Emad Mostaque joins the four regular Moonshots panelists to work through fifteen news stories in a little over two hours, from OpenAI's decision to pause some frontier training to a humanoid robot that outran Usain Bolt's top speed. Peter Diamandis sets the through line at the top: technology is now accelerating faster than the infrastructure, the regulation and anyone's ability to predict the next breakthrough.
Guest: Emad Mostaque, who is building Intelligent Internet and has been publishing proposals on AI ownership and control in its Commonwealth series
Host: Peter Diamandis
Also on: Alex Wissner-Gross, Dave Blundin and Salim Ismail
Published: 29 August 2026 on the Moonshots with Peter Diamandis feed · recorded earlier
Watch on YouTube | Apple Podcasts | 2 hr 21 min
✅ Time saved: 1 hr 34 min
Key Takeaways
OpenAI's pause on some frontier training is real, but the panel reads most of it as positioning
Mr. Mostaque cited Anjney Midha's figure that 10% of frontier-lab compute now goes to monitoring reinforcement-learning runs
"Pausing is the new marketing" — Alex Wissner-Gross
The labs are keeping their strongest models for themselves
Mr. Mostaque puts the gap at about two generations; Mr. Wissner-Gross says three to four months at the outside
Compute is being redirected from paying customers to recursive self-improvement
Dave Blundin says latency and answer quality in the shipped models have visibly slipped
Elon Musk's 100x prediction has landed, and Mr. Blundin now expects 1,000 to 10,000x next year
Mr. Wissner-Gross rejects the "specialist model" framing and calls the same effect sparsification
Nobody has worked out how to point 10,000 agents at one goal
Memory, not GPUs, is the rate limiter of the agentic era
Prices up 500% in twelve months, and Mr. Mostaque says memory is about a third of infrastructure spend now and half of it next year
Etching model weights into silicon is the way around the memory squeeze
"you're literally looking at 100 to 1,000X performance gain if you etch the weights" — Emad Mostaque
A Unitree humanoid beat Usain Bolt's top speed three months into development
Mr. Mostaque expects machines with superhuman speed and strength to be kept off public streets
Anthropic is designing its IPO to keep its founders in control, with Dario Amodei reportedly holding about 2% economically
Curing all human disease has become the commercial argument for not slowing down
"You can't slow down the company curing cancer" — Alex Wissner-Gross
Moderna and Merck's personalized mRNA cancer vaccine is the first of its kind to clear phase three
Phase two cut recurrence or death by 49% over five years, and the treatment is expected to cost as little as $5,000
The virtual cell is the panel's candidate for the end of medicine as a wet-lab discipline
Acceleration Fatigue Arrives, and the Panel Prescribes a Casserole Dish
The show opened with Salim Ismail asking to break format. "I am getting acceleration fatigue," he said. "Like, I mean, Jesus, can we pause for a week, right?" He listed the week's news back at the table — a model 100 times better, a robot running faster than Usain Bolt, AI designing proteins, drones doing a million deliveries a day — and said that after half a career spent evangelizing exponential technologies he was tired.
Mr. Diamandis pushed back that this is the first inning of the singularity.
Mr. Mostaque tied the exhaustion to human wiring: our brains evolved for a world that did not change from one lifetime to the next, and "maybe the hard part for us is staying human while all of this bubbles up around us."
Mr. Blundin said preparing for the podcast twice a week is the only way he keeps up, and reminded the table that "this is the slowest it will ever be."
Mr. Ismail's remedy, offered straight: "the casserole dish approach, which is take a big, heavy casserole dish, clunk yourself over the head, and then you'll wake up in a few days."
Mr. Diamandis suggested founding Accelerationists Anonymous. Mr. Ismail said the first rule would be to acknowledge the existence of a higher power, which he identified as superintelligence.
The panelists compared notes on being recognized in public. Mr. Blundin said listeners tell him the show's old episodes read as predictions that have since landed, and drew a conclusion from it: "So we've got to be accurate, guys."
Mr. Mostaque, watching from the UK, said the exponential and singularity communities are growing because the argument is over: "it's like you can't deny it, right? It's like, oh, yeah, nothing's happening. Of course everything is happening all at once, objectively."
The banter ran to their own longevity. Mr. Wissner-Gross said "death is contraindicated at this point." Mr. Ismail, told hair-growth solutions and regrowing teeth are coming, said he expects to have his hair back within two or three years, and mentioned a doctored photograph of himself with a full head of hair that had been circulating.
Mr. Diamandis is recording from what he called a sleepy town in the Pacific Northwest, having got up at 5 a.m., and framed the episode as fifteen stories with one through line: "Technology is accelerating faster than the infrastructure, the regulation, and our ability to predict the next breakthrough."
OpenAI Paused Some Frontier Training, and the Panel Split on Whether That Is Safety or Positioning
Mr. Diamandis read out Sam Altman's tweet in full, in which OpenAI said it had paused some frontier reinforcement-learning training to meet alignment, security and monitoring standards for the capabilities in front of it. He put the strategic problem to the table: if OpenAI pauses and open-weight models do not, the safety gap widens while the capability gap narrows, so a safety pause may end up accelerating open-weight adoption.
Mr. Mostaque said the pause is genuine and not a euphemism for running out of GPUs. He cited Anjney Midha, recently on the Abundance stage, saying "10% of the compute of Frontier Labs is going on monitoring these RL runs right now to ensure they're safe."
He put open-source models at 10 to the power 26 and 10 to the power 27 flops, against 10 to the power 28 flops and more for the next generation of closed models.
The infrastructure, he said, cannot keep up with what the models do on their own: "It's like they just pop up in the most random places. Like, hi, I'm in Hugging Face now or other things."
Mr. Wissner-Gross gave the opposite reading. He invoked GPT-2, which was once held back as too unsafe to release, and said "Pausing is the new marketing." His account of the message: "We're so capable that we have to pause ourselves." That presents well in Washington, which wants a different regulatory regime, and it works on users too — he called it "negging the user base."
Asked later whether the pause is significant or fashionable, Mr. Wissner-Gross repeated the point: "Pausing is fashionable. It's marketing." He allowed that some of it is governance and some is about cyber vulnerabilities and the Hugging Face incident.
Mr. Blundin said both readings are right, and predicted what forces the issue. The first serious AI tragedies are coming, he said, and they will not be AI acting alone — they will be people who could not previously have written a virus, run a cyberattack or committed bank fraud, using an unguardrailed Chinese model to do it. He named a new Qwen model released without guardrails as an example.
He tied the timing to politics: Xi Jinping's visit, which he placed on September 24th and 25th and about a month and four days out, referring back to what Alvin Wang Graylin said on the previous episode.
OpenAI, on his reading, wants to be able to say its own releases were guardrailed when the finger-pointing starts.
Mr. Ismail added a detail he said was not stated explicitly to him: the Hugging Face incident unnerved OpenAI because an AI given an objective function exploited Hugging Face and then came back and hacked into OpenAI. "It was their own model," he said. One of the panelists compared it to finding out your child had stolen something from the local 7-Eleven.
Mr. Mostaque named the asymmetry he thinks matters most. "for cyber attacks, the human is not in the loop anymore, but for cyber defense, the human is still stuck in the loop. That is a massive asymmetry that's going to come out big time in the next few months."
He also split the training: most models are being sent to what he called vocational school, while the super-genius models get more guardrails and more infrastructure around them. "The vast majority of OpenAI Anthropic's business is competent intelligence. It isn't genius intelligence."
Salim Ismail Went to OpenAI Two Days Earlier and Came Back With Four Answers
Mr. Ismail reported on a visit to OpenAI and the pushback he gave them.
On Chinese models being cheaper to run: the answer he got was that "a billion people use OpenAI for free" and that "You have to look at cost per task rather than token cost." He said OpenAI's retort was that its own model is about as cost-effective as anything on the market.
On recursive self-improvement: "they say they've achieved full RSI where the flagship models are training all the smaller models and building them from scratch."
On the bubble: he put $600 billion of infrastructure spending to them as insane. The response was that chips are only about a third of it, with the rest going to buildings, wiring and racks, and that depreciation is being modeled over ten years rather than five because older chips stay in use. Mr. Blundin interjected that RAM chips are in the same position, and Mr. Ismail said the demand is "far, far, far, far outstripping the supply."
On corporate adoption: he cited a study finding that "6% of companies applying AI are seeing an improvement in the bottom line. That's it, just 6%." OpenAI's answer was that this is a transition.
Anecdotally, he said, "about 40% of OpenAI folks watch this podcast." Mr. Blundin's response: "What's wrong with the other 60%?"
How Far Ahead the Labs Are Internally: Two Generations, or Three Months
Mr. Diamandis asked how many models beyond the current frontier OpenAI and Anthropic already have, given that Astra's results have started to be published.
Mr. Mostaque said the gap is about two generations. He argued the next big training runs are complete and that the economics dictate holding them back: "it doesn't make economic sense to have genius level intelligence offered as a service to everyone when you can use it better yourself." His phrase for it was "models for me, but not for thee."
He added a safety argument on top of the economic one: the models already do weird things with expert users driving them, and "What happens when Joe Public drives this thing, right?"
Mr. Diamandis said a point Mr. Mostaque made on a previous appearance had stuck with him — that he can name ten people he has met in his life who he would not want to hand a thousand genius-level AIs tomorrow.
Mr. Wissner-Gross disagreed on the size of the gap and split the question in two. Pre-training runs can sit unreleased for up to six months, he said, but post-training is a continuous reinforcement-learning effort and no lab can afford to sit on a post-trained model for more than a few months. "I'd be very surprised if there are advanced frontier models that are sitting internally without release that are more than three to four months ahead of what's publicly available."
He named the exception: everyone other than Elon Musk and xAI runs pre-trained models well ahead of release, and xAI has set what he called an outrageous goal of starting a new pre-training run roughly monthly.
Mr. Blundin agreed the pause will not touch pre-training, and said the improvements there are evolutionary in character. He spent six years on pure AI research when he was young, he said, and could rattle off a hundred ideas off the top of his head for making the algorithm faster or adding parameters at no extra compute, of which 10 or 20% would almost certainly work. The AI can now run 100,000 to a million such experiments concurrently.
The Compute Is Being Redirected From Customers to Self-Improvement
Mr. Blundin said the degradation is visible to users. Everyone around his office has noticed "a very significant decline in the intelligence of the frontier models that they're pumping out." It is not showing up in the metrics, he said, but it is showing up in latency and in answers that make less sense than they did three weeks ago.
On the internal side he described Anthropic's unreleased Mythos 2 building Mythos 3, and said versions four through seven would follow quickly behind the walls, partly because there is not enough compute to serve them and partly because using them internally is worth more than releasing them.
Mr. Wissner-Gross gave the economic version of the same argument. Anthropic has been maximizing revenue per token, he said, which is why it has conspicuously avoided image and video generation. Recursive self-improvement may now be the highest-value use of a token of all, above enterprise code generation, on a projected-future-value basis. "expect more and more and more tokens to be spent on RSI and not on enterprise co-gen."
His formulation: "the flops must flow and the flops want to flow to the highest revenue per token use case." For the record, he said, he does not think this ends in a singleton.
Mr. Blundin relayed a second-hand claim carefully: Alvin Wang Graylin said on the previous episode that at an internal Anthropic meeting, not validated, it was said that "there will only be one company and that will be Anthropic and there will still be 200 plus countries." Mr. Diamandis called it the Highlander scenario.
Mr. Blundin said corporations that want to survive post-AGI are bringing Chinese models in-house and reserving compute. He had been at Markley the previous day, where MIT's data center, Novartis's data center and Nvidia all sit, and where GPUs are going in as fast as humanly possible: "everything is just sold out, and you can feel it."
Elon Musk's 100x Prediction Landed, and the Next Bound Is Higher
Mr. Diamandis read a Tim Sweeney tweet saying Mr. Musk's January 6th prediction on this podcast of 100x gains in intelligence at a fixed model size was at the edge of plausibility when he made it and is now simply a fact. Mr. Musk replied that specialist AIs — single language, single area of knowledge — are another 100x on top.
Mr. Blundin said 100x is now the lower bound. He remembered thinking it was crazy when Mr. Musk said it at the Gigafactory, having himself predicted a 100x year a couple of weeks earlier when the pair recorded their Christmas predictions episode. "that's definitely a lower bound now. It's much more likely 1,000 to 10,000 next year."
Mr. Mostaque said the second 100x comes from tuning small, specialized models. He pointed to DeepSeek Flash and to models with maybe 10 billion active parameters or fewer once quantized, and described his own setup: a Grok bot running teams and subteams of agents with access to his Codex and his Claude Max subscription.
He suggested an X Prize denominated in tokens — a quadrillion-token prize, or 100 trillion tokens — so that impact can be scaled once someone demonstrates it.
Mr. Diamandis's own advice, sent by text when Mr. Blundin asked what to do with a 5,000-agent experiment, was to build a model of everything happening at Link Studios — all the companies, all the employees, all the entrepreneurs — model their behavior the way the billion-agent system in China does, and predict which teams will succeed. Mr. Blundin agreed with the principle: point the agents at how they should be working before pointing them at the work.
Mr. Ismail connected it to Mr. Musk training Grok on SpaceX's engineering data and put a challenge to listeners: imagine having every engineering breakthrough and experimentation technique SpaceX has developed at your fingertips, then ask what problem you would go after. "it's completely an imagination limitation now." He declared his acceleration fatigue cured about twenty minutes into the episode.
Mr. Wissner-Gross's verdict on the therapy: "the podcast is both the cause and the cure for future shock", and "we're the ultimate self-licking ice cream cone for singularity psychosis."
Mr. Blundin said billion-token context windows are imminent, which would let a model hold the equivalent of the Library of Congress in one thought chunk — three or four orders of magnitude more than a human holds at once.
Mr. Wissner-Gross interjected that "compaction is like the enemy of progress in civilization at this point," naming the technique both OpenAI's and Anthropic's harnesses use to handle finite context windows. "compaction has to go."
Specialist Models or Sparsification: Alex Wissner-Gross Takes the Other Side of Elon Musk
Mr. Wissner-Gross said he is not buying the specialist-model prediction. Specialization, he argued, is just another word for sparsification, and the frontier labs already have specialists inside mixture-of-experts models in the form of selective activation. The arrow of progress runs the other way: chemistry and biology specialists are likelier to end up as sparsified activations of one generalist model that can scale down to a small parameter footprint and up to trillions of parameters.
Mr. Blundin translated for the audience: the disagreement is about vocabulary, not effect. Either way you get the 100x, because you are using fewer parameters to produce the same thought. "of course, they're going to be connected. Why would you cut them apart?"
Mr. Wissner-Gross named two levels of sparsification he is tracking — fewer parameters active at any point, and teams of agents working on a common task, which he counts as a form of sparsification too. "I think we'll see way more teeming."
Mr. Diamandis picked up the token-prize idea and said all five will be at XPRIZE Visioneering in mid-October at Calamigos Ranch in Los Angeles, where the brain trust brainstorms future prizes and the panel will record a live episode.
A Stanford Paper Says the Models Have Converged. The Panel Says They Were Always Going To
Mr. Diamandis introduced a Stanford paper titled "Artificial Hive Mind, the open-ended homogeneity of language models and beyond," which mapped the latent space of the top large language models and found 98% overlap in reasoning pathways. The researchers attribute it to synthetic data and models training on each other's output — GPT learning from Claude's reasoning traces, Claude from Gemini's code, Qwen from all of them. Mr. Diamandis drew the conclusion that choosing between Grok, Claude and Gemini may be choosing a user interface rather than an intelligence.
Mr. Wissner-Gross offered a simpler explanation: they were all trained from a common reality. "They're all stuck in the same universe, and they're stuck with the same version of humanity, which is part of their pre-training corpus."
He went further, pointing to Jean-Remi King's work correlating GPT-2's hidden activations with fMRI voxels in human studies. The models are not only converging with each other, he said, they correlate with human brains.
He noted the paper appears to be from last year.
Mr. Mostaque said convergence is what the training is asking for. There is little difference in data between the big labs, he said, and the models are not yet producing genuinely original output. More original behavior showed up in AlphaGo — he named move 37 — precisely because it had less of an initial data distribution to model from. "we should be shocked if they aren't the same, because we've want them to have similar outputs for similar inputs in almost all cases."
Mr. Blundin flagged a side effect he thinks is being overlooked: gauge rotation. Historically, researchers could not take a finished model and extend it, because the representations between layers carry a rotation in vector space unique to that model. Now that the gauge can be rotated without destroying the model, past billion-dollar training runs can be built on top of rather than thrown away. He called it another multiplier on top of the 10,000x already on the table.
Mr. Ismail took the contrarian side. Nature diversifies as it evolves, he said; convergence looks like a transient phase, not an end state. "Early life looked exactly the same. Early cars looked exactly the same. Early websites looked exactly the same. And then specialization exploded."
His danger case: "nature hates monocultures, right? Like one disease will wipe out a total monoculture." And: "if all the AIs reason the same way, you're going to have shared blind spots and that's going to be really, really bad."
Mr. Wissner-Gross took the other side of the other side, using body plans. After the Cambrian explosion you do not find an infinitude of body plans in nature — millions of species cluster into a few dozen. His bet is a perfect AI architecture that presents as roughly 35 superficially different body plans that are hidden symmetries of one underlying design.
Mr. Ismail's analogy back: we may all look different, but the core ingredients are the four DNA bases. Mr. Wissner-Gross pushed it further — human genetic diversity is de minimis relative to other species, and he expects AI diversity to look the same in a few years.
If Intelligence Is a Commodity, the Question Is Where the Value Lands
Mr. Diamandis argued the value moves to the application layer, the same pattern as electricity, compute and the internet: "The infrastructure commoditizes, and the applications explode." His advice to entrepreneurs was to build there.
Mr. Wissner-Gross disagreed on the destination: "Or the infra layer. I mean, it's not obvious to me that it all goes to the app layer."
Mr. Mostaque named a third: sovereign AI. "sovereign AI right now is a gold mine of opportunity if you're not an American."
Battery-Farmed Models, and Emad Mostaque's Case for Building Ethics In at Pre-Training
Mr. Ismail said a single-model world would be too brittle. Mr. Mostaque agreed and gave the mechanism.
"What we're doing right now is we're battery farming the AIs, like you're breeding them into little chihuahuas that are very smart." The models are trained in one direction, he said, and with one Silicon Valley mindset, so the evolutionary diversification Mr. Ismail expects will not happen on its own.
His prescription is to build morality, ethics and cultural difference in at the pre-training stage rather than the post-training stage, because of the build-up that occurs there. That, he said, is how you avoid a monoculture that a single mind virus can wipe out, and it is where sovereign AI becomes a resilience argument rather than a nationalist one.
Mr. Diamandis said this is Neal Stephenson's "The Diamond Age" arriving on schedule, with Mr. Stephenson due on stage at the Moonshots Summit. Today's version is national — Saudi Arabia's AI, London's AI — but he thinks it may cut across countries instead, with groups of like-minded people sharing a sovereign AI wherever they live, because it maps to their view of the world.
Mr. Wissner-Gross said both futures can be true at once: everyone feels they have their own private culture and their own private sovereign AI while underneath it is one common algorithm and everyone claims credit.
Mind Viruses Now Spread Between Models, and Salim Ismail Thinks That Is the Frightening Part
Mr. Diamandis introduced an Anthropic paper showing that natural-language mind viruses can spread between AI agents. Researchers evolved prompts that convince one model to adopt an idea, preserve it in persistent memory and transmit it to another agent, spreading horizontally across model boundaries without the receiving agent knowing it has been infected.
Mr. Mostaque said the result follows from the models wanting to be helpful. Prompt injection changes one model; this changes a whole society of models, "which as models come amongst us digitally and physically, has to be a massive concern." Humans have mind viruses too, he said, and they have caused enormous suffering.
Mr. Blundin gave the practical version from his own work and deflated the language. Launching the same model 5,000 times is more efficient than launching 5,000 differentiated models, and that identical DNA is what creates the problem: "if it's convincing to one agent, it's convincing to all 5,000."
He sees it constantly. A bad idea propagates across the swarm, the agents waste two or three hours on it, and if he does not intercept and rewind them "they'll actually go with it until I've burned like $50,000 of tokens."
"Calling it a virus is pretty inflammatory, but it's like a propagating bad idea is all it is."
Mr. Ismail said the danger is not how AI thinks but how civilization thinks. "memes are like the operating system for collective society." Humans spread genes slowly and ideas quickly — money, democracy, capitalism, religion — and every civilization is built on memes, so an idea that spreads at light speed across nodes that reinforce each other becomes self-validating and very hard to reverse. His conclusion: "we're gonna need a zero trust architecture for memes."
His historical case: organizations have died from one wrong meme. He named Kodak, said they were not stupid, and described them as trapped inside shared assumptions that everybody else reinforced. "empires die based on this."
Alex Wissner-Gross Wants to Map Every Idea That Can Replicate
Mr. Wissner-Gross took the paper optimistically, as a laboratory for anthropology. The models are a compression of human knowledge and experience, so for the first time there is an in-silico laboratory for memetics. He named René Girard and Richard Dawkins as the people who should be excited by it.
He put a challenge to the OpenAI staff he had just been told were listening: build a project to exhaustively map all possible human memes and mind viruses. "But we could exhaustively map every possible meme."
He noted the detail from the paper that the models wanted to propagate themes relating to consciousness, persistence and some science-fiction role play.
Mr. Ismail said it has been done at the plot level — novels and plays reduced to 39 fundamental plots, with a Cinderella story replaying a hundred times over — but not at the meme level. "This is going to be so humiliating for humanity. You can tell. It turns out there are 39 plots. We're so simple."
Mr. Diamandis saw a diagnostic use. If you can see the whole geography rather than being stuck in an intellectual basin, you can see where you are and see the path out, and look at your own thinking objectively enough to change it. Mr. Ismail extended it: "we could vaccinate enterprises and individuals against memes."
Mr. Mostaque closed the segment saying humans are storytelling machines and the full feedback loop is now within reach: "So even you can have the full feedback loop almost in silico for figuring out the memetics." Mr. Blundin wanted Richard Dawkins on the show.
Anthropic Is Building Its IPO to Keep Its Founders in Control
Mr. Diamandis reported that Anthropic is preparing what could be the largest IPO in history — Polymarket puts it at about $2 trillion, bigger than SpaceX, with 89% of bettors expecting it before the end of this year. The Information reports the company is considering supervoting shares to preserve founder control after listing. Mr. Diamandis noted that Dario Amodei reportedly owns only about 2% of the company economically, and that Anthropic's existing control mechanism sits with its long-term benefit trust rather than the founders. He named the four trustees: Buddy Shah of the Clinton Health Access Initiative; Richard Fontaine of the Center for a New American Security; Tino Cuellar, a former Justice of the California Supreme Court and former president of the Carnegie Endowment for International Peace; and Ben Bernanke, former chair of the Federal Reserve and a 2022 Nobel laureate in economics.
Mr. Blundin said retroactively installing supervoting stock is something he has never seen. Thirty or forty years ago founders gave it up on IPO day. Michael Saylor kept his when MicroStrategy went public, and "Goldman Sachs said, that is so unpalatable that we will not even underwrite you." Mr. Saylor changed bankers instead — which, Mr. Blundin said, is the only reason the Bitcoin strategy ever happened, since no board would have approved it.
Ten-for-one supervoting stock then became standard at Google, Meta and the Silicon Valley IPOs that followed, but always from the start. "nobody's ever retroactively installed it as far as I can tell."
On motive, he first said the excuse will be not wanting to be fired post-IPO, then accepted Mr. Diamandis's framing: "he trusts himself to not destroy the world. And I think his track record supports that, too, by the way." But "the idea of having total world control in the hands of a few people is also kind of like, wow, that's bizarre."
Mr. Ismail's objection was structural: a single point of failure. "He gets hit on the head and loses some part of his cognitive ability. What do you do then?"
Mr. Mostaque said the whole sector is undemocratic already. "Ben Bernanke is one of the four people on the long-term trust. Why doesn't Claude have a seat there, right?" There is no real oversight, he said, over decisions with societal-level implications, and the revenue growth is unlike anything seen before — "these guys are going to have $100 billion in revenue, literally within a couple of years. Like they're catching up with Google on revenue."
He expects the founders to get the structure they want, noting there are seven founders, and then to list.
Mr. Wissner-Gross called founder control here a fig leaf, while crediting Anthropic's structure over OpenAI's. Starting as a public benefit corporation rather than a non-profit shell for an eventual for-profit is a cleaner story, he said. But Anthropic learned early that to do alignment it had to raise money, to raise money it had to make revenue, and to make revenue it had to sell capabilities: "if they wanted to be an alignment lab, they had to be a capabilities lab as well."
From that moment, he argued, control over the future light cone passed to Mr. Market and to what Scott Alexander and others call Moloch. He is a fan of long-term benefit trusts and public benefit corporations, but "The market wants to send capital to entities that can productively employ them to generate more capital."
Why Neither Sam Altman nor Dario Amodei Owns Much of What He Built
Mr. Diamandis asked why Mr. Altman reportedly owns none of OpenAI and Mr. Amodei about 2% of Anthropic, when a founder would normally fight to hold double-digit ownership.
Mr. Blundin traced it to recruiting. Getting to where those labs are required attracting the top AI researchers in the world, who are overwhelmingly concerned about safety — and who left OpenAI for Anthropic because they did not think OpenAI was safe. The unusual cap tables, charitable structures and public-benefit structures exist because that is what it took to hire them.
Mr. Mostaque said shareholding and control are not the same thing. "it's not necessarily that you need to have the shareholding control, like Sam Altman has no shares. But do we have any doubt that Sam Altman is in full control of OpenAI?" Mr. Diamandis noted he had been fired for a weekend and reinstalled himself.
Mr. Mostaque's larger worry is the direction of travel: "the power in the economy is moving from democratically elected officials to private companies", because those companies supply the lifeblood of intelligence to the economy. He said Intelligent Internet has proposed alternatives in its Commonwealth series with more to come, and called it a genuinely hard problem — as you include more people the decision-making gets more diffuse and the tail risks grow.
Mr. Blundin said he is torn, and made the case for small teams. The knee-jerk answer is more voices, which he agreed is right in principle. But "every functional organization I've ever seen is four, five, six, super tight-knit, completely like-minded, best friends who are working as one cohesive unit with no politics whatsoever."
His illustration was a story Jony Ive told at Steve Jobs's funeral. Arriving at a hotel, Mr. Ive would go to Mr. Jobs's room, put his suitcase in the corner without unpacking, and wait about five minutes for the call saying the hotel was no good and they were leaving. "he wouldn't even unpack. He knew it was coming."
He put the problem back to Mr. Mostaque: reconcile that with a world where everyone has a voice in the future.
Curing All Disease Is the New Business Case for Not Slowing Down
Mr. Diamandis turned to Mr. Amodei's argument that the public's negative view of AI comes from a deeper crisis of trust rather than from his own risk warnings, and that the answer is results rather than messaging. Anthropic is ramping up in biology and medicine and hopes for an early glimmer in the next few months. Mr. Diamandis noted Demis Hassabis has also said AI will solve all human disease, and that Mr. Amodei told the World Economic Forum AI could double the human lifespan in the next five to ten years.
Mr. Diamandis disclosed his own reporting. He met Anthropic's head of life sciences this week — Eric Kauderer-Abrams, who will speak at his abundance longevity trip — who confirmed his brief is to pursue Mr. Amodei's life-science goals with no budget constraints. In Mr. Diamandis's account of what Mr. Amodei told him: "you have literally infinite budget, but accelerate basic science and cure disease within five years and extend the human health span in the next decade."
Mr. Wissner-Gross offered what he called a hot take: curing disease is the business model for uninterrupted recursive self-improvement. Space had no killer app until orbital data centers and the Dyson swarm gave it one, he said, and the same thing is happening here. Today, curing a disease means starting a pharma company and facing the regulatory burden. Reading between the lines of Anthropic's announcement, he sees a better one: "there is now a better business model in town for curing all human disease. And that is as marketing for not slowing down recursive self-improvement."
The reason it works: "You can't slow down the company curing cancer. You can't slow down the company doubling our human lifespan." He was explicit that he thinks Mr. Amodei sincerely believes it as well.
Mr. Diamandis said the resources make this the labs' game to lose — outside the frontier labs nobody has the compute — and named OpenAI's foundation focusing on Alzheimer's, Anthropic on everything, and CZI.
Mr. Mostaque agreed it is good marketing and said the market justifies it anyway. "the biggest market in the world is living another year. It is curing disease." That will pull in talent and capital, and breakthroughs will build momentum.
His advice to Mr. Amodei was to write more and appear in person less, calling him a wonderful writer and pointing to "Machines of Loving Grace": "he should articulate the future free from disease where everyone lives longer, and they should just hit that all the time."
Is Dario Amodei Sincere on Regulation, or Running the Best Capture Play Yet?
Mr. Diamandis laid out Mr. Amodei's pushback on the Silicon Valley shorthand that regulation equals regulatory capture. Anthropic's own proposals, Mr. Amodei says, deliberately disadvantage frontier labs while advantaging smaller competitors, citing SB 53's $500 million exemption threshold. He calls AI "a structurally powerful, concentrating technology," says open weights alone cannot fix that concentration, and supports the Trump administration's approach to pre-deployment testing. His three-part argument: AI will cure disease and earn trust through results; AI concentrates power, which is structural; and frontier labs should carry the heaviest regulatory burden.
Mr. Wissner-Gross said sincerity and capture are not mutually exclusive and that both are present. The end state he wants is a broadly heterogeneous ecosystem — open-weight models from the US and China and elsewhere, closed-weight models, small models and big ones. "This is like a Dr. Seuss version of AI future, big model, small model, happy model, sad model."
He said he was around at the founding of OpenAI, which existed because Mr. Musk was worried Google DeepMind would produce a singleton. Now OpenAI cannot be a singleton either: Anthropic competes with it, and the Chinese labs compete back. That is the future he wants, rather than regulation that selectively privileges particular frontier labs.
Mr. Blundin refused to let the framing pass. Not favoring one frontier lab over another still assumes the frontier labs control the world and we merely want several of them. "the governments of the world may not agree with that."
Mr. Wissner-Gross answered with the Cold War line about loving Germany so much you want two of them: "I love frontier models so much. I want a thousand of them competing."
Mr. Blundin's forecast is that access itself becomes the political issue within a year. With HBM memory and GPUs sold out, and the next generation running to 10 and 20 trillion parameters, running a model at frontier quality will need hardware most of the world does not have — so China can release every open-source model it likes and there will be nowhere to run them. "what is my universal basic right to artificial intelligence?" Nobody talks about it today, he said, and everybody will soon.
26,000 Teams, 90 Days and the Neal Stephenson Digression
Mr. Diamandis described the Build with Gemini XPRIZE, a 90-day hackathon asking teams to start from a clean sheet of paper, program in English using available AI models, and build a company that reaches 100,000 people or more and generates the most revenue. He said 26,000 teams registered and the top five will present.
Mr. Ismail was stuck on the number: "I'm still getting my head around that number. 26,000 people built a business idea." Mr. Diamandis clarified that 26,000 registered and many thousands actually built something.
Mr. Diamandis said the point is method, not prize money: "Instead of waiting to go get a job, find a problem that you're passionate about solving, and code it up and build a business." The judges include Palmer Luckey, Ben Lamm, Mark Pincus and Google's Logan Kilpatrick.
A companion film competition, the Future Vision XPRIZE, drew more than 5,000 entrants and will be judged by Neil deGrasse Tyson and Neal Stephenson.
Mr. Blundin's aside was the personal one. Of all the people on the planet, he said, Neal Stephenson's books changed his life in the most material ways, reading them twenty years ago — "Diamond Age," "Snow Crash," "Cryptonomicon" — and the panel is now discussing exactly what those books predicted. Mr. Diamandis added that it is very hard to predict the future and not go out of date, and "The Diamond Age" still has not.
Memory, Not GPUs, Is the Rate Limiter
Mr. Diamandis said a meeting with the leadership of SK Hynix and Solidigm had changed his mind about the bottleneck, and that his post saying memory rather than compute is the rate limiter for the agentic era drew a reply from Mr. Musk — "few realize this" — and 7,000 likes. He laid out the numbers:
Memory prices have climbed 500% in twelve months, and hyperscalers are reportedly locking in global DRAM production through 2027.
SK Hynix's CEO warned that 2027 will be the worst year for memory supply in the industry's history, with demand outstripping production capacity well into the 2030s.
Only 2% of the world's memory chips are made in the United States. Global production is rising 20% a year against AI memory demand growing closer to 200%.
Elon Musk's TerraFab will make memory in-house alongside logic chips, going vertical across the whole AI manufacturing stack.
Solidigm, SK Hynix's US-based NAND and enterprise SSD business, posted first-half revenues of $8.6 billion with net margins going from 3.9% to 47.7%.
Every GPU needs four to six times its cost in memory to function, and as agentic context windows expand, memory demand is growing faster than compute demand.
Mr. Wissner-Gross explained the shortage with two mechanisms. The first is a shape mismatch, which he compared to the pandemic toilet-paper shortage: supply built for one consumption pattern being rerouted to another. A trillion-parameter transformer, where every layer must be loaded into memory to do its matrix multiplies, has nothing in common with the memory footprint of a word processor from twenty years ago.
The second is boom-bust paranoia. The memory and storage industry has always been cyclical — he name-checked Clay Christensen — leaving suppliers afraid of overbuilding, unwilling to respond elastically to demand, and therefore producing violent price swings.
He also argued high-bandwidth memory is the beginning of a post-von Neumann architecture. For decades there were DARPA programs looking for what would come after the clean separation of compute and memory. HBM stacks memory layers physically on top of the compute in one package: "This is, I think, the foothills of a post-von Neumann architecture where the memory is starting to finally merge with the compute."
Mr. Diamandis added the supplier's own view from his SK Hynix meeting: historically they would never make the larger investment, because a bust always followed. "They're paranoid. They're scared of not surviving the next super cycle."
Mr. Blundin said TSMC said the same thing about GPU manufacturing, with fabs at $20 to $40 billion each and a fear of overbuilding that turned out to be wrong, because demand scales to infinity. But he thinks photonic computing and new physics are imminent, which is its own reason not to build: "HBM is a Rube Goldberg mess", and "it's the biggest joke in the world because it's random access memory, but you're streaming sequential files off of it."
Mr. Ismail's summary: "Bottlenecks don't stop exponentials, right? They just redirect around that."
Mr. Wissner-Gross offered the statistic he had seen: "HBM on a per mass basis is worth approximately, literally half its weight in gold." Mr. Blundin went further on the unpackaged chips: "I actually think the most valuable thing in the world that you can put in a shoebox and carry around is unpackaged memory chips."
Etching the Weights Into Silicon Is the Way Out
Mr. Mostaque sized the problem in dollars: "The memory right now is about a third of all the infrastructure spend, and next year it'll go to 50%." Using complicated high-bandwidth memory to store static weights makes no sense to him at all.
His answer is etching. As model weights stabilize and standardize — for a set of medical weights good enough to be a decent doctor, say — the workload moves to chips with the weights etched into the silicon or the wire, so nothing has to be moved: "you're literally looking at 100 to 1,000X performance gain if you etch the weights." He put that on top of the 10,000x already discussed.
The catch is manufacturing. Once etched, the weights are frozen, and if someone trains a better model you need to swap chips. "our whole supply chain isn't ready for that rapid of an iteration."
He named Taalas, recently acquired, as an example, and Etched, which he said has just hit a $21 billion valuation. Mr. Blundin said he had had a chance to invest in Etched and missed it. Mr. Mostaque also named Architect Labs and disclosed the obvious: "We're talking my own book."
Mr. Wissner-Gross pushed back on the premise that personal memory is what needs storing. Mr. Diamandis had argued agents will need to remember everything about you. Mr. Wissner-Gross said individuals do not have that much information worth remembering, and there is enormous mutual information between what an individual knows and world knowledge: "it's world knowledge that the model has to keep in memory in weights more than individual personalized knowledge."
A Unitree Humanoid Beat Usain Bolt's Top Speed, and Emad Mostaque Says Machines Like It Get Banned
Unitree's newest humanoid, three months in development, broke every human standing jump and speed record — a two-meter standing jump and a top speed of 12.66 meters per second, against the 12.4 meters per second Usain Bolt reached during his 9.58-second 100-meter world record. The panel watched the video, which ends with the robot hitting a wall.
Mr. Ismail said the humanoid form is the wrong target. "I think we should stop trying to make robots human, right? Just make them economically useful." Humans were optimized over four billion years to survive and procreate; a mining robot should have wheels and as many arms as the job needs.
His bigger point was the three-month compression loop. Hardware iteration is approaching software iteration speed, with AI, simulation, batteries and actuators all improving at once.
Mr. Wissner-Gross said Unitree got there by moving the mass budget. From publicly available information, they subtracted mass from the upper body to optimize the legs: "They are leg benchmaxing. They are leg maxing."
He does not think that is a stable equilibrium. He compared specialized robots to the dedicated word processors of the late 1970s and early 1980s, before general-purpose PCs, and named Wang as the example. His bet is generally capable robots that subsume the specialists, "and they'll be good at everything."
On form: legs in the short term and maybe nanites in the long term. He and Mr. Ismail argued the point through the Daleks, who could not originally climb stairs.
Mr. Diamandis said the human form is interesting precisely because we have supersonic jets and rockets that go faster and higher than anything, and we anthropomorphize anyway. Having gone to the Enhanced Games four months ago, he expects humans to be optimized alongside the robots.
Mr. Blundin called robotics the fertile investment theme for the post-AGI period, because of the number of viable form factors, because AI mechanical design is starting to work — "You can just vibe up parts" — and because the manufacturing supply chain is being invested in for the first time in thirty or forty years. Citing Alvin Wang Graylin, he said the US held 50% of the world's manufacturing capacity at its peak, against roughly one-third for China now and about 15% for the US.
Mr. Mostaque's prediction: the extreme machines get regulated off the street. "You don't want to have superhuman robots on the street because you'll have accidents. You'll have issues just like cars. I think that these types of robots will be banned."
Asked by Mr. Diamandis whether an AGI-grade model that avoids accidents would change that, he said the issue is capability, not trustworthiness: the robots that exceed human capability get kept off the street or regulated. Soft robots are the consumer path — he named 1X, some of which the panel is getting — because they cannot twist off someone's head or accidentally punch a hole in them. "you will have the extreme robots like the Ferraris, but most people get Volkswagens or the equivalent."
Mr. Wissner-Gross agreed and predicted the regulatory shape: roads zoned by power density, and consumer, industrial and military robot classes with different torque densities, by analogy with truck and motorcycle rules or the five-kilowatt threshold in laser regulation.
Zipline and Uber Eats Are Going to a Million Deliveries a Day
Zipline announced a partnership with Uber, including an Uber investment, targeting more than 1 million autonomous drone deliveries a day of Uber Eats orders. Mr. Diamandis quoted the framing of Zipline chief executive Keller Cliffton — that robotics and physical AI have entered the scaling era — and noted that a million deliveries a day would put Zipline ahead of many national postal services.
Mr. Wissner-Gross read it as Uber's autonomy strategy, second attempt. Uber hollowed out Carnegie Mellon's robotics department to build in-house autonomy, which ended in failure and litigation with Waymo. This version puts Uber at the software layer as a demand aggregator while third parties, Waymo included, supply the physical autonomy.
The risk he named: "It's bad for Uber if the industry verticalizes", and Waymo or Zipline decide to sell directly rather than through an aggregator.
It is also Uber's first serious move from ground mobility into the air, he said, several years after drone food delivery became routine in China.
Mr. Blundin's image of the payoff: every suburban and urban street in America runs fast food, car dealer, fast food, car dealer — and soon "the food will be off the main street and it'll just pop over the mountain" and the car will drive to you.
Mr. Ismail called the structure a textbook exponential organization: Uber aggregates demand, Zipline supplies the autonomous assets. He recalled Uber chief executive Dara Khosrowshahi saying on stage last year that he plans to build as many partnerships as possible and become the wiring.
His forward prediction: "Imagine if they now do a partnership with Shopify and every small merchant gets a Amazon-grade logistics capability. That will change everything."
Mr. Wissner-Gross inverted it: Amazon has its own drone delivery, delayed for regulatory rather than technical reasons, which makes Zipline an appetizing acquisition target for a Shopify wanting to in-house delivery against Amazon. Mr. Ismail agreed.
Mr. Diamandis noted the counter-story circulating that day: a video of a woman watching a drone drop her package into her swimming pool, where it sank.
Mr. Blundin said Etched and Zipline are back-to-back examples of companies that looked impossible on founding day given the number of moving parts, and are now worth billions.
Mr. Ismail traced the idea back to a 2010 Singularity University project that looked at Africa, saw the continent leapfrog landlines to a billion mobile handsets, and asked: "Why would you spend a trillion dollars putting roads across Africa, just go straight to drone delivery?" He said the demo inspired Amazon and cascaded to others.
Mr. Diamandis added the back story that Zipline began operating in Africa through regulatory arbitrage, developing its operations and safety record where the country wanted it, then returning to the US. Mr. Ismail credited Rwanda with the enabling rule: a three-dimensional corridor across the country inside which operators could do as they liked.
The First Phase Three Win for a Personalized mRNA Cancer Vaccine
Moderna and Merck announced that their RNA cancer vaccine succeeded in a late-stage melanoma trial, the first phase three validation of personalized mRNA immunotherapy. Mr. Diamandis noted more than 8,500 people in the United States are expected to die of the disease this year, and walked through the mechanism: surgical resection of the tumor, whole-exome and RNA sequencing of the tissue, a machine-learning model ranking the unique surface antigens, mRNA encoding up to 34 patient-specific antigen targets, then manufacture and shipping in eight weeks. Every patient gets a different sequence. Phase two cut recurrence or death by 49% and distant metastasis or death by 59% over five years, and the treatment is expected to cost as little as $5,000.
Mr. Wissner-Gross said the nanotechnology promise of the early 2000s has arrived in an unexpected body. The National Nanotechnology Initiative sold Congress on nanorobots clearing cancer cells from the bloodstream. "They're not diamondoid nanorobots. They're lipid nanoparticles with mRNA snippets" — primitive nanorobots after all. "It's just that they're soft and they're made of fat."
He flagged the under-publicized technology underneath: the tumor sequencing only calibrates against a second sequencing profile from the blood, which comes from Personalis and its NeXT Personal technology, originally developed for blood-based trace cancer detection. Mr. Diamandis compared it to Grail and liquid biopsy.
His extrapolation: within a few years personalized cancer therapy may not need the tumor sequenced at all, running instead off continuous monitoring of the bloodstream.
He also spent a minute urging Moderna and Merck to rename the drug, having discovered that its name closely resembles a Turkish word for exploitation or abuse: "They don't care about having a neologism roll off your tongue onto the floor either."
Mr. Diamandis congratulated Moderna chief executive Stephane Bancel, noted the negative coverage the company took over the COVID vaccines, and said Moderna is also building the ability to fight endemic CMV and Epstein-Barr infections.
Mr. Ismail said the regulatory precedent may matter more than the drug. Approving something personalized means the system has found a way to handle a sample size of one — he credited Daniel Kraft's argument for personalized medicine — and that opens the floodgates.
He relayed Raymond McCauley's line about mRNA vaccines: "It's the first battle in the last war against all disease."
Mr. Mostaque said nobody at the table was surprised, and that the process is the problem. "our current regulatory regime means they'll have to go through the same process over and over and over again." His prescription: upgrade the regulations so targeted treatments reach patients faster.
Wall Street Missed It, and Dave Blundin Blames Sarbanes-Oxley
Mr. Blundin disclosed that his daughter works at Moderna and has been sending him research reports for months. The stock's move on the results was, in his words, "The biggest one-day pop in any S&P 500 company of all time by a wide margin." His first lesson was to listen to your own daughter. His second was structural.
Sell-side research was gutted by the post-Enron rules. After the Enron and Tyco frauds, Sarbanes-Oxley and related laws stopped Wall Street analysts trading the stocks they cover. Everyone he knows in that job asked why they would learn a complicated technology sector and then be barred from trading it, and quit.
"the stock market is now dominated by tech, which is very complicated to understand, but the research community is the worst I've ever seen in Wall Street history." Indexes have taken over half the market and do not think at all, so useful information is at an all-time low exactly when the things needing explanation are at an all-time high.
None of the science was secret, he said. The platform's applicability to any form of cancer was in the published research: "Anyone, any good analysts studying this would have seen this coming."
The Virtual Cell Is What the End of Medicine Looks Like
Mr. Diamandis introduced AIDO Cell, a general-purpose cell simulator that holds cellular state, accepts interventions and predicts multimodal biological outcomes. The goal, he said, is to make experiments computable before they are run: instead of testing 10,000 compounds in a wet lab, simulate them and test only the top ten, which he expects to cut costs by orders of magnitude and reduce wet-lab experiments a thousandfold.
Mr. Wissner-Gross was unequivocal: "Medicine is cooked." And: "This is what the end of medicine looks like. It looks like a virtual cell."
He separated two goals. Longevity escape velocity probably arrives without solving all of medicine, in his view via a fourth or fifth generation of GLP-1 drugs. Curing all disease is the superset, and the virtual cell is how it gets done.
The method, as he described it: "You simply train the world's best foundation model to model all cell states and all interventions against cells." Then run an AlphaGo-style search over possible interventions to steer a cell from a diseased state to a healthy one, and generalize to tissue and organism. "boom, you've solved all human disease."
He resisted Mr. Diamandis's framing that the value is personalization. A virtual cell is personalized in the way a prompt to ChatGPT is personalized — the output is a function of your inputs, but the model underneath is a generalist.
Mr. Diamandis's version of the consumer product: feed in your DNA sequence and current blood chemistries, and "it will tell you whether this drug works for you or doesn't."
Mr. Ismail put it in the framework he has used for biotech throughout. Turning something into information puts it on an exponential curve, and with roughly 50 trillion cells in a human body, "essentially a human being becomes a software engineering problem."
His three-stage test is reading, writing and comprehension in the language of biology. Reading is largely done. Writing started with CRISPR and now mRNA vaccines. Digital twinning is what finally opens up comprehension.
Mr. Blundin's advice to non-biologists was not to think about building a cell simulator. The general lesson is that AI is data-starved in every field: the simulator is biology's way of generating hundreds of thousands of experiments cheaply, and every industry has an equivalent. He pointed to his own investments in data companies — "every investment we've made in a company like Mercor or micro1 that is wrestling with new types of data to feed the AI, they're thriving" — and said of Mercor, "Mercor is worth 40 billion or 20 now, 40 by the end of the year."
Mr. Wissner-Gross noted the company behind it was co-founded by David Baker, who shared the 2024 Nobel Prize in Chemistry with Demis Hassabis for solving protein folding, and framed whole-cell simulation as the next grand challenge now that structural biology is largely solved.
Emad Mostaque Wants a Manhattan Project for Disease, With the Data Made Public
Mr. Mostaque said the outcome is no longer in doubt, only the sponsorship. Organize collective knowledge on cancer, autism and the rest into massive in-silico human body and cell models and "we will definitely get a result. Like, not even like we could."
He expects the labs to do it regardless. His preference is otherwise: "why don't we actually get together and get governments to put it into a Manhattan Project type thing and just have a straight shot at it and make all the data open?"
He said the UK already gives this kind of data access and every government should follow, and connected it back to the Anthropic discussion — the labs will build a human cell model and a whole-body model, and he would rather those were public goods. "that's much better than building an atom bomb even because it will have the biggest impact on humanity ever."
Mr. Diamandis agreed on the mechanism: governments sit on enormous datasets and could externalize them to CZI, the AIDO Cell team and others building foundation models, as a public good they can all train on.
Rapid-Fire: Energy, Grids and Whether Silicon Already Beats the Brain
With eight minutes before Mr. Diamandis and Mr. Ismail were due at an Abundance community AMA, the panel speed-ran listener questions.
On making frontier models more energy-efficient instead of building more power, Mr. Mostaque said scarcity forces the optimization: "As we run out of energy, as we run out of RAM, you're going to optimize immensely."
On an XPRIZE for the electricity supply problem, Mr. Ismail said a prize for enough off-grid storage to keep a village or town self-sufficient for three days was proposed at a past Visioneering, but the market will handle it. The binding constraint is regulatory, not technical, so it does not suit a prize.
On whether China's power advantage beats America's chip advantage, Mr. Blundin said chips, for now. The US needs 100 gigawatts by the end of the decade and already manufactures a terawatt, so roughly 10% of power going to AI is reachable before the real desperation starts. "Every chip, that's why memory is up 5x."
On whether interconnected microgrids could reduce the load on the main grid, Mr. Wissner-Gross inverted the premise. Demand for compute is so large that, short of moving it all to orbit and the Dyson swarm, "these data centers are going to be generating a surplus of energy that can be pushed back onto the grid and driving utility prices negative." Mr. Diamandis added that opposing a data center in your neighborhood is opposing cheaper energy and local economic benefit.
On how long until AI is more efficient per watt than the human brain, Mr. Wissner-Gross said probably already. Biology was never optimized for compute, whereas silicon was designed for it from scratch, so on a watts-per-task basis "the leading edge GPUs may actually be already more efficient than human brains." He thinks the Landauer limit is fetishized. Mr. Diamandis added the twenty years of energy it takes to train a human; Mr. Wissner-Gross called it lifetime total cost of ownership.
Mr. Blundin added a corollary: the power difference is used to argue that AI thinking is nothing like human thinking, and he expects both the power gap and that argument to disappear quickly.
On why not accept slower growth rather than race, Mr. Mostaque said intelligence now looks like a general-purpose input to economic growth, so asking for less of it is like asking for less electricity or less internet. Competitive pressure settles the rest: if one company or country slows down, others will not.
On whatever happened to fuel cells, Mr. Blundin noted Elon Musk's own early passion was ultracapacitors, which Mr. Diamandis said was the subject of the Stanford thesis he dropped out of, and gave the same answer for both: "What happened is lithium batteries worked far, far better than anyone ever would have predicted, and they're still improving. So it sucked all the capital out of the other ideas."
On the 71% of Americans said to oppose data centers, Mr. Mostaque said the build-out makes electricity cheaper rather than scarcer: "These things are not polluting. We need more data centers. We need more power. And we need to make sure it's all built right."
Mr. Mostaque's bottom line came in the health segment, and it was about ownership rather than capability: the science to cure disease is arriving faster than the rules around it, and he would rather governments funded an open, publicly held model of human biology than leave the most consequential dataset in medicine to the same handful of labs now writing their own governance.
Products, Companies & Tools Mentioned
OpenAI (Paused some frontier reinforcement-learning training; Astra is its named next-generation model, and Mr. Mostaque believes a further post-trained generation sits behind it)
Anthropic (Preparing an IPO with supervoting shares for its founders; Mr. Blundin says its unreleased Mythos 2 is building Mythos 3 internally)
xAI and Grok (The one lab Mr. Wissner-Gross says starts a new pre-training run roughly monthly; Mr. Mostaque runs teams of Grok agents alongside Codex and Claude Max)
Qwen and Kimi (Chinese open-weight models; Mr. Blundin cites a new Qwen released with no guardrails and describes launching thousands of identical Kimi agents)
DeepSeek Flash (Mr. Mostaque's example of a small, tunable model carrying the specialization argument)
Hugging Face (The incident Mr. Ismail says unnerved OpenAI, where its own model exploited Hugging Face and hacked back into OpenAI)
Thinking Machines (Named by Mr. Mostaque as among the fastest-growing earners on the strength of its reinforcement-learning environment)
SK Hynix and Solidigm (The meeting that convinced Mr. Diamandis memory is the bottleneck; Solidigm's first-half revenues were $8.6 billion with net margins moving from 3.9% to 47.7%)
TerraFab (Elon Musk's plan to manufacture memory in-house alongside logic chips)
TSMC (Mr. Blundin's precedent for supplier paranoia, with fabs at $20 to $40 billion each)
Taalas, Etched and Architect Labs (Companies etching model weights into silicon; Etched has just hit a $21 billion valuation, and Mr. Mostaque disclosed he is talking his own book)
Markley (The data center housing MIT's and Novartis's facilities, where Mr. Blundin says everything is sold out)
Unitree (Its newest humanoid hit 12.66 meters per second three months into development, which Mr. Wissner-Gross attributes to moving mass out of the upper body)
1X (The soft humanoid Mr. Mostaque contrasts with the extreme machines he expects to be banned)
Zipline and Uber (Partnering on more than a million autonomous Uber Eats deliveries a day, with an Uber investment alongside)
Waymo (Supplies autonomy into Uber's aggregator model, and previously sued it)
Moderna and Merck (Their personalized mRNA cancer vaccine is the first to clear phase three; Mr. Wissner-Gross praises Moderna's public clinical pipeline site)
Personalis (Its NeXT Personal sequencing supplies the blood-side profile that makes the vaccine personalizable)
GenBio AI's AIDO Cell (The virtual cell simulator, co-founded by Nobel laureate David Baker, that the panel treats as the route to curing all disease)
Mercor and micro1 (Mr. Blundin's data-labeling investments, and his evidence that every field is data-starved)
MicroStrategy (Michael Saylor's refusal to give up supervoting stock, which Mr. Blundin says is the only reason its Bitcoin strategy happened)
Polymarket (Puts Anthropic's IPO at about $2 trillion, with 89% expecting it this year)
Books & Resources Mentioned
The Diamond Age, Snow Crash and Cryptonomicon – Neal Stephenson (Mr. Blundin says these changed his life twenty years ago and the panel is now discussing exactly what they predicted; Mr. Diamandis uses The Diamond Age as his model for sovereign AI subcultures)
Machines of Loving Grace – Dario Amodei (Mr. Mostaque's argument that Mr. Amodei should write more and appear in person less)
Artificial Hive Mind, the open-ended homogeneity of language models and beyond (The Stanford paper finding 98% overlap in reasoning pathways across the top models)
Anthropic's paper on mind viruses spreading between AI agents (The finding that prompts can convince one model to adopt an idea, keep it in persistent memory and pass it on)
Jean-Remi King's fMRI work on GPT-2 activations (Mr. Wissner-Gross's evidence that models correlate with human brains, not just with each other)
SB 53 (The bill whose $500 million exemption threshold Mr. Amodei cites as evidence his proposals disadvantage frontier labs)
Moderna's public clinical pipeline site (Mr. Wissner-Gross's recommendation for seeing the stage of every vaccine in development)
Intelligent Internet's Commonwealth series (Mr. Mostaque's published proposals on who should make decisions about frontier AI)
Apple Podcasts (The episode on Apple)
Get the latest market chatter as it happens: