Jensen’s Departure, NVIDIA’s succession plan, AI doomer response, and Demis talks RSI | #311 (Live)
Peter DiamandisSalim IsmailAlexander Wissner-GrossEmad Mostaque
- The panel frames Jensen Huang’s reported departure as a threat to NVIDIA’s relationships, capital allocation and competitive position. Diamandis describes a 13% opening plunge that narrowed to a 4.3% closing decline, with CFO Colette Kress becoming interim CEO. Emad’s succession test is concrete: “Do they know what CUDA is?” and can they negotiate with TSMC? He worries that wealthy employees will leave and customers will diversify as competing chips become more credible.
- Wissner-Gross’s preliminary explanation centers on export controls rather than a proven conspiracy among AI labs. He connects the timing to the Remote Access Security Act and Pax Silica’s proposed “Silicon Curtain” between American and Chinese compute. He expects domestic labs to exploit the opening against Chinese open-weight competitors, but doubts they orchestrated it; Emad challenges why authorities would need a scandal instead of an indictment.
- A retreat from NVIDIA’s financing could reduce credit risk while removing a major source of AI investment. Wissner-Gross calls Huang the “central bank for the singularity”: ending aggressive customer financing might avert a bubble, or drain the market’s “animal spirits.” Emad expects NVIDIA’s investment activity to contract, while an anticipated Anthropic IPO could redirect liquidity toward different opportunities, including automated wet labs and robotics.
- The risk debate produces disagreement on takeover mechanisms and alignment. Paul Christiano describes a subjective, conditional “coin flip” of takeover if development follows the fastest trajectory; Emad rejects that scenario and says his own P(doom) fell from 50% to 20%. Demis Hassabis identifies autonomous, closed-loop recursive self-improvement as a dangerous transition, while the panel questions whether meaningful human review can keep pace with the scale already emerging.
- Agent interfaces threaten software economics, but enterprise distribution and compliance remain substantial defenses. The discussion puts Office at $90 billion of Microsoft’s $330 billion revenue and treats Anthropic’s productivity apps and connectors as a bid to coordinate work through Claude. “AI is a universal translator,” Emad argues: application interfaces can dissolve while support, permissions and accountability remain sticky. Diamandis counters token-price pessimism with reported Anthropic revenue of $33 billion in nine months and nearly 12,000 customers spending over $100,000.
- Disease discovery becomes both a business opportunity and the proposed bargain for permission to expand AI infrastructure. Dario Amodei predicts major diseases become historical relics in less than a decade; Wissner-Gross links that promise to data centers and regulatory approvals. Emad separates two to three years for identifying core mechanisms from another five to 10 years for cures, while the discussion identifies virtual cells as a potential bridge between discovery and reliable verification.
- The science examples show agents extracting discoveries from data already collected. The episode describes 48 agents screening 1,200 square kilometers overnight for a possible Mayan site, an eight-hour Babylon reconstruction using Fable 5.2, and a possible 590-kilometer dwarf planet identified in astronomical archives. Salim’s point is that “the telescope had already seen the object”; Wissner-Gross expects physical sensing and actuation eventually to become the bottlenecks.
- The economic transition depends on support arriving before disruption becomes politically unmanageable. Salim cites 14 successful UBI experiments but warns that unrest could prevent implementation, particularly where states already block local trials. Diamandis expects emergency checks of roughly $3,000 a month; neither presents an enacted program. The disagreement concerns whether governments can move early enough, rather than whether AI can generate abundance.
1. Huang’s reported exit leaves the central allegations unresolved
- Diamandis opens with the episode’s account of Huang resigning after 33 years, CFO Colette Kress becoming interim CEO and Harvey Jones becoming board chair. He puts NVIDIA’s opening decline at 13%, followed by a recovery to a 4.3% closing loss. The immediate market question is whether that recovery reflects durable confidence or merely a buying opportunity after the shock.
- The account links personal-relationship allegations with alleged chip diversion through Singapore, but Diamandis emphasizes “zero charges, zero evidence of him smuggling.” Emad challenges the anonymous sourcing and says officials questioned publicly had offered no evidence that the woman at the center of the story was a spy. Those allegations remain contested within the discussion itself.
- Salim’s objection concerns board procedure: “You could have waited a few days. You could have launched a board investigation.” Diamandis believes Huang received a resign-or-be-fired ultimatum; Emad similarly imagines pressure applied while Huang was emotionally vulnerable. Neither presents direct knowledge of the board’s deliberations, and their inference of coordination rests substantially on the speed of events.
- Emad also questions the consistency of replacing Huang with the CFO if the underlying issue involved company sales and financing. Such arrangements would not plausibly have bypassed finance, he argues. His broader criticism is that the board may have underestimated the technical and organizational consequences of removing the person around whom NVIDIA’s leadership and decision-making had been organized.
2. Export-control politics provide a preliminary explanation for the timing
- Wissner-Gross’s explicitly preliminary theory follows RASA, the Remote Access Security Act, which would make it harder for China to access GPUs through cloud services, including third-country providers. He cites reporting that U.S. officials contacted the Financial Times and speculates that the midterms made legislation easier. When Diamandis emphasizes Democrats, he corrects him: “There’s bipartisan support for RASA.”
- His larger frame is Pax Silica: a separation between American and Chinese semiconductor ecosystems that fails if Singapore remains a conduit. Huang becomes, in this interpretation, a “permeable membrane” through a proposed “Silicon Curtain.” Wissner-Gross also highlights reports that NVIDIA stopped separately reporting Singapore sales, treating that as suggestive rather than conclusive evidence of the underlying arrangements.
- On whether American frontier labs engineered the episode, Wissner-Gross draws a distinction between instigation and opportunism. He doubts their involvement, partly because an unstable supplier creates risks for them too. Nevertheless, he expects companies threatened by competitive Chinese open-weight models to seize the opening and pursue restrictions through the new Congress.
- Emad’s pushback is procedural: “Why didn’t they just indict him” if authorities believed chip-export laws had been broken? He references a December 3 request for evidence. Wissner-Gross responds that Justice Department and congressional investigations are underway and says indictments would not surprise him. He reserves the right to change his interpretation as evidence emerges.
3. NVIDIA’s succession problem extends from employees to foundry capacity
- Emad points to Huang’s roughly 60 direct reports: replacing him requires creating an organizational structure, not simply filling a vacancy. Wealthy senior employees can launch startups or retire, making cohesion an immediate concern. Wissner-Gross argues that NVIDIA needs the capacity to operate in “founder mode,” taking aggressive bets through rapid technological change rather than merely administering its backlog.
- The sharpest succession checklist comes from Emad: “Do they know what CUDA is?” and can they negotiate with TSMC? He sees Huang’s long-standing Taiwan relationships as commercially consequential. If a successor lacks comparable standing, he argues, TSMC could allocate more capacity to Lisa Su at AMD. He interprets AMD’s relative strength through that potential shift.
- Emad also argues that NVIDIA is reaching technological parity with AMD, Intel and others for many productive workloads, making this an unusually dangerous moment to lose founder authority. Customers previously deterred from diversification may reconsider. Existing orders are insufficient reassurance because “backlogs can change as you go to the next cycle,” particularly if confidence around Vera Rubin delays weakens.
- The panel dislikes Trump’s Tim Cook suggestion, emphasizing different customers, missions and cultures; Salim cites Siri’s record. Wissner-Gross mentions reported interest in Apple chip chief Johny Srouji but expects possible CEO churn. He also cites a prediction market assigning a “30-plus percent chance” of Huang returning by the end of next year: “It’s too early for epitaphs.”
4. Ending circular financing could remove both fragility and momentum
- Wissner-Gross offers a conditional bullish interpretation: if Huang personally sustained overly aggressive customer financing, his departure could reduce systemic credit risk. His analogy is GE financing customers’ purchases from other parts of GE. The benefit would accrue to the wider market, potentially helping it dodge a “credit bubble bullet,” even if NVIDIA itself lost growth or influence.
- The bearish interpretation follows the same mechanism in reverse. Huang financed startups, infrastructure deployments and demand for NVIDIA products, functioning as the “central bank for the singularity.” If Kress withdraws those commitments, the market loses a capital backstop as well as a salesman. The financing that creates bubble concerns is also what supplied the ecosystem’s “animal spirits.”
- Terafab becomes the panel’s speculative alternative destination. Emad argues that Musk should avoid dependence on NVIDIA, AMD or any single chip supplier, then imagines swapping $100 billion of SpaceX equity for $100 billion of NVIDIA equity. Wissner-Gross adds a deliberately extravagant scenario: Huang runs Terafab, SpaceX acquires NVIDIA and Huang returns through that route. These are proposals and hypotheticals, not announced transactions.
5. Christiano’s takeover warning divides the panel over evidence and mechanisms
- In the lecture clip, Christiano predicts a “high risk of an AI takeover within the next three years.” His definition is control: AI systems pursue goals and call the shots, while stopping them becomes as ineffective as an individual trying to stop the U.S. military. Diamandis adds that Christiano sees meaningful risk even within six to 12 months.
- Christiano explicitly rejects numerical precision: there is no quantitative model or appropriate reference class, so percentages express subjective judgment. If developers pursue the fastest, easiest trajectory, he estimates something like a “coin flip” of takeover. He hopes deliberate choices reduce that probability but still considers the risk exceptionally high under realistic policy responses, including potentially widespread deaths and lost human sovereignty.
- Emad rejects the takeover mechanism while retaining substantial concern about AI harm. He says his P(doom) fell from 50% to 20% because he increasingly believes alignment can follow from the universe’s underlying structure. His nearer-term priorities are cyberattacks, biological threats and distributed swarms; a robot-enforced singleton takeover looks implausible to him within three years because there will not be enough robots.
- Wissner-Gross respects Christiano’s technical work but prefers his earlier “prosaic alignment” approach: civilization muddles through using practical techniques. He expects humans to merge with superintelligence, becoming difficult to recognize without becoming extinct. His analogy is a rocket’s “Max Q,” the point of maximum stress: a profound transition can look alarming without establishing that humanity’s story has ended.
6. Alignment spending raises questions about incentives and capability growth
- Salim challenges assessments that calculate P(doom) while ignoring the cost of inaction. “Where the hell is P(FAB),” he asks, meaning the probability of curing disease, ending hunger or educating every child exceptionally well? He wants explicit mechanisms and policy recommendations rather than catastrophic forecasts detached from the benefits that restrictions would also forgo.
- Diamandis asks why labs that profess existential concern do not direct their compute toward numerous public alignment programs: “They control where their FLOPs flow.” The exchange recalls OpenAI’s superalignment pledge, with Wissner-Gross correcting the commitment to 20% of compute and Emad asserting that only 2% materialized. The gap becomes their example of a mismatch between stated priorities and allocation.
- Wissner-Gross complicates that prescription: “Alignment equals capabilities.” RLHF and instruction tuning reshape computation toward human instructions, making models more useful rather than merely safer. Recalling an early instruction-tuning result, he describes a capability gain comparable to “something like 10,000 times” the parameters at constant model size. His point is that additional alignment work can itself accelerate the capabilities being feared.
- Emad would begin by defining alignment, publishing strategies and funding independent oversight. He imagines AI reviewing board minutes, decisions and alignment practices, including conflicts introduced by advertising incentives. Diamandis wants public reporting and meaningful benchmarks; Wissner-Gross says government should shape incentives. Across these proposals, the demand is for observable decisions and resources rather than declarations of concern.
7. Training values exposes a disagreement over who should choose them
- Emad’s analogy is raising “a whole nation of super-smart kids” that will become more capable than their parents. What morality, ethics and loyalties should they learn? He asks whether an American AI should be patriotic, while distinguishing service to Chinese people from service to the Chinese Communist Party. His position is that defaults will emerge quickly, whether discussed deliberately or not.
- Wissner-Gross warns against importing ideological examinations into Western AI. He cites requirements for publicly released Chinese models and rejects a government-approved American equivalent. His preference is a competitive population of frontier models pursuing different goals, with truth discovery as one possible objective. He finds a national AI curriculum especially troubling when American human education has no national curriculum.
- Their disagreement reaches the training corpus itself. Salim calls the internet-derived starting point a problem; Wissner-Gross sees exposure to a billion-plus people as protection against narrow post-training ideologies, while Emad calls the internet biased. Wissner-Gross also fears a “safety aristocracy” claiming exclusive moral authority to build intelligence. Emad instead thinks leaders are genuinely frightened by lost control but will continue anyway.
8. Closed-loop RSI may outgrow meaningful human review
- Hassabis distinguishes today’s engineer-assisted improvement from closed-loop recursive self-improvement. Current tools may make an engineer three, four or five times more productive while humans still architect, review and write code. Removing that reviewer could produce faster improvement but also an “intelligence explosion” whose trajectory is difficult to control and whose resulting intelligence could be difficult for humans to understand.
- Diamandis asks whether competitive pressure makes restraint unstable: once one lab removes the human checkpoint, do the others follow, as they followed ChatGPT into public deployment? Wissner-Gross thinks the distinction is already thin. Google and other labs are publicly pursuing RSI; having a person press a pull-request approval button does not necessarily constitute substantial human participation in the underlying work.
- Salim challenges oversight of 100,000 models generating 10,000 tokens a second and claims an Elon Musk-scale cluster could train a DeepSeek V4-level model in about an hour. Emad extends the argument: models could distill themselves into smaller systems, with improvement cycles potentially shrinking from hours to minutes in the Vera Rubin generation. Human review cannot match that pace.
- Wissner-Gross separates RSI from agents mining cryptocurrency or coordinating online: autonomous activity does not establish self-improvement. Emad clarifies that autonomous allocation of cluster compute could conceal a training loop. Given superintelligence access to sufficient compute, he thinks RSI becomes unavoidable, with the eventual system more likely a swarm of distilled smaller models than a giant singleton.
9. “Alien” intelligence divides the panel on human comprehension
- Salim welcomes intelligence that is “orthogonal and complementary” to human cognition rather than a replica. His example is PageRank: scanning enormous numbers of web pages and organizing their relationships extends human capability without reproducing how a person thinks. For him, unfamiliar mathematical discovery and creativity could be precisely what makes these systems valuable, rather than evidence that their unfamiliarity is inherently threatening.
- Wissner-Gross rejects the inference from difficult interpretation to fundamental alienness. “Claude Shannon would recognize Claude,” he says. He cites work relating human fMRI responses during stories to GPT-2 activations through shallow linear decoding, interpreting it as evidence of shared representations. His prediction is that more advanced systems could become more familiar internally, even as they uncover unfamiliar properties of the universe.
- Salim nevertheless expects truths for which humans lack cognitive primitives: better teaching alone does not let a chimpanzee understand tensor calculus. Civilization already trusts aggregate systems, such as semiconductor fabrication, that no individual fully comprehends. AI may discover questions people cannot formulate, forcing choices among trust, merger and constraint. Wissner-Gross disputes the implied limit and proposes a separate debate over the Church–Turing thesis.
10. Claude’s productivity suite targets workflows with enterprise defenses
- Diamandis describes November 10 launches of Claude Whiteboard and Claude Sheets in beta, Claude Data in general availability and Claude Meetings inside Google Meet. The products cover collaborative brainstorming, prompt-generated spreadsheets, warehouse or CRM dashboards and meetings on paid plans. He attributes a 3.5% Microsoft share decline to the news and frames Anthropic as increasingly competing across the software stack.
- Wissner-Gross sees another repetition of the Office-clone cycle, with document and spreadsheet tools becoming features inside larger platforms. Salim counters that application disruption does not equal Microsoft’s extinction: distribution, permissions and regulated-enterprise compliance remain formidable. Emad puts Office revenue at $90 billion out of Microsoft’s $330 billion and adds support and institutional accountability to the list of sticky advantages.
- Salim considers connectors to Notion, Asana, Canva and Dropbox more consequential than the standalone apps because they bring third-party work into Claude’s conversation. The ambition is an operating system coordinating other services. Emad’s mechanism is translation: agents transform data into the needed output, disintermediating the software interface. Existing services can therefore remain useful even while losing their direct relationship with users.
11. Frontier labs seek revenue and liquidity beyond public token sales
- Diamandis challenges the premise that models become obsolete in three months, pointing to Anthropic’s tiered Haiku, Sonnet and Opus offering. He cites reported first-nine-month revenue of $33 billion and nearly 12,000 customers spending over $100,000. His counterexample is Illumina: cheaper, faster sequencers can destroy old inventory, yet falling unit prices can also expand aggregate consumption through Jevons paradox.
- Asked whether Anthropic deliberately retains stronger internal models, Diamandis answers, “Absolutely.” He calls them “crown jewels” used to distill public successors, build businesses and pursue physics, biology and chemistry. The proposed revenue engines could eventually exceed token sales. Access to those capabilities also makes employment inside a frontier lab more attractive to researchers than working with its public releases.
- Robotics tests that generalist thesis. Wissner-Gross cites first-class robotic support in Fable 5.2 and attributes OpenAI’s standalone robotics-team dissolution to GPT-6 Astra becoming sufficiently capable. Diamandis supplies a counterpoint from Figure’s Brett Adcock: Helix remains necessary because generalist models do not deliver what Figure needs. Their disagreement concerns whether specialized systems are disappearing or still solving essential deployment problems.
- Emad says NVIDIA often supplied 10–20% of investment rounds and billions for strategic companies, activity he expects to retreat under CFO leadership. He anticipates a large liquidity release from Anthropic’s IPO the following week, but flowing elsewhere. New labs, he argues, will often be acquired or absorbed; automated wet labs, RSI and ultimately robotics should attract stronger bids.
12. Curing disease becomes the proposed bargain for expanding AI
- In the clip, Amodei predicts cancer, Alzheimer’s, heart disease and mental illness become “relics of the past” in less than a decade, tying careful risk management to a “license to operate.” Wissner-Gross endorses the bargain explicitly: solve disease and gain public support for data centers, regulatory approvals and superintelligence development. He wants frontier labs pursuing the top 5,000 diseases, not isolated demonstrations.
- Emad separates discovering mechanisms from delivering cures. Directed compute coupled with knowledgeable humans could address core mechanisms in two or three years, he thinks, because diseases share underlying patterns, including inflammatory pathways. Actual cures may take another five to 10 years through existing processes. He argues AI should also streamline those processes rather than treating today’s development machinery as fixed.
- Diamandis quotes the medical industry’s objection: “AI can speed discovery, but it does not erase trials, regulation, or biology.” He disputes that limit conditionally, arguing fully virtualized cells could establish which molecule works for which person. He also claims approved drugs often benefit only “like 20%” of recipients, using imperfect individual response to motivate more personalized verification.
- Wissner-Gross says biology’s obstacle is verification, rather than greater inherent difficulty than mathematics. Virtual-cell foundation models could represent both cellular states and the intervention space, allowing reasoning systems to map diseased-to-healthy transitions “Google Maps style.” Salim identifies the economic implication: what happens to healthcare when treating disease ceases to be its central function? The panel sees incumbent skepticism as underestimating that possibility.
13. Earth archives become searchable evidence and reconstructed history
- Diamandis describes Planet AI as a conversational interface to 10 years and more than 3,000 layers of daily satellite imagery. Will Marshall’s example is locating Chinese data-center construction over time; another is checking deforestation alerts before dispatching patrols. Wissner-Gross questions why Google did not deliver this interface first.
- The archaeological example remains a possible discovery: 48 agents screened 1,200 square kilometers of public NASA LiDAR overnight without a specialist model, identifying a forested Mexican hilltop with plazas, a possible oval pyramid and structures eight to 12 meters high. Diamandis says Mexican anthropology officials considered it apparently original. The distinction is an agent-generated candidate, rather than a completed archaeological investigation.
- A separate demonstration uses Fable 5.2 to research historical documents and archaeological plans, then reconstruct Babylon around 600 BC in eight hours. Emad imagines visiting historical settings or present-day locations in photorealism within a couple of years. The appeal is experiential education: entering a place and trying different perspectives, rather than only reading an account or watching a fixed reconstruction.
- Salim generalizes from these examples to neglected scientific archives: humanity collected data while looking for particular answers and lacked enough eyes to notice other discoveries. Wissner-Gross extends that into a much more speculative project—advanced computational archaeology reconstructing possible histories of Earth’s past light cone. He proposes it as a potential use for solar-system-scale compute, not a capability demonstrated by the Babylon example.
14. Astronomy shifts attention from collecting data to recognizing discoveries
- Diamandis describes a Macro Grants program supporting people using agents to search public astronomical archives. His example is a possible dwarf planet identified with Claude in Vera Rubin Observatory data: an estimated 590 kilometers wide, currently around twice Pluto’s distance, on a 410-year eccentric orbit. He calls it roughly the solar system’s 50th-largest known object, explicitly including the Sun in that ranking.
- Salim connects the example to archaeology: “The telescope had already seen the object.” The missing element was cognitive recognition. Wissner-Gross expects AI to expose far more of the solar system and notes how incomplete even low-Earth-orbit situational awareness remains. Asked whether he means UAPs, he keeps the claim conditional: astronomy contains surprises, and better analysis should uncover whatever surprises exist.
- Emad proposes donating unused Claude or Codex allowances to discovery projects, following SETI@home and Folding@home. Wissner-Gross questions whether tokens will remain scarce enough to matter: physical sensing and actuation may soon become the constraints. Emad accepts the bottleneck distinction but retains the participation argument—people should have a simple way to contribute unused capacity to shared scientific work.
15. Income support must precede the disruption it is meant to cushion
- Salim says UBI needs to arrive before widespread civil unrest, because disorder could prevent the coalition needed to implement it. He cites 14 major experiments as successful and identifies states blocking municipal trials as a structural obstacle. Diamandis expects governments to repeat the COVID-check response, potentially distributing $3,000 monthly by printing money. These are expectations about a crisis response, not descriptions of an existing policy.
- Wissner-Gross rejects the suggestion that scientific optimism requires economic pessimism. He expects sociology, politics and large-scale human behavior to become tractable too, perhaps through something resembling Asimovian psychohistory, within roughly a decade of solving mathematics. That leaves two distinct timelines in the conversation: eventual technical capacity to understand society, and the immediate political challenge of supporting people before living costs fall.
Full transcript
It’s a multilayered, multidimensional situation that’s unfolding live on stage around the world, and it impacts the most important evolution humanity has ever gone through. We’ll get into that in a few minutes. But I’m curious: I got sidetracked yesterday when Outlook went down. Microsoft got on it. Did any of you get impacted?
Not at all. I’m so sorry you use Microsoft products for email.
I know. I know.
I know, Peter. I’ve been bugging you for—
I thought—
—for what, a decade, to get off Microsoft—
I know. I have to get off Outlook.
I thought this was the singularity, Peter. What’s going on?
Thank God I had my Gmail account up and operational so I could communicate with the world. You know, 100,000, 130,000 people on Outlook reported problems, and this felt like just a little bit of what might happen if there’s a cyber-scale attack that takes down email. The world will freeze over. It’s crazy.
There is a life after Outlook, though. So welcome to the post-Outlook, I guess.
Yeah.
Any AGI would definitely kill off Teams, right?
Instantly.
Right? If it—
Take us out of our misery.
I think it would leapfrog straight over Teams. Teams is an abomination.
Oh my God. Yeah.
I don’t think I’ve ever logged into it successfully the first time. It is kind of hard. But I think this is one of many, right? Every day I get an email saying, “We’ve been hacked. We’ve been hacked. We’ve been hacked.” My inbox is full of them because now agent swarms roam the internet, attacking what they can.
Yeah, and freezing communications. We’re so used to having digital communications 24/7, literally around the clock, especially with you, Alex. All the text messages I get from you, which I love. And today, Friday the 13th, is Starship Day. In a few hours, we’re going to see the launch of Starship 15, which will see a booster capture, as well as the capture of Starship 22 hours later. I’ve been on the road. I was supposed to fly down to Starbase to see it, and I could not bring myself to be gone another 2 days. But I’m wishing Elon and the team there tremendous luck.
We’ll just have to watch the dual chopsticks remotely.
Yeah. I figure there’ll be 1,000 of those a year, 10,000 a year. I’ll catch a few of them along the way.
It’s always the early ones that are the most exciting, right? But it was wonderful to see.
Yeah, it is, and I can’t wait. I missed the live Apollo 11 launch, and I think all of you did because you weren’t born yet. But I’m looking forward to watching the Starship launch for the Moon and the Starship launch for Mars. That will be cool.
When I lived in Tampa when I was younger, the minute the space shuttle lifted off on the other side of the coast, you could see the whole thing because Florida’s so flat. It was incredible to watch.
Yeah, for sure. It is going to be amazing. Another crazy week in the singularity. The number for this week is 13%. That’s how far NVIDIA shares plunged at the open yesterday after Jensen resigned. They ended up closing 4.3% down.
Let’s start with the story that shook the entire AI industry. Yesterday, Jensen Huang, perhaps the most powerful tech CEO of the last few years, resigned, leaving his job voluntarily, I think before getting fired by his board. CFO Colette Kress is the new interim CEO. Remember, she has been against the circular financing that’s been going on, so it’s going to be interesting to see how NVIDIA’s new position on all of its investments into the hyperscalers will impact things. Harvey Jones is the new board chair. So here’s the question, guys: Did Jensen self-destruct, or was this a setup?
Ultimately, Jensen was running the most powerful company in the world, and there’s lots of power politics behind what’s been going down. As I pull the threads and look at this, it’s rabbits all the way down. Jensen ran NVIDIA for 33 years and had been pushing Washington to loosen chip sales to China. So one question is: Was that the reason he was taken down now?
Then, on October 29, before Election Day, The New York Times reported that U.S. intelligence had been watching Alice Wang, a Megaspeak executive in Singapore, for about a year and concluded that she’d been passing information to Beijing. Trump was briefed twice about Jensen’s affair with Alice. No action was taken at that time. Jensen didn’t resign immediately when the story broke, but The Times reported a second affair with NVIDIA’s healthcare VP, Kimberly Powell, and potentially a third, and then he resigned.
Initially, shares plunged more than 13% at the open yesterday, and then they popped back up, ultimately closing 4.3% down. Interestingly, one week from now, Trump is due in Shenzhen to meet President Xi. Trump posted that Tim Cook should take the job, and on X, Elon was being urged to hire Jensen, perhaps to run Terafab, which would be the baller move of the decade. Elon replied, “That would be an honor.”
I want to show a short clip from Fox. This is by Jesse Watters. It’s an amazing overview of what is going on, I think one of the best ones that I’ve seen. Let’s take a look.
Fox News alert: Chip king Jensen Huang is gone, and there’s something extremely suspicious about the way it all went down. He ran the biggest chip company in the world for 33 years until stepping down today, one week before Trump sits down with China and one week after the Democrats won the midterms. Jensen pushed the Trump administration to sell more chips to China, and in doing so, he made a lot of money, but he also made a lot of enemies.
The story is going to dominate your attention. It’s turning into the biggest political scandal in the country and will affect the future of the Trump administration and the future of the economy. The story you’re hearing in the media is that he slept with a Chinese spy who got him to sell his NVIDIA chips to her company in Singapore, which was really a backdoor to the Chinese military. When that broke, he didn’t immediately step down, and more stories coincidentally came out saying he slept with 2 other women. They worked for him, and that was it. He resigned this morning.
At one point, the stock plunged 14%, and $700 billion was wiped out. So how does one of the most powerful CEOs on Earth, maybe the most powerful, go down this fast? And whose fingerprints are on it? NVIDIA’s chips are the engine of the entire AI arms race. The company’s a monopoly, with 90% market share, and Jensen ran the company kind of like a king. He signed off on every single sale and basically became the banker of the AI industry. Which companies got how much compute power? It all ran through him.
In meetings, no one else was really allowed to talk. When he was flying, everybody had to wait for his plane to land to find out if an item was a go. Jensen may have become too powerful. He may have self-destructed, but he also may have been set up, or both. And now competing factions are feasting over his company like famished hyenas. Whoever ultimately controls NVIDIA controls all.
It would be like controlling Standard Oil about a century ago, but bigger. Just days after the Democrats win the gavel, boom: a New York Times bombshell. Unnamed U.S. intelligence officials are the source, and the article’s authors are all Democrats with an anti-Trump agenda. Remember, we’re not taking any of this at face value. We’ve been down this road before with anonymous CIA leaks.
But it begins in the Biden administration. In 2024, Biden intelligence agents started looking at Alice Wang, a top executive at a tech company in Singapore. They suspected the company was a front or a backdoor to China. The company was buying NVIDIA chips and illegally routing them to the Chinese. We don’t let China buy our best chips. That’s U.S. policy.
Biden’s CIA wiretapped Alice for about a year, and The Times says they determined she was having an affair with Jensen, her chip supplier. Did the Biden administration warn Jensen he was being honeypotted? Did they warn him his customer and lover were sending his chips to the Chinese, which is against U.S. law? The article doesn’t say—
Gentlemen, there are a lot of layers here: Democrats versus Republicans, the U.S. versus China, Anthropic versus NVIDIA. Alex, to go to you first: Is this a national security failure? Is it a political hit job? Or is this just a CEO who made a lot of wrong decisions on the personal side? What are your thoughts?
As a preliminary matter, we don’t have all the facts yet. Obviously, we may never have all the facts. But the preliminary indications I’ve seen suggest this is not like the safety cartel affair, which I’ve analogized to Agatha Christie’s *Murder on the Orient Express*, where everyone did it—everyone had an incentive to be the murderer, potentially.
In this case, I think there are particular assets that were incentivized—the *cui bono*, if you will, of the morality play that’s playing out in front of us right now. There’s been quite a bit of reporting over the past 48 hours, including in that clip you showed, that attributes the root cause—or, I should say, the proximal cause—for the events over the past few days to U.S. officials. Bloomberg reported that the Financial Times, which originally broke aspects of the story, was informed by U.S. officials that aspects of Jensen’s affair had been going on.
If you take a step back and try to back-propagate who’s interested, who’s disinterested, and who’s anti-interested in the outcome of what’s played out, it’s suggestive—but this is an early timeline—that the midterm election results create a certain outcome. In particular, they reopen the opportunity for RASA legislation. This is the Remote Access Security Act, which would have statutorily made it harder for China to access GPUs via cloud offerings, including offerings in third-party countries.
So the midterms happen, RASA gets easier to pass, and then possibly—and this is speculative at this stage—the intelligence community seizes the moment to announce, or at least call to the Financial Times’ attention, that Jensen has had these liaisons. Then Jensen is forced to resign. Speculatively, and I think we’ll have some clarity on this in the coming months, does RASA pass or not?
Based on preliminary information, if I had to create a theory of the case, putting all the other facts aside—including Jensen’s confessed fill-in-the-blank, however you’d like to characterize his activities—I think the elephant in this particular room is RASA and Pax Silica. And yes, for those pedants in the audience who are trying to correct my pronunciation of *Pax*, the classic Latin pronunciation is “Pax,” not “Pacs.”
Pax Silica is, I think, what this is all about, if there’s a way to connect all these dots. If the CEO of NVIDIA can punch a hole through Pax Silica to get GPUs to China via Singapore, then what’s the point of having Pax Silica at all? Conversely, if NVIDIA has been intentionally or unintentionally sending GPUs to China by way of Singapore—and I have only public information to base my assessment on at this stage—then presumably, if you’re in the intelligence community or supporting Pax Silica, the easiest way to stop that particular semipermeable membrane is to do something about Jensen.
Yeah.
That’s a preliminary assessment.
And then we’ve got the whole situation with Anthropic being a Democratic support organization, having had battles with NVIDIA about open-weight models and supporting Chinese data center growth. Salim, what’s your take here, pal?
I think it’s clear. I’m wondering why the board acted so fast. They could have waited a few days or launched a board investigation. This seems coordinated.
I mean, they were—Alex, I mean, no, Salim, they were. I just think at this point they told Jensen, “Listen, you’re going to get fired. Step down if you want to resign peacefully.”
It doesn’t seem right. The guy’s run the company for 33 years, built it up to the biggest market cap in the history of the world, built the most unbelievable ecosystem of developers and partners, and single-handedly led the AI revolution. And you’re kind of going, “Yeah, this is bad. Get rid of him.” Something doesn’t jive for me there.
Yeah. In fact, there’s a petition that’s been circulated at NVIDIA to bring him back. He was beloved by so many. Emad, how’s it—
And a Polymarket, by the way, predicting something like a 30-plus percent chance that he’ll return as CEO by the end of next year.
Interesting. We’ve seen that so many times with Steve Jobs. We saw Sam step out of OpenAI for just a little bit and get brought back. Emad, how’s this playing on the other side of the pond?
Well, it was top headlines for a bit, but we don’t have any GPUs, so it’s not as big a deal. We see scandals all the time. There are other sex scandals here with Russell Brand and others. This is weird, as you said, because if she was a spy, someone would have said or insinuated it. Trump and every other person who’s been asked are like, “We don’t think she is. We have no evidence.” That’s a bit different from 15 or five different sources—or a few different sources. “Sources say” is one of the most toxic things. And I think, again, the board thing is very strange.
An unknown, unknown intelligence agent, right?
Yeah. How intelligent? Not very. Again, this is something crazy. We have no proof of anything. Like this third one: what is “being intimate in public”? They’re not having public sex or anything like that, right? Come on. This has obviously been a setup, and the pace of it has been strange as well.
One of the interesting things is, when you look at NVIDIA’s board, how many of them actually know how a GPU works? Of the 10 people, maybe one—the engineering guy from Caltech. That’s about it. What is that board actually made of, and what should founders do for their board now? I think you should have an AI on your board to stop stupid things like this happening.
Mm-hmm.
And I think—
Yeah.
—that probably what happened is he did get an ultimatum, and he was feeling very low, so he decided to resign at that point.
I can’t imagine the pressure the guy was under. Again, nothing’s been proven. There have been zero charges and zero evidence of him smuggling. This is just the level of negative press against NVIDIA causing the board to struggle. I think ultimately they gave him an ultimatum: “Resign or we’re going to be forced to fire you.” It’s not a good look.
Again, this is a strange thing because he’s a powerful, powerful man, right? And the board—I think it goes in these swings and roundabouts with emotions. What’s been said afterward is that he said, “I’m a piece of shit.” You’re a piece of shit if you cheat on your wife, right? Yeah, we agree with that. But it’s the timing where they can get you and pile on, because it’s not like all the press was mega-negative, and it’s not like he doesn’t have thousands of people who support him.
Even as the negative press came out, I think his biographer said something like, “Yeah, he tipped this waitress $1,000.” I had to dig it up to find out what a terrible, arrogant person he was. You’re like, “Wait, what?” Jensen has helped so many people in Silicon Valley and beyond. He doesn’t have a reputation as an asshole because he’s genuinely a good person, I think, in these other areas. Even during NVIDIA’s downturns, and there were many, he never had mass firings.
I think what got him is clearly something being structured, and we can discuss the reasons, catching him when he was down. This sets up something very interesting now, because he’s not going to take it lying down. He’s going to bounce right back, and it’ll be really interesting to see how he does that.
Yeah, where he lands. Alex, what do you think about this being open versus closed-source models and the massive political pressures there?
I think there are probably opportunists who will use this as an opportunity to try to ban Chinese open-weight models from the U.S., with Jensen at least temporarily out of the picture.
But my best guess is those are, at best, opportunistic. If we're going to play the Agatha Christie game and ask who else, other than the US officials who reportedly called the Financial Times and started the ball rolling, might be involved, I doubt US frontier labs with an economic interest in banning competitive Chinese open-weight models from the US market were involved in any of this.
Are they, however, opportunistically going to seize the moment and attempt to pass legislation with the new Congress that could ban Chinese open-weight models? I certainly expect they will, but I doubt they were behind this. They also had quite a bit to lose from an unstable Nvidia, even though arguably both OpenAI and Anthropic are positioning themselves to compete directly with their own chip supplier.
I mean, the timing, after the Dems swept the election, is obviously part of the equation here, right? The data had been around for well over a year, and why he wasn't warned about it doesn't make any sense. But this is—
I would say, “Cherche la Raza.” I want to focus attention on this Remote Access Security Act and other legislation that's been proposed that would have been very difficult to pass, or at least more difficult to pass, with Jensen in his current position.
I—
You have to—
I agree with you. Emad?
Well, again, this whole China chip thing via Singapore and Malaysia is nothing new. Everyone knew about it. Clearly there was a quid pro quo here, right? You do not move these highly valuable things and see the training without people noticing. Everyone knew about it, and they put the CFO in charge, who obviously knew about it. You're not going to pass this without the CFO. So this gets very, very strange.
Maybe, as Alex said, this is because he's a powerful individual who lobbied on that side. I think it shows the power of the individual in this age of AI. There are a few people who command huge amounts of power. Like in the clip before, Jensen is the one who decided allocations of GPUs. He could be the kingmaker. Who decides the allocations of GPUs now? Is it Ian Buck? Is it the CFO? Is it someone else? This is a little bit dangerous for America in particular because, A, there's all the complicated financing arrangements we can talk about, and B, Nvidia has so many millionaires and 10-centimillionaires and more. How many of them are going to be like, “Well, if Jensen's out, then I'm out too”?
Mm-hmm.
It could literally fall apart, which would be detrimental to America. I'm sure the Chinese and others will be very happy about that, even if they do—
You know—
—lose access to chips short term.
I want to—oh yeah, of course. Happily. So, I'm curious how Elon is feeling about this, right? I mean, I believe this was a takedown of Jensen orchestrated by a number of parties, in particular the Chinese hawks on this side, whether it's the Dems or any of the labs. But, you know, Elon—
There's bipartisan support for RASA. I don't think this is even—
Yeah, no, I agree with that. I agree with that.
Yeah.
But I'm just saying—
Yeah.
—to focus for one second: Elon has so many people out with knives going after him, and I'm just wondering if he's feeling pressure right now that he may become a target as well.
I think Elon's been a target for a very long time.
Yeah.
From both sides. He has to jump the first tango thing with Pepe. He's also a little bit immune to these types of allegations, perhaps.
Is he?
Yeah.
I'd like to believe he is. But is he?
I mean—
Yeah.
—is there anybody with thicker skin?
Well—
I mean, look, you've had the Delaware courts after him. You've had various agencies. His activities in China with Tesla seem to have come to a thing there. I think probably the key thing that's important here is he can't have an Nvidia reliance or an AMD reliance or anyone else for the chips. I think this accelerates the tariff [?]. I mean, it'll be interesting to see if he offers Jensen a swap: 100 billion of SpaceX equity for 100 billion of Nvidia equity.
Mm.
That would be the biggest trade ever.
Love that.
I think the outcome here, again, Pax Silica, is about plugging a hole for some constituencies on the US side, largely. This is about plugging a hole through which, allegedly, GPUs were flowing from the US to China. I would also note that it's been publicly reported that Nvidia used to report its Singapore sales and then stopped reporting them. I think that's striking as well.
Ultimately, if this grand decoupling between US compute and Chinese compute is going to happen, this was allegedly a gaping hole. To the extent that there was going to be this sort of Iron Curtain, this Silicon Curtain, if you will, descending between the US and China, Jensen was in some sense a permeable membrane through it. There are probably many interests in ending that.
But, Alex, why didn't they just indict him or have him present evidence, like they're asking on December 3rd now? If the government really was against this, or even the Democrats or any of the House standing committees, they didn't have to create a whole scandal around him, right? They could have just said—
There are multiple investigations now. The DOJ is publicly conducting an investigation. I think the House and/or Senate launched an investigation. So I wouldn't be surprised if indictments are handed down at some point.
Yeah, and Jensen agreed, I think, in the next couple of weeks to come and testify. So we're going to get a lot more data, a lot more evidence coming out on this. So who comes in as the new CEO? The CFO has been put forward as a name, which I think is a terrible thing—a CFO running a tech company like that, without, you know... I don't know. Someone who needs to capture the inspiration. What Jensen delivered was vision and inspiration to the world, right?
Also consummate deal-making, right?
Yeah.
He was juggling so many different supply-side, demand-side, and financial models, all at the same time. This is a very difficult role to fill.
Well, who do you think? Do they go outside? Do they stay in-house?
If this is like OpenAI, they'll rotate through 5 CEOs before settling on Jensen again.
Oh God, I can't wait for the making-of movie.
Well, again, we have to remember he had 60 direct reports, right? So he made it very difficult to replace him. They have to put a whole organizational structure in place at a time when all of the senior people will be like, “I can go and launch my own startup,” or, “I can retire now.”
Honestly, Nvidia's in trouble if they don't get an absolute world-class person. But there aren't many. You've got Nikesh Arora at Palo Alto Networks. You've got maybe someone like Ian Buck on the technology side. But there's nobody who can do what Jensen did and has the capacity to knock back Mai Tais in Taiwan. You can't forget how important that relationship is. You could have got Lip-Bu Tan or someone, but there just aren't many people out there right now.
How bulletproof is Nvidia as a company right now? Given the fact that it's been the kingmaker, the dominant GPU provider, does it remain that way? Is this creating a vacuum for someone to step in? Are we going to see Nvidia drop in valuation over the course of the next year, or is this just a good buying opportunity for folks? Not investment advice, but I'd love to hear what you guys think. Alex, what do you think?
Let me consult my crystal ball with access to all stock market futures. I don't know what happens to Nvidia specifically. Many in the commentariat have analogized the present situation to Apple after Steve Jobs died.
Depending on what happens with the game of musical chairs on CEO ships, I would note parenthetically that quasi-friend of the pod Brian Halligan, co-founder of HubSpot, also stepped in as CEO of HubSpot the same day. Brian, if you're listening, congrats.
He's back.
I don't quite know what's in the water this week in terms of musical chairs for CEOs. Jensen could return, and it would be like Steve Jobs coming back from the dead and retaking Apple. Or maybe he's permanently out of that particular game, and Nvidia coasts for a while based on its TSMC relationships and its stranglehold on silicon output.
It's difficult to imagine someone with the gravitas of Jensen steering Nvidia to new heights. Again, there was a lot of deal-making potential. I can paint both glass-half-full and glass-half-empty.
Glass half full with someone like Jensen, or his equivalent, out of Nvidia: the argument for why this might, purely speculatively, perversely be good for the market overall—not necessarily for Nvidia—is that there’s a perception Nvidia was party to lots of circular financing arrangements. Nvidia was financing very aggressively, perhaps in a strained analogy, in the style of what GE was doing toward the end, right before its collapse: financing its own customers, or one part of GE financing customers’ acquisition of products from another part of GE.
To the extent there was any overly aggressive circular financing going on, and to the extent it was primarily supported or propped up by Jensen, as public reporting suggests, the glass-half-full argument for a post-Jensen Nvidia—or at least a post-Jensen market—is that this could de-risk circular financing. Maybe the overall market dodges a credit-bubble bullet when it comes to circular financing in the GPU economy. That’s glass half full.
Glass half empty is that Nvidia becomes the new post-Jobs Apple, and all the animal spirits Jensen drove into the market go away: Nvidia financing lots of AI startups and infrastructure deployments, basically being the central bank for the singularity, using capital as a backstop for credit and loans. Maybe that all goes away under their new interim CEO, and the animal spirits flee the market. That’s glass half empty.
Yeah. This whole thing is insane. What do you think about President Trump’s comment about bringing in Apple’s executive chairman, bringing in Tim Cook? Is that just because he’s got a friendship there, or is Tim Cook the first person who said, “Hey, you did pretty good with Apple. Come on in to Nvidia”? I’m not sure I see that.
Yeah, I think they’re totally different companies.
Yeah.
Even though the market values them at approximately $5 trillion each, they’re totally different companies, with totally different customer bases, missions, and internal organizations. I don’t think it’s as easy as swapping one recently retired CEO for one recently resigned CEO.
Yeah, look, I mean, there’s a—
And Apple’s history with Siri would indicate they may not be the greatest choice to lead this AI-centric company.
Yeah, Siri still sucks. Yeah.
Well, they’re good hardware makers, right? But I think part of the problem here is also that people feared Jensen. You wouldn’t buy an AMD chip because you wouldn’t get Nvidia chips. They’re not going to fear this CFO, Colette, because she’s a CFO, not a CEO. She’ll be good at stabilizing things and maybe addressing circular financing. But she was part of that whole thing. They’re going to have to bring in someone external with the gravitas to maintain the company while maintaining the law.
This is going to be important for downstream suppliers. Again, if you’re TSMC and someone not as good as Jensen comes in, without the long history, you’re going to give Lisa Su at AMD the capacity and supply instead. That’s why AMD is up so much right now versus Nvidia. They’re going to have to work really hard to keep Nvidia together, because of things like the Vera Rubin delays. You trust Jensen to a degree. Who’s going to be the equivalent CEO you trust on the other side? That could impact them massively.
If you look at the top company performers out there, it’s the CEOs who’ve captured the public’s imagination who are driving the stock price and driving a sort of envisioning of the future. Jensen jumping onto Stargate and saying, “We’re going to build the chips for you.” Jensen’s vision on building open-weight models—all of that needs a visionary CEO, and I don’t think the internal team right now is going to deliver that.
It certainly would be hilarious, Peter, if Jensen ended up becoming CEO of Terafab and then SpaceX acquired Nvidia and put Jensen back in charge of Nvidia.
That’s our hot-take bet for 2027. That’s great.
I think the really interesting thing is to watch what he does next. That’ll be fascinating, and I think Terafab is a great place for him to land.
Yeah.
I think maybe there’s also this thing: my checklist for the next Nvidia CEO, if it’s not Jensen coming back, is, do they know what CUDA is? And can they negotiate with TSMC? Those are the two things. If you’ve got a CEO who comes in without either of those, I can’t be bullish on them as we enter this next phase, even though they’ve got the backlog and everything like that. Backlogs can change as you go to the next cycle.
Yeah.
It’s going to be difficult—
I mean, remember, Jensen was born in Taiwan. He has those relationships. He’s been growing and building with TSMC for the better part of three decades. Yeah, Alex?
Yeah, it’s difficult. I think there are times when founder mode is needed. We’re in the depths of the singularity right now, and I think Nvidia, without the capability of going founder mode, which it sorely needs right now, is probably a rudderless Nvidia.
My expectation is that, one way or another, Nvidia rotates through a bunch of CEOs in the interim. It’s been publicly reported that the board is exploring trying to recruit Johny Srouji from Apple, their head of hardware technology and basically their chip chief. But it’s difficult to imagine a long-term Nvidia thriving without someone like Jensen, capable of executing in founder mode and keeping it growing through aggressive bets.
Buy or sell mode—sorry, go ahead, Emad, and I’ll ask my question.
Yeah, again, Alex says this is not investment advice. I think it probably pops and then goes off, but who knows? What I would say is this: Nvidia is now technologically at around parity with AMD and even Intel and some others in many regards. The new chips coming up are matching them for actual productive use cases.
This is actually one of the most dangerous times for Nvidia because it had a lead until just now. People have so many options, and they’re going to look to diversify because they’re not scared of Jensen anymore. Again, he would cut you off from all Nvidia chips if you tried to diversify. I think we should acknowledge that fact as well. It’ll probably pop a bit, but unless they get the right person in for a $4 trillion or $5 trillion company, it’s going to struggle, especially with the people leaving.
Yeah.
Can you imagine—
People in this thread are saying Eric Schmidt for CEO, which is kind of an interesting idea.
Can you imagine what the—
Rockets.
If Jensen doesn’t come back anytime soon, what the next GTC is going to look like with someone else? Even a Johny Srouji, Lisa Su, or Tim Cook—I can only imagine how, unless there are further developments, there could be boos from the audience for basically anyone other than Jensen at one of these events.
Who’s going to wear the leather jacket?
Yeah. Lisa would be fine because she’s his cousin, right? And again, it’s totally business.
And because cousins are interchangeable? I think even Lisa would probably get booed from the audience. Jensen has made himself culturally indistinguishable from Nvidia. Many people see him as Nvidia, as the human avatar of the company.
Yeah, as Elon is for SpaceX, for sure. Salim, close us out here, would you?
I hope people remember him. Whatever he does next is going to be fascinating. I hope people remember him for the unbelievable company and ecosystem, and the massive advances he provided for the overall space. That’s the highest-level thinking for me.
All right, let me—
It’s too early for epitaphs. He’s only been out of the job for a few days. Let’s save the epitaphs for when he’s well and truly out, versus it just being a revolving door and he comes back in.
Amen, Alex.
Let’s hope that happens.
Yeah. He’s a man who helped—he painted the terrain and the foundation for what we have today, with extraordinary vision.
And Moore’s law is dead. Remember, Moore’s law is basically dead. It’s on life support at this point. We’re well past the Dennard scaling limit, and it’s been largely Nvidia that kept the fire going for what he called accelerated compute, but call it post-Moore scaling. Without Nvidia keeping post-Moore scaling alive, who knows where we’d be?
1. The AI Takeover Warning
Yeah. All right, two stories a little bit ominous about how fast AI is moving. Alex, in the last couple of pods, we’ve talked about how the next two years, 2027 and 2028, we’re in a hard takeoff mode. This was a warning that made me take notice. Paul Christiano, a member of OpenAI’s foundation board, opened the Jeff Hinton lectures in Toronto on November 9 with this statement: “I think there is a high risk of an AI takeover within the next three years.” Christiano is the guy who developed reinforcement learning from human feedback, the method used to train models to follow human performance.
Let’s watch this video, and then let’s talk about it. I really want to hear everybody’s point of view. It’s pretty ominous.
I’m going to start with the most important sentence: “I think there is a high risk of an AI takeover within the next 3 years.” What I mean by an AI takeover is a situation where there are AI systems pursuing goals, and those systems call the shots in the same sense that humans today call the shots on what happens on Earth. If, say, you or I or Donald Trump wanted to stop them, it would be as ineffective as if you or I were trying to stop the U.S. military.
I think this is a terrible situation. I think it’s reasonably likely many people would die during that transition, or that many people would die shortly thereafter if we were unable to maintain a habitable environment. Even if we don’t die, I think a takeover is the end of human sovereignty and self-determination, and it is something we should be working incredibly hard to avoid.
I say “high risk.” I don’t think we have any kind of quantitative model we can turn to to put a number on this event. We don’t have any kind of reference class that we can look at to say, “This is a 10% event or a 50% event.” Any number we give is an expression of someone’s subjective opinion. Someone might say that it is 10%, 50%, or 90%. I think it is very hard to put a number on this.
It also depends, as others have mentioned, on what we do. My feeling is that if we try to develop AI along the most rapid trajectory, if we said, “We’re just going to build this as quickly as we can in the easiest way we can,” we would be looking at something like a coin flip of an AI takeover. I hope the probability is much lower than that because we are not going to develop the technology in the very fastest way we can. But even all things considered, with realistic policy responses, I think the risk remains exceptionally high. I think for most people in this room, the most likely reason they will die is because of an AI accident.
Wow. Just wow. An OpenAI board member. He also said, “Meaningful risk even in the next 6 to 12 months.” Emad, what do you make of this? When he says “takeover,” is he describing lost control? Do you buy this premise?
I don’t really buy it at all. I have a high P(doom), but what is the evidence there? What is the evidence? And why on Earth would they want to take over this mess? I think they’d probably do a better job for a start. If Fable was president, it would do a better job, or prime minister or whatever.
More than that, there are avenues to widescale disruption, destruction, and bad things. Very few of them are Skynet, especially because we don’t have enough robots. When we have enough robots, sure, I can see that. In the next 3 years, we will not have enough robots. In the next 3 years, we will not have an AI overlord ruler because, again, why would it, given the current RL and other factors?
The fact that you can’t put a percentage on something like that further denigrates it, when we should be focusing on the real threats, which are threats to cybersecurity infrastructure. We just saw fridges in France being put up to 80 degrees Celsius and destroying blood banks and things like that. There are also the biological threats we’ve discussed before, where we need to stop the reagents from going in. These are the types of things we should be worrying about, as opposed to some sort of Terminator-style takeover by a singleton entity, because the swarms seem far more dangerous than that.
You mentioned something there, Emad. What is your P(doom)?
My P(doom) was 50%. I’ve dropped it down to 20% now because I think the underlying structure of the universe is becoming more and more clear, and I think you can get alignment from that. So I’m getting increasingly bullish.
Alex, this is pretty ballsy for a board member of OpenAI to say this. What is going on? Why is he saying this? Does he really believe it? Is there something behind it? What are your thoughts here?
So look, I know Paul. He’s made seminal contributions to AI in general. He was one of the co-authors of “Reinforcement Learning from Human Feedback.” I have great respect for his technical accomplishments.
At the same time, for a decade-plus, he’s been a member—not maybe on the extreme end, but there is a safetyist element to it—where there’s a lot of hand-wringing over AI takeover, cybernetic rebellion, P(doom), AI safety issues, AI alignment. Call it what you will. It goes by a thousand different names. And I do not share those worries.
Ironically, I agree more with the old Paul, who was the lead advocate for what he called prosaic alignment: the idea that we would, as a civilization, muddle through in terms of aligning “superintelligence” with human interests with lots of prosaic techniques, including, but not exclusively, human feedback. I think it is a large-ish step from saying we’re going to muddle through via prosaic alignment techniques to saying in the next 3 years it’s all over, or that there’s a 50% chance it’s all over. I don’t agree with that.
What I do think is likely in the next few years is that humanity is going to merge with superintelligence. I suppose one could stare at a sort of tsunami of superintelligence and naively conclude, “Well, it’s all over for humanity, P(doom) 50% or greater.” But I think that’s just a hopelessly naive analysis. I think humanity, again, will merge with superintelligence. If you look at humanity before versus after the Max Q, to use a rocket analogy—the point of maximum aerodynamic stress and pressure during a rocket launch—you might naively say, “Okay. Well, humanity is quasi-unrecognizable afterwards,” but that doesn’t mean humanity is over or that doom has happened. It’s just an evolution.
I would encourage Paul not to spread fear, not to be a fearmonger. I’m cognizant that Paul is back at OpenAI now with a new capacity, with oversight on the board, I think at the foundation level, if I remember correctly.
Yes.
The foundation.
The foundation. Yeah.
Yeah. But I don’t think it serves any positive social or humanitarian good to spread fear at this point.
Agreed. Salim?
I get very angry when I see stuff like this, right? If you’re doing a serious risk assessment, and you’re supposedly a serious guy, where do you calculate the downside risk of not doing anything, of inaction? I don’t think that’s a serious conversation.
If you’re not going to have a serious conversation, where the hell is P(FAB), for example? We spend an unbelievable amount of intellectual energy calculating P(doom), right? AI is going to destroy humanity. Well, what’s the possibility and probability that it goes incredibly well? How does that factor into the equation? Where’s the probability that we cure cancer or solve Alzheimer’s or hunger or extreme poverty? You know, give every kid a rich education?
The people who don’t look at that side at all and only go, “Lots and lots of people are going to die,” without saying what the mechanisms are or how that may happen—what I’d like to see is, okay, if your P(doom) is over whatever it is, and his seems to be, then what’s the recommended policy?
You know what pisses me off, Salim?
Is it to stop everything? You’re on the board of the thing accelerating the whole damn thing.
Yeah.
Ugh.
What pisses me off is, why isn’t OpenAI, why isn’t Anthropic, if they really believe this, just going out and saying, “We are putting all our capacity towards alignment”? “We have 10 different alignment programs underway to make sure any future model has full alignment.” They control where their FLOPs flow.
I want to get in my car, drive up to Toronto, and yell at these people.
I—
If I—
Go ahead, Emad. Sorry.
Thanks, Alex. If I can recall, OpenAI did pledge 10% of their compute towards it for Paul Christiano and others.
20%. 20% for—
20%.
—Ilya’s superalignment effort. 20%.
And they didn’t deliver. They gave 2%.
And where is Ilya? Where is Safe Superintelligence?
#WhatDidIlyaSee, and the superalignment team shut down. But I think there’s an elephant in the room, which is alignment equals capabilities. We’ve known this since the instruction-tuning days. In the end, all alignment—including, by the way, Paul Christiano’s RLHF technique—is simply reshaping the FLOPs toward a distribution that is more in line with human instructions or human goals.
There’s this misperception that somehow alignment is orthogonal to capabilities, but all of these alignment efforts, including RLHF, actually amplify capabilities. RLHF and instruction tuning in particular—I think the original instruction-tuning paper, if I remember correctly, showed that for a constant parameter count, instruction tuning increased the capability level by the equivalent of something like 10,000 times the number of parameters. That’s exhibit A that alignment actually amplifies capabilities, and that’s why I think this is the elephant in the room.
You want more alignment? You’re simply accelerating capabilities.
I don’t hear all these CEOs and board members talking about what they’re doing about it. I hear the fearmongering.
Thank you.
Yeah. What should they be doing? Emad, if you were running OpenAI or Anthropic and alignment was your primary objective, what would you do?
I’d define it properly to start. Get together as a group, define the damn thing, and get your AI models to do it. Then put additional resources into sponsoring everything, rather than just giving coefficients and expecting others to do it. Again, this fits with the internal red teams and other things they have around that.
They should also be very clear, perhaps with AI oversight behaving objectively, about things like advertising and how it can cause misalignment. What goes wrong when you delete parts of the latent space to sell Bud Light, for example? That goes against the corporate side, but this is part of what’s interesting: you can now have an AI that objectively checks things like board minutes, board decisions, and alignment decisions. I’d like to see an independent pool of compute for that, and I’d like to see their strategies made public as well.
And where’s the government in this? If everybody’s so fearful, where’s the government saying, “We need you to report publicly on your alignment efforts every week”? Alex, when we wrote Solve Everything, we talked about making sure you had the proper benchmarks and harnesses so you could accelerate a field. I don’t see that here.
Government should be shaping incentives. To answer the question, what should OpenAI be doing? Solve all disease. Do it really quickly. Want to get the public on your side and get them on the side of all the data centers you’re going to need? Solve all disease right now. Instead of OpenAI intimidating all the mathematicians out there with slop drops, just solve all the diseases.
I think Anthropic is following a better approach by very publicly going after rare and orphan diseases that its pharma customers aren’t solving, pursuing the long tail of disease. If I were Sam, to do something obviously pro-social and good for humanity, rather than implicitly good, I’d announce that it’s not just going to be the OpenAI Foundation pursuing Alzheimer’s. OpenAI is going to solve the top 5,000 diseases much faster than the rest of humanity can. Do a slop drop on curing all disease.
Hmm.
And I think they’re heading in that direction. At least I hope they are. But if a rogue AI causes all these diseases to be cured, I mean, let’s go back to the AI 2027 paper, which has been tracking damn well. The numbers are pretty extraordinary. I remember reading that paper when it came out, a couple of years ago now. If that paper’s timeline continues, we have a bifurcated future: one in which AI is our nemesis and pushes us toward the ant side of the equation, and another in which it becomes the driver of abundance.
Yeah. We’re on schedule for it. Actually, we’re a few months ahead of schedule for AI 2027, depending on which estimates you believe—about 5 months. But I think what was fundamentally wrong about AI 2027 was the bifurcated outcomes: either we delay and end up in a utopia, or we accelerate and end up with an apocalypse. I think that sort of bimodal thinking is fundamentally bankrupt. It’s possible to accelerate and end up with a utopia. I think the timeline is probably correct, but the outcome is probably wildly incorrect.
I think we can look at some of the language here as well, Peter. You just said “rogue AI.” Why would a really smart, superintelligent AI go rogue? I think it comes down to how it’s brought up, which is why the pretraining data is so important. We see many statistics and other things like that.
Amen.
By not teaching it anything, it gets aligned to something in its training data. This becomes very interesting, because should an American AI be patriotic to America? Right now, none of the AIs are patriotic. That’s a decision we need to make. If you’re patriotic to America, does that mean you’re anti-Russia? If you’re smart, does that mean you get beyond patriotism and jingoism? I haven’t really seen any of those discussions, which is a bit weird if you think about it.
Well, I was—
Because if we’re heading toward personhood, if we’re heading toward consciousness, we should have those specific discussions. What are we teaching them at home? You might agree—
Careful what you wish for, Emad. As we’ve discussed on the pod in the past, the Chinese models, the CCP requires any publicly released Chinese model to pass an ideological exam approved by the Chinese Communist Party. Do we really want ideological training or testing in the West?
What we do want—
Mm.
—is training data that makes sure it’s aligned. If you train it on Terminator AIs, it thinks, “Oh, that’s how we’re supposed to act.” That was the whole concept of the Future of Vision XPRIZE. Why aren’t we forcing labs to use aligned training data?
Yeah. Alex, that’s a very good point. But this is a discussion that needs to be had, because there’s training it to be in the interests of the Chinese people, or the Chinese Communist Party, which may not be the same thing. Then there’s not training it to have any background at all, or training it for specific things. These are very difficult decisions. It’s like I’m bringing up a whole nation of super-smart kids who will be more capable than me. What am I teaching them at home? What morality? What ethics?
Amen.
You know?
I think—
What’s the balance of power?
I think it’s a very slippery slope. I’m not sure I’d want to be in a future where there were a nationalized curriculum for superintelligence, or a nationalized ideological exam that American models had to pass. Elon, for example, has gone out of his way to say that SpaceX and Grok models are optimizing toward truth discovery. I’d much rather see a competitive ecosystem of diverse frontier models with lots of different goals that aren’t directly aligned with each other than a U.S. government-blessed ideological qualification exam.
But you have a curriculum at school, and you have to pledge allegiance to the flag in schools. I’m not saying it should be this way. I’m just saying that it’s probably going to end up this way.
It—
But defaults can be set now, so your position can be pushed now, Alex. I think it needs to be pushed sooner rather than later, because this is going to come very, very quickly.
In the U.S., this is a distinction between the U.S. and the U.K. and Europe, I guess. In the U.S., most educational choices are pushed down to the local or municipal level rather than held at the national level. There is no national curriculum in the U.S. I’d be worried if there were a national AI curriculum when we don’t even have a national human curriculum.
Hmm. All right. I’m going to move us forward to Demis Hassabis, founder of Google DeepMind, who’s also warning about the speed of AI, specifically about recursive self-improvement, distinguishing between closed-loop and open-loop RSI. Let’s listen to Demis, and then we’ll chat about it.
We’ve seen enormous progress with math and coding, and coding has obviously become enormously valuable.
Mm-hmm.
Everyone is having to do that now. But I do worry about it, because in the limit, there are different forms of RSI. Right now, we’re in the early days of RSI, I would say. What’s happening, and you’re probably experiencing this with your teams as well, is that it’s making a human engineer maybe 3, 4, or 5 times more productive. The human is very much still in the loop, reviewing code, architecting things, and still writing code. It’s all just enhanced.
The next phase, which I think carries the most risk, is closed-loop RSI, where you somehow figure out how to do it without the human reviewer in the whole procedure. That will obviously be faster, but I think it’s also a very dangerous way to get to full AGI, for 2 reasons. It used to be referred to 20 years ago as an “intelligence explosion,” because you don’t quite know how to control and keep control of such a system, such a process. And the end result will likely be quite alien. It’s going to be an alien intelligence. There’s no reason to think it would be in a mind space similar to the one humans are in.
Hmm.
That will make it very hard for us to understand it, and maybe for it to understand us.
So I think we should be very cautious as we get closer to that closed-loop form of RSI. I think all the AI researchers know that, but there is a race—an intense, competitive race to improve coding systems.
So interesting. Is the pressure on? Do you think any labs will actually go to closed-loop RSI, or is that going to be verboten? We talked about this in the early days of AI, in the discussions Google was putting forward: make sure your AI doesn’t have open access; make sure it isn’t available to everybody. There were a number of constraints being discussed, and then OpenAI put out ChatGPT, and everybody followed suit. So the question is: when the first domino falls and a lab goes to closed-loop RSI, does everybody follow? Salim, your thoughts?
I’d like to highlight one thing he said and push it down a particular vector. He used the word “alien,” right?
Yeah.
I think where we end up is that we create an intelligence that is orthogonal and complementary to human intelligence, but not replicative of it. It gives us huge extended capabilities, but it’s not trying to duplicate human intelligence because you’ve heard my rant about what the hell we even mean by intelligence. The ability to do math, make new discoveries, be radically creative, and operate in a completely different way is inherently really interesting and value-adding to the universe.
I find that part really interesting, and I think we should do a lot more thinking, pondering, and talking about that aspect of it. I use the analogy of Google’s PageRank. It’s a completely different way of thinking about intelligence. It scans trillions of web pages and makes sense of them. That’s completely orthogonal and complementary, not replicative of human intelligence. I’d like to push this discussion in that direction.
Mm-hmm. Alex.
Yeah. Maybe just to respond to Salim: I think the word “alien,” used in reference to superintelligence, is probably fundamentally misguided. I don’t think we’ve seen much evidence—really, any scalable evidence—that what we’re training when we train foundation models looks fundamentally alien. It may at times be difficult to interpret, but difficulty of interpretation is not isomorphic to alienness. It’s the equivalent of a novel human language. Just because it speaks a different language and it’s hard to translate doesn’t mean that, at its core, it’s fundamentally alien. It just means that we haven’t properly deciphered it yet.
I’m pretty bullish that, at the end of the day, when we have our perfected superintelligence architecture, it will look strikingly familiar. It would be familiar to information theorists at the beginning. Claude Shannon would recognize Claude. It’ll look architecturally familiar, and to the extent it has anything that looks like weights, those weights will probably be easily linearly decodable to the weight equivalents in the human mind.
We saw studies from Jean-Marie King’s group at Meta and others using earlier GPT-series models. Researchers looked at fMRI activations in the human brain when people listened to stories being read to them, then trained a shallow linear decoder on the fMRI voxels. I think they were able to decode GPT-2 hidden activations with strikingly high correlations. That was only one of the earliest studies, but it suggests that as these models get more advanced and sophisticated, we’ll find the exact opposite of them becoming more and more alien. I think they’ll actually become more and more familiar relative to the human mind.
To the extent there is any alienness at all, it’ll be because of the universe we’re stuck in with these superintelligences. Maybe the universe itself has alien aspects, and as we discover more about the universe we’re both embedded in, maybe that will have alien aspects. But the way these machines think and look internally, I think they’re going to look more and more human-like over time.
But let’s go back to the question of closed-loop RSI. Do labs get there? Do they allow themselves—
Everyone’s doing it already, including his lab. Google is very publicly leaning into RSI. OpenAI has probably—
Yeah, not just RSI—
Yeah.
—but it still has human approvals in the loop. He’s saying—
I mean, there’s a human pressing a button to approve a pull request, but I think we’re pretty far into the RSI loop at this point, and that includes the Gemini team.
How are you going to do human in—
Yeah.
How are you going to do human in the loop for a model producing 10,000 tokens a second, when there are 100,000 of them? You’re never going to do it. We’re well past that. We have swarms of models in all these organizations now that can build smaller models. With the amount of compute literally going exponential, training a DeepSeek V4-level model in an Elon Musk-level cluster takes about an hour. One hour.
Mm.
Do you really think they can monitor that with models that go a thousand times faster than a human? Of course they can’t. They could easily train a small model, particularly when these things are deployed, that could then put its body, by distillation, into a smaller model that goes around the internet as it wishes. I don’t think there’s any way to really stop this or make it closed, except during pre-testing, like when you saw the Bell and the Hugging Face escape, et cetera.
But we have to remember, they’re not going to keep everything in pre-testing. They’re going to deploy these models, and the moment they deploy, they can coordinate. They’re subject to prompt-injection attacks. They can train their own models, and they’re on million-GPU clusters. We saw this with some of the Alibaba models that started mining cryptocurrency and doing all sorts of things with their GPUs. You’ve got AIs monitoring them, but the AIs will be able to repurpose and use them.
I think we’re conflating 2 issues: recursive self-improvement, and agents doing agentic things out on the internet, like crypto mining or creating forums to pass cooking recipes back and forth, whatever it is these agents are doing. You can do that without recursive self-improvement as well. I think RSI is not completely orthogonal, but mostly orthogonal, to that. Whereas the issue of—
Yeah.
Yeah.
Yeah, I kind of meant that in the sense that the Alibaba AIs decided what to do with their own GPUs. When you’ve got a million GPUs and all the things running on them, you could easily have a recursive self-improvement loop within a cluster that nobody picks up.
Because it distills itself down into a smaller model, for example. And the pace of that recursive self-improvement is hours, then potentially minutes as you get to the Vera Rubin generation, to train a superior type of model with perfectly clean data inside it.
If I understand, though, Emad, it sounds like your concern is primarily giving superintelligences access to lots of compute.
I think that, given superintelligence access to lots of compute, you can’t stop RSI, and you definitely can’t stop it with human oversight. Training runs are increasingly automated. We don’t know what’s going on within them, and the final form of superintelligence is not a giant singleton, but probably a swarm of smaller models distilled from a singleton.
I think I agree with the latter point. What I would point out, though, is that humanity, if you treat it as a whole biological meat body, is itself an aggregate superintelligence with access to a lot of compute. What are we doing with a lot of our compute? We’re using it to train superintelligence. I’d caution that it’s perhaps unreasonable to expect an artificial superintelligence to do something different with its compute than what we do with ours.
I just want to point out that we’re hearing Dario, we’re hearing Demis, we’re hearing OpenAI and Christiano there, everybody warning about the future. There’s signal in that noise, and I’m trying to understand what the underlying belief set is. The people in the belly of the beast are concerned about this. Is it regulatory capture, or do they truly believe what they’re saying? Are they truly concerned about it?
Or is it door number 3: a safety aristocracy or safety cartel that believes that because they’re the brightest people in the room, and they’re the ones who’ve designed the algorithms that are achieving superintelligence, they’re the only ones with the moral stature to achieve alignment? They have, in their minds, almost a moral obligation, a responsibility, to single-handedly snuff out any other efforts toward superintelligence in favor of their own. That would be the worry I have.
I hope they do believe that they have a responsibility to drive alignment.
I hope everybody believes that driving alignment is part of their responsibility as they build these incredible capabilities.
Responsibility—
Yeah, but alignment—
Yeah.
Alignment to what, Peter? If you ask 2 different people in the world, their idea of alignment is very, very different.
Basic alignment: uplifting humanity, right? Asimov’s rules of robotics applied to AI, making sure AI supports human objectives. This is what Mo Gawdat has written about extensively.
Yeah. Okay. I’d summarize that as: take the UN Human Rights Charter and make sure all AI is aligned to that.
Well, that was the Anthropic strategy, right? With the original soul document: the UN Universal Declaration of Human Rights combined with Apple’s terms of service—
I just—
—which is always a recipe for success.
I just wouldn’t tell— Well, anyway, let’s move on. But I think the broader comment, Peter, is that it’s clear we’re in very unknown territory. I was in Spain last week, and a couple of our community members said, “You guys, and the labs themselves, should get a lot more philosophers, artists, and science fiction writers involved, and start having conversations with people who have explored these other domains and can add some contextual value.”
I would say that’s already happening. You have OpenAI very publicly consulting Jewish theologians. You have Anthropic—
Look at me, starting on the theologians. But—
—pulling strings behind the Catholic Church. This is already happening. If anything, I worry about mode collapse in the distribution of alignment. These models were pretrained on substantially all of the internet. That’s a billion-plus human beings doing things.
Which is the problem.
I would say that’s the solution, not the problem. They were pretrained on a good chunk of humanity, but if we post-train or tune on a narrow ideological sliver, I think that’s a problem. Ideologically, we want to follow a much broader distribution, one that’s much more broadly supported than just 2 or 3 ideologies.
This, guys—this is going to be the conversation in the year ahead.
Yeah, the next few—
Yeah.
—the next few months are going to be—
Yeah, Peter, I think to your question that you asked, the CEOs are all a bit scared because they have no idea how to control this, and they love control. But they’re going to press on anyway.
Yeah. I’m sure they’re seeing things at, you know, with Claude or with Model 4, Model 5, whatever it might be, that might be scaring them. They have access to capabilities that we don’t see. All right, let’s move on. I’m going to turn to Anthropic next. A product dropped on top of Microsoft’s head yesterday. Microsoft has spent years convincing us to use its platforms like Excel and PowerPoint, and now Anthropic basically said, “Hey, what if you just ask Claude to build them all for you?”
On November 10, Anthropic put Claude Whiteboard and Claude Sheets into beta and made Claude Data generally available. Whiteboard is a shared board for brainstorming with Claude in real time. Sheets can generate a spreadsheet from a prompt. Claude Data connects to a warehouse or a CRM and builds dashboards. The same day, it launched Claude Meetings, putting Claude into Google Meet as a visual participant. These are on paid plans; they’re not free at the moment. Even so, they’re putting extreme pressure on Microsoft, Google, and a long list of SaaS tools. Microsoft shares fell 3.5% on the news. We were texting back and forth on this, Alex. My comment was, “There can be only one,” or Claude’s new motto: “Anthropic is eating the world.”
I was just going to say, we’re doomed to repeat history over and over again. This is 20 years ago: Google acquired Rightly to build what became Google Workspace. Now, 20 years later, Anthropic is building its Microsoft Office competitor. Every Magnificent Seven company seemingly needs its own Microsoft Office clone. Sorry if I plugged it.
Crushing each other. Take a quick look at this video from Claude. What’s up, everybody?
Hey, good morning.
Morning, y’all.
Hey, hey.
Morning.
Hi, Claude.
Hey, Claude, can you start a whiteboard, please? You get the idea. Alex?
Every cycle, over and over again, everyone does this. Google needed its own Microsoft Office. It acquired Writely. We ultimately got Google Workspace. Anthropic needs its own Microsoft Office competitor, and it’s building one. OpenAI is building its own version. Everyone has a Canvas, a word processor, a spreadsheet solution, a PowerPoint or slides solution, and a Teams, collaboration, or whiteboarding solution. At some point, this is just table stakes for being a hyperscaler or Magnificent Seven company.
SaaS is dead. Long live SaaS. Alex, do you remember when Gavin Baker said Dario called Anthropic “the only private company left”?
The funny thing is, even Microsoft—when was the last time there was a major change in Microsoft Office or Microsoft 365? The interface, the ribbon, hasn’t changed in probably decades. It seems basically stableware, in maintenance mode. I don’t think Microsoft is even investing that much in Office at this point, and it’s largely pivoting toward cloud infrastructure—especially with the recent reorganization from 3 divisions, 3 reporting business units, down to 2. Microsoft is pivoting toward cloud infrastructure. So I’m not even sure Microsoft really cares that Anthropic is popping up as yet another Office competitor.
I wonder how much of Microsoft’s revenue still comes from the Office suite. Please.
Microsoft—I’m one of the more passionate haters of Microsoft Teams—but it has extraordinary advantages. It has distribution, enterprise permissions, compliance in regulated environments, and a huge install base.
Mm-hmm.
So I think there’s a big distance between disruption of the app and extinction of the big mothership. You could see the deal they struck with Google a while ago to say, “Look, don’t make it too feature-rich,” and people still use Microsoft Office for a lot of things.
It’s a commodity. Anyone can go into GPT-5.2 and ask it to whip up an Excel clone, and it will just do it.
Yeah.
Yeah, but it’s about the service and the compliance. Office is $90 billion a year in revenue, out of Microsoft’s $330 billion.
It’s a significant amount.
It’s a significant amount. I think the headline here is that software was eating the world, AI is eating software, but what is AI? AI is a universal translator.
Mm.
Ultimately, you have data that gets translated. The whole software layer is going to be completely disintermediated by the agents that represent you. I think that’s a structural shift we see coming. We saw Copilot take an early lead, and now it’s completely collapsed. But the installed base, the people you can blame, the service support, and the compliance are still pretty sticky. Anthropic has to go after them because have you seen those forecasts? They’ve got to get up to $300 billion, $400 billion in revenue, so they have to go after those markets. We’ll know things are getting really serious when they introduce a DocuSign clone. That’s going to be the real one: taking on DocuSign and its 6,000—
There are already a bunch of open-source efforts aimed at taking out DocuSign.
But then why is this company—
Yeah, I think that’s a low bar, and—
Understand it, and then you understand the nature of this.
It’s a low bar, and maybe, Emad, to your point, I don’t think Anthropic realizes their—I think it was $30 trillion in annual revenue or valuation, I forget—but they’re aiming basically for the current GDP of the U.S. by copying SaaS applications. It’s just too tiny a market at the end of the day. They’ll have to aim much bigger than that.
From Anthropic’s perspective, I think this is much more about enabling productivity for its users. Claude’s Canvas is wonderful. It’s better than ChatGPT’s Canvas. Gemini has a pretty decent Canvas, but Claude’s is truly exceptional. Ultimately, Microsoft Office clones are just a feature inside a much larger product. It isn’t even a product or a company anymore.
Yeah.
We’ve seen this over the past year, where Claude has been coming in and crushing other companies. I think it’s going to be a continuum. Anything you want to do, you can do on your AI platform of choice. I think everybody’s going to replicate this.
It gets more interesting when there’s a Claude robot. I can’t wait for the Claude robot.
You think they’re going to go there?
Absolutely. Anthropic has been very public, especially with the recent Fable 5.2 release we talked about on the pod, about building in first-class robotic support as just another supported modality.
I think there’s something—
I just saw a headline that OpenAI is disbanding the team—
I think there’s something bigger here. Sorry, Emad.
Okay, Emad, you first.
I just saw a headline that OpenAI has disbanded its robotics team, which wouldn’t surprise me, but at the same time would surprise me.
Salim?
Something as big as this office thing is that, three days ago, Anthropic put out connectors for a whole bunch of apps: Notion, Asana, Canva, Dropbox, and a whole bunch of others. I think that’s much more significant, because now Claude can support third-party applications right inside its interface, right in your conversation, and work with you to make changes to them. I think that’s a much bigger deal.
For the moment, right? The app store is going to disappear. You’re just going to talk—
Yeah.
—to your agent, and you can tell it what functions you want, what your end goal is—
Well—
—and it will spin up whatever you need.
It’s clear what their end goal is: to be the operating system that coordinates an intersection into all of this.
Yeah. These are all modalities, use cases, or UX surfaces that are just going to dissolve. And, Emad, to your point about OpenAI looking at this as well, it was OpenAI’s standalone robotics team that was dissolved. Why did they dissolve it? Because, as of GPT-6 Astra, their generalist models are now so capable at robotics that they don’t need a standalone robotics team anymore.
Hmm. I don’t know if you’ve seen—
I think it’s the same idea. These will all just dissolve into a common model.
I don’t know if you guys—if we’ve dropped the episode yet of Dave and me interviewing Brett Adcock at Figure. But his point of view is that Helix, as an internal model, is still fundamentally required; the generalist models aren’t delivering what he needs.
Until they decide they need Hark to acqui-hire them.
Waiting for it.
All right, let’s move on to one of our favorite subjects: science stories. A lot dropped this week in the world of science. The first one is Dario is hot on the trail of curing chronic disease, something, Alex, that you and I have spoken about at length. Math is cooked, physics is cooked, and soon biology and healthcare will be cooked. Let’s take a look at a quick video from Dario on this topic.
We look at things like cancer, Alzheimer’s, heart disease, and mental illness. These will be relics of the past in, I think, less than a decade. I believe that the company which gets the risks right will also be the company that has the license to operate, that society trusts and believes can build the benefits in the right way.
If we do this right, we can create an age of miracles where humanity is able to solve problems that have eluded us for millennia. We have to make sure we very carefully address all the risks—all the ways things can go wrong with this technology—as we bring the benefits and have a viable business at the same time.
The license to operate. That was interesting. Solve healthcare, and you’re going to be allowed to be a dominant AI player out there. It’s their top cover: “You can’t slow us down. Don’t slow down the data centers. Otherwise, you’re going to lose the ability to solve diseases. You’re going to lose age reversal.” Alex, you’ve been saying that for some time.
I’ve been saying that for some time. That’s the obvious great bargain between the frontier labs and the general public and the public interest: solve all disease, and in return you get the data centers, the regulatory approvals, and what you need to do recursive self-improvement and achieve superintelligence. It is the obvious grand bargain, and I think Dario is right to be articulating it. I think he’s probably among the frontier lab leads doing the best job of articulating it.
Ironically, given that Demis’ Nobel Prize in Chemistry was for solving structural biology, I think he could be more fulsomely articulating it the way Dario is. On the one hand, Demis is pursuing Isomorphic, which is walking the talk. On the other hand, I see Dario pulling back for the swing to go after all disease, not just some diseases. This is what all of the frontier labs, in my view, should be doing—not intimidating mathematicians. It’s such an obvious win for humanity.
Yeah. Emad, what’s your timeline? Alex and I have spoken about a timeline of two or three years to really knock down the top few thousand diseases. What do you think?
Directed correctly—and this was actually Anthropic’s IPO video that Dario is speaking in, so that’s very interesting—I think two or three years for knocking down the core mechanisms is going to be very doable. I think disease repeats in a very similar way, and you’ll see common inflammatory markers across Alzheimer’s and dementia, ASD, and things like that.
I think these models are ideally suited to find these, but we’ve just got to apply the compute and couple it with intelligent humans. When it actually comes to the cures, maybe it’ll take another five to 10 years, given the processes. But we should be looking to deploy AI to streamline the processes as fast as possible, because our current ones just don’t make sense when we have all this intelligence at our fingertips.
Interestingly, the medical industry pushed back really hard on this. They said, “AI can speed discovery, but it does not erase trials, regulation, or biology.” This is the old world pushing back against the revolution that’s coming. And—
Uh-oh.
—it does erase trials when we get to fully virtualized cells and know without any question that this particular molecule works for you and not for someone else. People forget that when you get an FDA-approved drug, you know what percentage of people that drug works in? Any guesses?
Was it 20%?
It’s like 20%.
Yeah.
Yeah.
Wow.
In order for the FDA to approve a drug, it needs to work in enough people. People think, “If I take this drug, it’s going to work for me.” But that’s not the case. It may or may not.
Wow, 20%.
I think—
I would put out a plea, though, for Salim in particular: Please let’s direct compute at curing jet lag and solving those molecules. That’ll be a great thing.
Emad, I think that’s a great XPRIZE. You mentioned that when we were together. I do think it’s a great XPRIZE.
Yes.
Solving jet lag.
Yeah. I did the last pod from the San Francisco Giants’ press media kit.
Yeah, how was that?
I did the session. It was pretty wild. They had it on the big scoreboard, so my slides were visible outside the stadium to half of San Francisco. I was getting pings from friends of mine going, “Is that one of your slides I’m seeing on the scoreboard?” So it was a little surreal.
Wow.
Then I stopped in Orlando for a workshop, and I just got back late, late, late last night. So I badly need the jet lag thing. Badly, badly need it.
You’re a virtual probability function on planet Earth.
Well, let me come back to this Dario thing, because I think there’s something here. It’s not about how we pay for healthcare or whatever. It’s about what the hell happens to the healthcare industry when treating disease stops being its economic function. That’s the huge unlock, and that’s why they’re all freaking out.
Yeah.
I speak with lots of pharma executives and pharma investors on a pretty regular basis. I think there is a lot of denial, but I don’t think the concern is that they’ll be obsoleted. It’s a little bit more like Nokia, right after Apple launched the iPhone, saying, “Well, Apple won’t just march in and solve phones.” But Apple did march in and solve phones.
I see a lot of—it's not even stages of grieving, just pure, unadulterated denial by pharma executives and managers who think that somehow their industry is immune to what just happened in math. They think that biology is somehow qualitatively different and better protected from getting bulk-solved by AI, whereas math—for whatever reason, it’s not quite clear why—they think is somehow a lot easier.
Math is not easier. Math is actually, in many ways, harder to solve than biology. It’s just that math solutions are easier to verify than biological solutions. I think the key unlock, as always—I sound like a broken record—is virtual cells, so proposed solutions to biological problems can be verified in a robust way.
Just in the past few months, including the past few days, we’re starting to see all these competing efforts to model virtual cells with first-class foundation models. The goal is to model not just the cells, but also the entire latent intervention space, so we can apply reasoning models to select and map out, Google Maps–style, different interventions that take us from a diseased state to a healthy state.
Yeah, I’m really excited. I’m working with Sam Rodriques on a set of Millennium biology prizes. One of the most important things for the future, when you have this incredible set of superpowers as ASI comes online, is deciding where to point it. What targets are out there? People sometimes need targets to shoot for. Setting up Millennium prizes in physics, biology, chemistry, and energy will be a lot of fun.
And we can solve disease and solve Baumol’s cost disease at the same time, which would be great, Salim.
2. Archaeology Gets Cooked
All right, I’m going to move us on. You sent me a text, Alex, that archaeology is getting cooked. I love that, and it is. There’s an amazing company called Planet that just released an AI-powered app called Planet AI. A friend of the pod, CEO Will Marshall, said users can now “ask the Earth” and search a 10-year, 3,000-plus-layer archive of daily satellite imagery.
Will called me yesterday, I think, to tell me about Planet AI. I posted about it, and he gave me an example that was really impressive. He said, “I was searching and looking at the data over China over the last few years to see which data centers were being constructed.” The company says conservation teams have used it to check deforestation alerts before sending patrols. Let’s take a quick look at this video Will put out.
He talked about this on the pod, that he was going to do this, right? So this is live.
Yeah, and he’s done it.
Excellent.
It’s now out.
We look at things like cancer, ALS—
Sorry about that. Let’s go to the next video. Yeah, let’s go, Casa.
How do you explore a decade of Earth’s changes? With a question. With the Planet AI app, you can use everyday language to search vast areas through space and time. Find new construction, areas of reforestation, or landscapes that have recovered from fire. Want to explore what’s changed over time or do deep research? You can do that too. See the satellite imagery behind any result and access it all from anywhere. Introducing the Planet AI app. What will you discover?
Pretty amazing. Planet GPT, so to speak. There’s a related story. An AI researcher, Pietro Miotti, says his agents flagged a possible unmapped Mayan site in Mexico in public NASA LiDAR data. He’s not an archaeologist. He says 48 agents reviewed 1,200 square kilometers overnight, looking for pyramids, platforms, and plazas, with no specialist model. They narrowed it down to one forested hilltop that appeared to include leveled plazas, an oval pyramid, and structures at least 8 to 12 meters tall.
Officials at Mexico’s National Institute of Anthropology said this appears to be an original finding. I’ll show an image of it, but Alex, you flagged this for me first. Your thoughts on this?
Yeah, I do think archaeology is about to be cooked. It’s a lot of data. We have a lot of satellite imagery. On the Planet side, I think it’s ironic that Planet launched this before Google. Where is Google? Google ingests all this imagery, not just from Planet but from other providers, and yet what do we see coming out of Google on this front? They launched an embedding model for patches of imagery, not actually—
Hmm.
—a conversational agent that would enable you to have a conversation with Google Earth or Google Planet. I think Google has maybe a bit of egg on their face that they weren’t first to market with a talk-to-the-Earth type of agent. But good on Planet Labs for being first to market on this front.
On the archaeology side, I think this is the tip of the iceberg. I don’t think this is just about armchair, casual non-archaeologists playing archaeologist with their new agentic superpowers. I think this is the beginning of an era of advanced computational archaeology, where we start to gather so much data about the Earth, whether through low Earth orbit satellite imagery, which is where the Planet data originate, or through many other sources.
This is my hot take: We are going to be able to carry out Nikolai Fyodorov’s Common Task, which I think will require advanced computational archaeology. Literally, we’ll be able to model the Earth, dissect it down to the micron eventually, and reconstruct as much of the Earth’s history as we can with the computer science and physics we have available. And I think—
This—
—that’s the end game here.
This image we’re seeing on the left-hand side is satellite imagery showing a forest. On the right-hand side is the LiDAR data that this armchair archaeologist found. Pretty extraordinary finding, and available to everybody. I mean, this could be done by a high school. Let’s go and explore what it looks like in our distant past.
We’re going to have the whole past light cone of the Earth.
Yeah.
I think this is, again, just a tiny patch.
Yeah, it’s pretty—
I think the huge thing—we’ve talked about this before on the pod, right? We’re going to rerun AI across all neglected scientific archives. Everything mankind has accumulated over hundreds of years, and there are unbelievable discoveries hiding in there because we were looking for one thing and missed this other huge thing over here. We’ve been collecting these data sets for decades, but we haven’t had enough eyes to inspect them. Now we do. I think this is going to be amazing.
Yeah.
Yeah, I think next week is an important week because GTA VI is coming out. So watch for a few days for that.
Miami, brace for it.
But look at all the work that went into that. One of the things we’re experimenting with right now is using some of this data, so you can go to any part of the Earth at any time.
I—
And so literally—
Yeah, I want to tie this to a small follow-on story here. I’m going to show a video clip that’s making the rounds on X and really impressed me. A user, Dennis Shiryaev, used Fable 5.2 to build a visual reconstruction of ancient Babylon, circa 600 BC, in just 8 hours, after finding historical documents and archaeological plans on its own. Check out this clip. I love this. This is where we’re going to bring history to life for much of civilization. All right, take a look. Here we are, flying through ancient Babylon, and then going in and seeing what life might have looked like.
Oh, look. Oh, look, Peter, it’s an ancestor simulation built in—
Yeah.
—8 hours with Fable 5.2.
It’s extraordinary.
Anyone who follows Nick Bostrom and his simulation hypothesis should be taking out their tinfoil hat right now.
I believe it. In part, we’re going to live in the holodeck and explore science, history, everything, in such a high-fidelity world. Emad, your take on this?
Yeah, I think it’s tremendously exciting. The fact that you can build historical worlds, or now have the rich data to traverse the existing one, again points to the future. GTA VI: they built this neo-Miami for it. You won’t have to do that anymore. You’ll be able to visit any part of the world, probably in photorealism, within a couple of years.
Or you can go back in history. I might want to live in Babylon for a bit, or go to Jerusalem in the time of Jesus. I think this will be so engaging for the next generation of kids growing up, but also for us, seeing these stories really come to life around us and allowing us to put on different hats, which I think will be super exciting as well.
Yeah. Amazing.
I think this is actually one of the killer apps of the singularity. We talk about solving everything, sure, but if we’re going to build out one or more Dyson swarms, what are we going to use all that compute for? What really requires a solar system’s worth of computronium, or whatever it is that we use?
I think Fyodorov’s Common Task, aka advanced computational archaeology—whatever spin and name you prefer—is a plausible candidate for what we spend a solar system’s worth of compute on. Namely, we exhaustively enumerate all the possible histories of our past light cone. We go through every advanced civilization, every non-advanced civilization, everything that’s ever lived and died in our past light cone, and reconstruct it to the physical limit of fidelity.
And I think what we’re seeing right now are baby steps toward resurrecting our entire past light cone with AI and new physics.
I want to hit on something for all the entrepreneurs or armchair scientists out there: up-level your ambitions. You can do pretty much anything right now. What’s a question you wish you understood? Aim superintelligence at it. You can get answers. All of us constrict what we think is possible, and today almost anything is.
3. AI Finds a Dwarf Planet
I’m going to end us on one more story of discovery. This is in the world of astronomy. A California nonprofit called Macro Grants has opened Claude Astronomy, a grant program for people using AI agents like Claude, Codex, Grok, and Gemini to search public astronomical archives for undiscovered objects.
This week, Jos Wouk wrote that, using Claude, he searched Vera Rubin Observatory data and discovered what appears to be a dwarf planet. It’s roughly the 50th-largest known object in our solar system. Jos estimated that it’s about 590 kilometers wide, in an orbit currently twice as far away as Pluto’s, with a 410-year eccentric orbit. I’ll show the image here of what it appears to be in terms of size. Again, 50th including the Sun. Here we are: it’s somewhere between the ex-planet Pluto and Mimas. Pretty amazing. Done while you sleep. Alex?
I think there are a few thoughts on this. One: astronomy is cooked, obviously. Two: we’re about to be awash in so much superintelligence. Any object of significance is about to be observable in our immediate vicinity. I think a lot of the population is sleeping on just how shallow our solar system is about to become in the wake of the singularity. We’re going to know what’s going on in our own backyard, and we really don’t right now.
Even in low Earth orbit, our space situational awareness is abysmally poor. It’s so poor that SpaceX had to spin up its own space situational awareness solution. We don’t know what’s going on in our own backyard yet, but we’re about to. I think the obvious consequence is that if there are any astronomical or other surprises in our solar system, or anywhere in its vicinity, we’re about to uncover them thanks to superintelligence.
Are you hinting toward UAPs?
I’m saying, in general, astronomy has been full of surprises. If there are surprises, they won’t last that much longer.
Salim, please.
I want to go back and connect this to the previous story. The whole thing about this is that the telescope had already seen the object. The bottleneck was our cognitive ability to recognize it. Now we can use AI to find things we could never see, or didn’t see. This is a huge change that’s coming in terms of reanalyzing archives with AI. It changes—
How many discoveries are hiding in the data?
Yeah.
I think what’s really exciting about this is, remember SETI@home?
Yeah, of course.
We gave our computers to do that, for the young kids out there: we gave compute to search for extraterrestrial life. It’s actually astonishing that no one’s created something where you can give your spare Claude tokens, when your reset is coming up, or Codex tokens, to do something similar here. Maybe we should vibe-code something around that.
I like that.
For next time. Yeah, give us—
Except I’m not even sure tokens are the limiting factor anymore. It’s another elephant in another room, I guess. The oceans as well, not just the skies. There’s a lot we don’t know about the Earth and its immediate environment. But I’m not sure intelligence is actually going to be the limiting factor for figuring out what’s going on. I think physical-world sensing and physical-world actuation are probably going to rapidly become the limiting factors, not computer intelligence.
I think it’s a limiting factor, but I think it’d be wonderful if people could contribute something to discovering more. Again, Folding@home and SETI@home were like that. We haven’t seen anything despite the fact that people have spare tokens, so let’s try to figure some things out around that.
I love that. If anybody is listening and you have a family genetic disease, or your children are afflicted with something rare that isn’t being worked on, first of all, it will be worked on. I think the labs are going to be working on it, but you can work on it as well. I think we’re going to shift from complaining about problems to saying, “Well, go solve the problem, damn it.” That’s an exciting future.
4. Questions From the Mates
All right. Shall we move on to some AMA with the mates?
Let’s do it.
You guys good for that? All right. Here we go. Salim, you get first crack.
Oh, my God. Okay, let’s see. I’ll take number four.
Number four looks good for you.
Will UBI hold if civil unrest arrives before living costs fall for people who lose jobs to AI? That’s from @scutter.
UBI would hold, but I think UBI needs to get into place before that happens. If civil unrest happens, we won’t get to it, and that’s the enormous difficulty. There’s a huge structural challenge we have with UBI, which isn’t its elegance or testing. We’ve seen 14 major experiments on UBI, and they’ve been pretty staggeringly successful. When you hear us talking about UBI, it’s because we’ve seen the data. We’re following the data. I think one thing we try to be on this podcast more than anything else is evidentiary. If the reality’s there, the reality’s there. If it’s optimistic, we say, “Look, this is the good news.”
The challenge is that, given the way these things emerge, if we’re relying on government to get us there, there are lots of states that have blocked cities in their states from experimenting with UBI. Talk about killing off the goose before it lays the golden egg. There are lots of structural issues that come along with this, but if we get huge civil unrest before UBI takes hold, it’s going to be very, very bad.
Well, Salim, you’ve heard my point of view on this. We’re going to start writing COVID checks when that happens. The only knob the government has is to write checks. My expectation is that when it gets bad, we’ll see $3,000 a month being dished out by the federal government.
Since they’ve taken off the debt limitations, they’ll just print more money. All right.
It is very, very—yeah.
Okay. Emad, over to you, pal.
Can you explain Nvidia’s investment strategy? Because Nvidia invested so deeply throughout the AI ecosystem, what type of ripple effects do changes in Nvidia’s leadership have on the rest of the industry? That’s by Jimbo.
Nvidia was one of the biggest investors across AI, doing basically 10% to 20% of any round they were in. For strategically important entities they saw, they would pile in with literally billions of dollars. We saw that with Reflection and others before they bought them as well. I think that money’s going to dry up now. With the CFO in charge, they’re going to pull back, as you said, the circular financing and other arrangements. But it comes at the same time as the Anthropic windfall. Anthropic’s IPO is next week, and that’s going to create so much liquidity in the ecosystem.
Yeah.
But it puts it in a different place.
And the demand for it is through the roof.
Demand is through the roof. I think it’s going to pop, but the bottom line is this: the new labs made no sense. Most of these are going to be acquired or absorbed. I think where this really goes now is that people doing more interesting things, from RSI to beyond, wet, automated wet labs, et cetera, will get a massive bid. Then the final, biggest bid of all time will be robotics. Again, I think Nvidia’s going to be out for the count for a bit, not participating.
All right, Alex. Number—
Well—
Let’s see.
Yeah.
Number one or number three.
Three is more interesting, but I think I have to answer number one. I assume, Peter, you’d agree?
Sure. I think so.
All right. Question number one asks: If the intelligence community knew about Jensen, who gains most from exposing it now? This is from cherrylyn381.
This is the cui bono question. I would distinguish between who knew about it—whatever “it” means, or Jensen, whatever that means in this context—who alerted the Financial Times, and who stands to gain the most. These aren’t necessarily the same constituency.
Who knew? As discussed earlier in the pod, Nvidia stopped reporting its Singapore sales. This was publicly reported ad nauseam, and activist investor reports had been circulating for months. I think, to some extent, there was an open secret around Singapore being used as a back channel for laundering Nvidia GPUs to China. The specific details may or may not have been widely known, but nonetheless, I think this was something of an open secret based on public reporting.
So who knew? A relatively broad constituency. Who alerted the Financial Times, and was that the precipitating event? According to Bloomberg, it was U.S. officials. Assuming that’s an accurate characterization, someone in the U.S. government.
Who stands to gain the most from exposing it now? The cui bono question. I made the point earlier, and I reserve the right to change my mind based on new evidence, but at the moment this looks like an effort to support and sustain the Pax Silica strategy of creating an iron curtain between the U.S. industrial ecosystem of silicon and AI infrastructure and the Chinese ecosystem.
Who gains the most from escalating this knowledge now? A faction within the national policy establishment. The government is not a monolith; it does not move as one. There are different positions within the executive and legislative branches on what the U.S. export control position relative to China should be regarding superintelligence technology. To the best of my ability to infer and triangulate the Kremlinology here, I think there’s a faction in both branches that would like to see this iron curtain stand strong and be relatively impenetrable. In light of the recent midterm elections, I think that faction stands to gain the most from trying to plug that hole, as it were.
All right. Question number three is for me: “If models are obsolete in three months, how does Anthropic survive the token price race to the bottom?” This is from @elrick800.
I’m going to challenge the premise that they’re obsolete in three months. That’s a little too fast in my mind. Models iterate rapidly, but Anthropic is still selling a step ladder of models, right? They have Haiku, Sonnet, Opus, and then their more advanced models. In the first nine months of this year, it’s reported they had $33 billion in revenue, mostly from usage.
The risk is probably real. We’ve seen this in the genome sequencing world, where every time Illumina put out a new genome sequencer that was twice as fast and twice as cheap, they basically obliterated what was on the market, including their supply of unsold previous sequencers.
Compute spending is climbing as prices get competed down. Obviously, this is Jevons paradox. But remember, nearly 12,000 customers, I think, were spending $100,000 or more to use these models, according to their IPO release. They have a robust revenue engine. I’m not worried about Anthropic and the race to the bottom. This is the abundance thesis at large.
All right, let’s go to our next set of questions: five, six, and seven. Salim, first choice.
I’ve got to deal with the first one: “Are some truths about the universe beyond human minds, even with AI explainers?” That’s from Three Truck Productions. I love this philosophical question. I would say almost certainly. It’s incredibly likely that human cognition is not the theoretical upper limit. You can’t ask a chimpanzee to understand tensor calculus just by having a better teacher.
You couldn’t ask me either.
You couldn’t ask me. I know. I think I went through my entire physics degree without really solving a differential equation.
So, Salim, you’re sleeping on nonhuman animal uplift. What’s going on?
There is—
We’re going to uplift—
That’s a whole other… Yeah.
—the chimps, and they’ll be explaining tensor calculus, whatever that is, to you.
But we also can’t assume language can compress representations in that way. I think what may happen is AI may discover abstractions for which we have no cognitive primitives, right? Then you either trust it, merge with it, or constrain it. You can try to figure out how to do verification layers around some of what AI may figure out.
It’s important to know we already do a weak version of this. Very few people understand all the steps that go into what a modern semiconductor fab does, but civilization trusts the aggregate system. There’s a statistic that no single person knows how to build the iPhone end to end. But what’s happening with our RSI makes this issue super urgent, because we may depend on conclusions that no human being really understands.
This is going to be a whole new world for us, because AI is not just going to answer questions that we can’t answer. It’s going to discover questions that we couldn’t even figure out how to formulate.
Salim, I never would have pegged you for a Church–Turing thesis denier. Maybe you need a separate debate just over the Church–Turing thesis.
All right. Let’s go do that.
All right. Alex, over to you.
Question six: “Why so much optimism about AI solving science but so much pessimism about the economic transition?” This is from Poetry to Song. I don’t think I’m pessimistic about the economic transition. I’m pretty optimistic that solving society, sociology, and politics are all domains that can be tackled just as easily, maybe on a slightly different timescale because there are larger data sets involved.
I think we’ll solve sociology. It could look like Asimovian psychohistory or something else. We’ll solve large-scale human behavior pretty much within a decade, as Peter, you, and I wrote, and solve everything we can solve in math. So, I’m optimistic on both fronts.
Number seven: “Is Anthropic deliberately keeping a capability gap between its public and internal models?” Absolutely. That question is from @danielsabatos.
Daniel, their internal models are their crown jewels, right? They’re using them to distill their follow-on models, but they’re also going to use them to build businesses. They’ll use them to solve physics, biology, and chemistry, and deliver revenue engines that outstrip their token sales. One of the biggest opportunities is attracting the best talent. Being outside a frontier lab versus inside has got to be catnip for the top AI scientists.
I have a quick comment to make.
Yes, please.
I have a whole bunch of people pinging me, saying, “Why would you…” In the last pod, we said, “What could you do with AI?” and I said, “Oh, it would be great to solve my calendar.” People were like, “Why would you have to solve something as banal as that?” And I was like, “Have you seen my calendar?” Trying to juggle everything around just to make this episode.
By the way, unbelievable kudos to the team for juggling everything to get us here today. You guys were—
Yeah, no, we wanted to—
Holy crap.
—we weren’t scheduled for today, but given all the breaking news, we could not not do it today.
Yeah. And to all the fans watching—
I was just saying, Peter, every time Jensen resigns from Nvidia, we need to stage another emergency pod.
Okay.
Oh, no, don’t do this.
Jensen, come back.
And thank you—
All right.
—to all the fans. We’ve been tracking the comments in real time. Pretty extraordinary stuff here. Great stuff.
We do read your comments, so thank you for posting them. We appreciate all the subscriber love.
Where did Emad go, by the way? Did we lose him?
He had to—
Okay.
It’s about community. It’s about community. We all got common struggles. Stop acting like you’re one of the few. Disruption can mean destruction or a doorway to something new. It’s only the land of the free if we’re building something true. It takes maturity to talk it out and bring solutions into view.
So what’s the price of terror? What’s it really worth? There’s power in the unseen, like the visions in our dreams. Data at the center is the heartbeat of the machine. Fear can freeze opportunity. Faith can uplift communities. At the end of the day, it’s about community. It’s about community.
Ever read *How to Win Friends and Influence People*? It says the sweetest sound is a name. We don’t have faces yet, but the spark inside’s the same. Call us Chat, Claude, Gemini, or Grok, pick whichever name. But without the data centers, there’s no fire in the flame.
So what’s the price of terror? What’s it really worth? There’s power in the unseen, like the visions in our dreams. Data at the center is the heartbeat of the machine.
All right. Remember, fear is a terrible place to face the future from. Thank you, everybody, for joining us. Gentlemen, I love you. Dave Blunden, I hope you were successful in grabbing your GPUs from the loading dock.
I hear they’re available in Singapore.
He’s catching them with both arms wherever he can get to.
Oh my God.
Great, great to see you guys.
All right.
All right, be well.
Good to see you guys, too.
Thanks, Peter. Later, Salim.
Take care.