
- 27:56new either. There's a historian,
- 27:59Gabriel Colco, who documented something
- 28:01similar in the American railroads of
- 28:03the late 19th century. These guys have
- 28:05probably read it. An industry with
- 28:08enormous fixed costs, price wars that
- 28:09broke everyone, and competition that no
- 28:11one could keep up with. That industry,
- 28:15that railroad, where the big companies
- 28:17failed to stabilize the market due to
- 28:19private agreements that were
- 28:20continually broken. They ended up
- 28:24supporting federal intervention. The
- 28:28regulation that the official story
- 28:30tells as a victory of the public
- 28:32interest against monopolies was
- 28:34actually largely the mechanism by which
- 28:37those companies incapable of sustaining
- 28:39competition achieved by law the
- 28:41stability that the market was denying
- 28:43them. Now replace railroads with data
- 28:48centers, and replace price wars with an
- 28:50intense race for training. And hey, you
- 28:53have an identical map. History rhymes.
- 28:59If the slowdown is also accompanied by
- 29:01audits, verifiers, certifications,
- 29:03compliance departments, security
- 29:05controls, mandatory reports,
- 29:07sophisticated relationships with
- 29:09governments, then every additional mile
- 29:11is more expensive for the next guy. The
- 29:15giant already has lawyers, it already
- 29:16has hundreds of security researchers,
- 29:18it already has teams to fix all that,
- 29:20it already has a direct line to
- 29:22Washington, it already has gig workers
- 29:24hired, and it already has investors
- 29:26capable of writing checks for tens of
- 29:28billions. But the challenger, the one
- 29:31who could become a good competitor, has
- 29:34to build all that while trying to catch
- 29:36up. And that's where security and
- 29:39finances stop being incompatible
- 29:40explanations. A general slowdown can be
- 29:43justified by genuine security reasons.
- 29:46Yes. Yes. And at the same time be the
- 29:49best possible rescue for the financial
- 29:51structure of those leading that race.
- 29:54Sure. And therefore you don't need
- 29:57Amodei, Alman, or Mask to sit in a room
- 30:00to decide that. You just need their
- 30:03incentives to converge somewhere. Now
- 30:06look at the calendar again. It's very
- 30:08important. Openi presents its S1 in
- 30:12June. In August it talks about 2027. In
- 30:14September, in the midst of a wave of
- 30:16concerns about security, Alman explains
- 30:18that this would not be a good time to
- 30:20list. That doesn't prove that the
- 30:23security argument is false, but it
- 30:25produces an alignment of interests that
- 30:26, man, would at least deserve some
- 30:28attention from those media outlets that
- 30:30are talking about how the end of the
- 30:31world is coming because of AI and that
- 30:33it must be stopped, regulated, and
- 30:34something done. Staying private longer
- 30:39allows you to face a massive investment
- 30:41phase without having Wall Street asking
- 30:43you every quarter how much you’re
- 30:45burning and whether an $ 852 billion
- 30:47valuation still makes sense. If
- 30:50Anthropic does go public with its S1,
- 30:52we’ll finally get a wide window into
- 30:54the plumbing of a cutting-edge lab.
- 30:58Margins, costs, compute commitments,
- 31:00capital requirements, cash flow—and
- 31:03then we can start to answer what’s
- 31:05still very much hidden. How much does
- 31:10it really cost to manufacture the
- 31:11intelligence that promises to transform
- 31:13the world? Because we don’t know yet.
- 31:18So yes, so far this explains what the
- 31:20company is in for. We’re missing the
- 31:22other partner, the other partner in the
- 31:25deal, the one who hasn’t put up a
- 31:27dollar of his own and is taking the
- 31:29lion’s share. Who gets to hold the
- 31:31key? Let’s start with the argument
- 31:34that justifies it all. China. Amodei
- 31:36acknowledges as much himself. Besides,
- 31:37a Western slowdown only works as long
- 31:39as it doesn’t hand a geopolitical
- 31:41competitor a decisive advantage. That's
- 31:44why Anthropic advocates both curbing
- 31:47certain capabilities in the United
- 31:49States and tightening controls on
- 31:51advanced chips to limit Chinese access
- 31:53to that border infrastructure and
- 31:55combat what they call unauthorized
- 31:57distillation by smaller labs that can
- 31:59extract capabilities from Western
- 32:01models. Let me clarify. The logic in
- 32:06Washington is obvious. If you're going
- 32:09to hit the brakes, make sure the guy
- 32:11behind you doesn't seize the chance to
- 32:14pass. And they think that guy behind
- 32:16you is China. But look at what's being
- 32:19built with this justification. Export
- 32:22controls on chips, restrictions on data
- 32:24centers, audit systems, independent
- 32:27evaluators, computation thresholds,
- 32:29model registries, certifications,
- 32:31controls on the weights of all this,
- 32:33verification of certain training,
- 32:36agreements between governments. The
- 32:39thing is that each piece, each piece
- 32:41separately, can have a reasonable
- 32:43justification. I'm not saying it
- 32:45doesn't, but all together they are a
- 32:47licensing regime for computing capacity
- 32:49. And that's where the second
- 32:51beneficiary of this story comes in, who
- 32:54by the way has been waiting for decades
- 32:56, because there is a technology that
- 32:58states have not been able to fully
- 33:00control since it existed and that is
- 33:03the one that allows people to produce,
- 33:05copy and distribute information without
- 33:07asking permission. Remember, look, they
- 33:11arrived late with the internet, they
- 33:13arrived late with social networks, they
- 33:15have been fighting and losing with
- 33:17encryption for 30 years, with
- 33:19artificial intelligence they have
- 33:21arrived on time for the first time and
- 33:23they have arrived, moreover, invited by
- 33:25the industry itself. What's happening?
- 33:28Nobody is telling you this. Charles
- 33:30Tily put it in a way that is very
- 33:32unpleasant for any government. State
- 33:35building is similar to the sale of
- 33:37protection with the particularity that
- 33:39whoever offers to protect you is
- 33:41usually also the one who defines what
- 33:42you have to protect yourself from.
- 33:45There is no need to believe that
- 33:46governments are evil, no. Just look at
- 33:49what each one takes. If the risk of AI
- 33:53is existential, then whoever grants
- 33:55licenses decides who can train a model,
- 33:57with what data to do it, up to what
- 33:59computational threshold it can be
- 34:00carried and with what content it can be
- 34:03blocked. We're no longer talking about
- 34:06computer security; here we're talking
- 34:08about who approves the capabilities of
- 34:09a country's cognitive infrastructure.
- 34:12And we don't have to go to China to see
- 34:14where all this is heading. In Europe,
- 34:17we've been seeing for years how
- 34:19perfectly presentable security
- 34:21arguments end up being converted into
- 34:24control infrastructures. Everything is
- 34:27always for our own good. Whenever
- 34:30there's a danger to humanity, a
- 34:33regulation appears.
- 34:37Against CO2, bottle caps. Protecting
- 34:39minors ended in proposals to scan
- 34:41everyone's private communications.
- 34:43Fighting fraud ended in mandatory
- 34:45digital identity. Combating
- 34:46disinformation ended in mechanisms that
- 34:48decide which content deserves
- 34:49visibility. In no case was an evil plan
- 34:52necessary. A legitimate reason, a
- 34:54frightened public opinion, and an
- 34:56administrative opportunity were enough.
- 34:59Now add it all up, everything,
- 35:00everything, everything. Four or five
- 35:02companies with enough capital,
- 35:04computing, and lawyers to comply. A
- 35:06handful of governments, all of them,
- 35:08all of them, with the power to decide
- 35:10who comes in and who goes out. and a
- 35:12technology that will cut across
- 35:14productivity, research, software,
- 35:16business automation, and a huge chunk
- 35:18of the work that you and I do. We will
- 35:23have reduced the likelihood that anyone
- 35:25can build an intelligence that is
- 35:27extremely dangerous to everyone, and we
- 35:29will have concentrated the ability to
- 35:31build the most powerful machine in the
- 35:33smallest group we’ve ever had at the
- 35:35helm of a general-purpose technology.
- 35:39And with the key to those who claim to
- 35:42be watching over us. And my criticism,
- 35:46mind you, is not to say that Dario
- 35:48Amodei is lying. I’m not going to say
- 35:50that because it’s not true. He
- 35:51probably genuinely believes what he’s
- 35:52saying. And Jacob Coxson, whoever he is
- 35:55, and so do the researchers who are
- 35:57concerned about alignment. And the
- 36:00official who writes the regulation
- 36:01probably believes that he’s
- 36:02protecting the public. All of that may
- 36:05be true; I’m not saying it isn’t,
- 36:06it may be true. But the mistake, the
- 36:10mistake, my friends, is to believe that
- 36:12a sincere conviction removes a
- 36:14financial incentive or that an honest
- 36:16official removes the natural tendency
- 36:18of any administration to keep the damn
- 36:20key to what it regulates and what it
- 36:22doesn’t. That's why we should also
- 36:27look at our part, because we live in an
- 36:29information ecosystem where alarm is
- 36:31immediately rewarded. This video is
- 36:35going to be watched much less than if
- 36:37it had been titled AI is going to end
- 36:38humanity. A scientist who explains that
- 36:41we still can't accurately estimate the
- 36:43probability of that catastrophe is
- 36:45actually competing against a headline
- 36:47that claims that AI is going to kill us
- 36:48all before the end of this decade. You
- 36:52have to share this. You don't need to
- 36:54know about algorithms to guess which of
- 36:56the two is shared more. Then comes a
- 36:59popularizer who turns uncertainty into
- 37:01certainty because certainty gets clicks
- 37:02. Fear is normalized. The politician
- 37:06takes advantage of terrified public
- 37:08opinion and offers protection and
- 37:09everyone is reassured. And he also ends
- 37:11up asking the companies that build the
- 37:14technology what rules they think are
- 37:15reasonable to impose. Don't you see the
- 37:19circle? That circle closes there. I'm
- 37:22not crazy, it's crazy, it's different.
- 37:27So that's the verdict. We may need
- 37:29external verifiers, it is possible that
- 37:31certain capabilities require limits,
- 37:32and some future models may be too
- 37:34dangerous to deploy without
- 37:35implementing strict controls. I'm not
- 37:37saying they won't. At the same time,
- 37:39this regulation solves two problems
- 37:41that have nothing to do with your
- 37:43security. It allows an industry that
- 37:46hasn't yet proven that its accounts add
- 37:48up to stop burning money without
- 37:50admitting that the numbers didn't add
- 37:52up. And it also gives governments
- 37:57control—Mrs. Ursula will be delighted
- 37:59with this—control over the most
- 38:01important technologies of the century,
- 38:02just when they still had time to get
- 38:04their hands on them. Hey, fear up top,
- 38:06invoice below, and key in middle. And
- 38:10this isn't going to come free. If the
- 38:13intelligence that's going to drive the
- 38:15software, the accounting, the research,
- 38:17and a good part of your work is
- 38:19manufactured by four companies behind a
- 38:21wall that they helped design, and a
- 38:23government decides who crosses it, you
- 38:25know that you're not going to choose
- 38:26the provider. You pay the rent they
- 38:30charge you and you'll use the tool they
- 38:32authorize. That they authorize you. The
- 38:36price of a mandatory input is not set
- 38:38by the market, it is set by whoever
- 38:40controls the gate. And the only thing
- 38:43we can do is not buy the whole story,
- 38:45man. When in the coming months you hear
- 38:48that we need new audits, agencies,
- 38:50licenses, and evaluators to protect us
- 38:52from the terrible artificial
- 38:54intelligence, ask yourself also what
- 38:56cash problem each of these solutions
- 38:57solves and how much power transfers to
- 39:00whoever signs them. And when Anthropic
- 39:04finally publishes its S1, let's read it
- 39:07, we'll do it here. There will be more
- 39:10truth about this industry there than in
- 39:12100 prophecies about the end of the
- 39:13world. You'll see, we'll find that
- 39:17things were not going well. Because the
- 39:21decisive question of this revolution is
- 39:23not who will be able to create the most
- 39:25powerful intelligence, it's who will
- 39:27have permission to do so and who
- 39:29distributes the permissions too. That's
- 39:32why, before handing over the keys to
- 39:34the wall, we have to look at who's
- 39:36already inside that wall. Let's take a
- 39:39look. Don't be impressed by some of the
- 39:43headlines, at least not yet. Let's
- 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.