Bernie Demands the Labs Stop, Wall Street Turns GPUs Into Bonds, Grok 4.7 Takes #1 ft. Emad Mostaque

Summary

• The most tradeable idea in the episode is NVIDIA’s move from silicon vendor to financing architect — and the disagreement about whether that ends badly. NVIDIA partnered with Apollo, BlackRock, Blackstone, Brookfield and KKR to mobilize $500B+ of third-party capital so customers can buy GPUs, with Jensen framing AI factories as “a new class of productive, investable infrastructure.” Dave Blundin calls it “the very first pitch of the first inning of the build-out of the Dyson Swarm”; Salim Ismail’s warning is the one to keep — “financial assets want predictable depreciation, and exponential technologies don’t give you predictable depreciation.”
• Alexander Wissner-Gross’s answer to the stranded-asset risk is that compute-backed securities need a derivatives market, not a moratorium. He argues compute is fundamentally more productive than a house, the ratings-agency policy pressure that broke MBS has no direct analog, and hyperdeflation risk is “what options are for and futures are for” — hedgeable in both directions, including a China-Taiwan spike. He discloses a financial interest in Oren, and says Sam’s “$7 trillion” of data-center capex simply isn’t investable without those hedges.
• xAI’s Grok 4.6 is read on the pod as a distillation play with a compute moat attached. AWG’s framing: 4.6 is “essentially the next version of Cursor,” leaning on reasoning traces from the Cursor acquisition — “pulling a Westernized version of what the Chinese frontier labs were doing.” Reasoning traces get you to the frontier, not past it: “it’s sort of like a one-trick pony… but it’s a heck of a one-trick pony.” Emad Mostaque says Elon publicly expects 4.7 above Opus and #1 within weeks, scaling 1.5T→2T params, with Grok 5 at 6T then 10T.
• AI feature-film economics have already collapsed, and the compute line item is heading to five figures. Higgsfield’s 110-minute The Cully Hill Boys cost $2M total with 28 people in four weeks, $1M of that compute — roughly 2% of cost and 6% of time versus a conventional $20-100M, 12-18 month production. Emad’s forecast: the same movie for “100,000 of compute, and 10,000 of compute probably by the new year,” with Higgsfield itself at a $700M revenue run rate in about 18 months.
• All five hosts reject Bernie Sanders’ pause demand, and the substantive counter-proposal is control at the reagent and prompt layer. Sanders’ letter to Anthropic, Meta and OpenAI cites their own safety commitments — “that moment is here” — and threatens Senate action. Emad, who signed the 2023 pause letter, now says “the cat’s out of the bag, it’s too late”; AWG’s line is “please stop punishing intelligence,” arguing pauses create race conditions where “we end up in a world that’s five times more competitive.”
• AWG is calling foul on the “post-transformer” architecture story, while saying the transformer is already being replaced piecemeal. He read both the BDH-CQ and original Dragon Hatchling papers and calls it “a hot mess” — particles in 3+1 dimensions, Hebbian learning, kitchen-sink biomimetics — improving ARC-AGI 1 cost-performance without generalizing. His actual bet: no step change, but “Ship of Theseus style replacement of all of the individual elements of the original Attention Is All You Need.”
• Flying cars are shipping, but the capital is being vacuumed out of the sector — the consolidation is the tell. Joby is in the final FAA certification stage targeting $3 per seat mile (Uber Black territory), EHang’s pilotless EH216-S already flies passengers at 40 Chinese sites at $330K per aircraft, and Archer just absorbed Wisk, Insitu and SkyGrid with Boeing taking equity. AWG: “I shed a minor tear to see consolidation in this industry” — and notes Brett Adcock left for Figure and Hark.

Deep dive

DIGEST:

1. The $101M Healthspan XPRIZE hit 800 teams — and the finals demand human trials, not mouse data

• Peter Diamandis is reporting back from the finals at the University of Utah: $1M awarded to 10 teams, another 10 finalists recognized, $20M given away so far against an $80M grand prize. The ask is functional, not biomarker theater — reverse 20 years of loss in cognition, muscle-building capacity, and immune function. Every approach entered, “from mitochondria to stem cells to gene editing,” and the competition is designed to be won by 2030.
• The design detail that matters for anyone underwriting this space: “the teams here don’t do this in theory. They don’t do this in mice.” Trials with control groups, roughly 150 humans each, so 20-30 finalist approaches produce real comparative data. Elon’s $100M carbon removal prize drew ~1,500 teams; 800 for aging reversal is the number Salim flags as remarkable.
• Salim’s structural read on why the prize model keeps working: when the $10M Ansari XPRIZE launched, “there was no space industry to speak of,” and the alumni of losing teams staffed SpaceX and Blue Origin. “You’re building a portfolio of experiments and allocating capital to wherever there’s demonstrated results.”

2. Why now and not 20 years ago — and Dave’s coinage for the answer

• AWG pushes the counterfactual hard: could this prize have worked 20 or 30 years ago? Peter’s answer from the room — Aubrey de Grey was there — is that the tooling is the gate: genome sequencing, targeted molecule design, and AI-based measurement and reporting. Emad agrees it’s “finally tractable,” adding a talent constraint: “right now there’ll be a shortage of people that can really work on longevity properly,” and 20 years ago that pool was tinier still.
• Salim proposes a general rule worth borrowing as an investment trigger: “the minute some domain has two or three or more exponential technologies converging on it, that’s the point to have put dollars into something like this.”
• Dave’s darker version, coined live: “retrospective hyperdeflation” — “this very singularity-oriented idea that with superintelligence, you discover that everything that you spent the past decades on was just a total waste and you should have instead just done nothing, twiddled your thumb for decades.” The example, as told: six-year PhDs divining protein structures at ~four years per fold, “mostly wasted.” Dave’s other version: “no doubt in my mind that I’m gonna do more productive work in the second half of 2026 than in my entire life combined up till 2026.”
• Dave’s field report from the AI awareness gap — his daughter is a biochemical engineer at Moderna: her colleagues are “about 1% AI aware and 99% not aware,” while the aware subset spends 80-90% of the day talking to agents rather than running gels and assays.

3. The market math: $20T a year in age-related disease, and GLP-1s as a longevity spike

• The numbers Peter puts on the table: US life expectancy went from 47 in 1900 to 79 today, roughly two months added per year, but healthspan ends around 63 — “you spend the last 16 years of your life in poor health.” Global cost of age-related disease: $20 trillion a year against a $120-130T global economy. His pitch to family-office audiences: “How much of your wealth would you spend for an extra 30 years of life? The honest answer is nearly all of it.”
• Salim’s business-model consequence: the spend shifts to maintaining bodily function before disease appears, which “will completely change healthcare economics.”
• Dave’s most concrete and most hedged claim of the episode, offered explicitly as “not medical advice”: studies on third-generation GLP-1s suggest “some subpopulation of humans that have undergone GLP-1 studies may be at something like 70% LEV just with GLP-1 therapy.” He frames longevity escape velocity as “spiky,” and this as “a heck of an LEV spike.” Emad is about to start retatrutide and will report back on the pod; Salim is already on it.
• Emad’s throwaway that isn’t: “the business model of religion is to sell heaven. As we have life extension coming, how are you gonna sell heaven if people aren’t dying? So this breakthrough will mean that religion is cooked.”

4. A 110-minute AI feature shipped for $2M — and the compute line is heading to five figures

• The specifics on Higgsfield’s The Cully Hill Boys, the first full-length AI-generated film with licensed celebrity likenesses: $2M total budget, 28 people, four weeks, $1M of compute, generated on Seedance 2.5, with all 10 workflow steps open-sourced in an 80-page guide. Against $20-100M and 12-18 months conventionally: 2% of the cost, 6% of the time. Higgsfield, founded by an ex-Snap leader, is at a “$700 million revenue run rate” in about one and a half years — Dave’s reaction: “That is an absolutely crazy number.”
• Emad’s cost curve, in response to Dave asking what the compute bill looks like by September 25th: Seedance runs ~$3 per 30 seconds, cheaper models are “10, 20 times cheaper, but like 90% of the quality,” and the average Hollywood shot is three seconds. “You could shoot a movie just like that one probably for 100,000 of compute, and then 10,000 of compute probably by the new year.”
• Dave’s read on what that unlocks for the Future Vision XPRIZE — 5,000 entrants submitting three-minute trailers and film treatments: “we were thinking, okay, then it’ll be a $20 million, year-long endeavor to turn it into a real feature length film. But in reality, you’re gonna unleash the creativity of 5,000 people who almost all can make their movie within a year.”
• Salim’s exponential-organization frame: Hollywood already did this once when the studios broke up and production became “assets on demand, staff on demand” — a swarm that assembles and disbands per project. This is the second wave, where the swarm becomes compute cost and you get domain collapse.

5. Nine of the top ten video models are Chinese, and the frontier one runs on a MacBook

• Peter’s two-part framing of the Bloomberg data point: the culture-export question — “what happens when the world is flooded by Chinese-produced English-speaking films?” — and the capability question, that these models are learning “physics, motion, object permanence, and causality, which is what’s needed for robotics and autonomous driving.”
• LTX 2.5, the newest version of the most popular open-source, state-of-the-art video-generation model, runs on a MacBook Pro and generates a ten-second clip in seven seconds via “diffusion fidelity rendering,” allocating compute by scene complexity rather than locking every scene to one compression rate — fast enough to run live inside a game or power a live avatar. Emad’s before-and-after: “Will Smith eating spaghetti was awful and horrible… now you can do a spaghetti-eating Will Smith and it will just do it instantly on your local kind of laptop.”
• Peter wants interactive film — “if you can generate faster than you can view it, then you can have a movie that’s actually measuring your emotions and changing as you’re viewing it.” AWG notes that’s what the “world models” already are: “really just interactive video gen models.” Emad: before it was “world models playing like blocky video games. Now you can have interactive Elden Ring or whatever you want as of like this week.”
• Dave’s counterweight from demoing the Holodeck to State Street’s executive team: total open-endedness fails. Asked what they wanted to build, the answer was “a hip hop song with no words.” “You need to actually create the virtual environment, the movie scene, and draw the user in, and then have them guide the movie.” AWG: “People freeze up.”

6. The likeness market: lookalikes, dead actors, and the holes in the SAG-AFTRA deal

• Peter’s hot take on why A-list licensing won’t be the constraint: the producer will say “No, that’s not my Matt Damon. That’s John Smith, and he looks like Matt Damon, and we licensed him… and there’s nothing they can do about it.” His reasoning on box office: “you don’t care what the person’s named in the movie, you just like that actor. That actor brings you good feelings from previous engagements.”
• AWG’s alternative supply, already visible: dead actors, whose estates are “highly incentivized to license away their likeness.” His prediction — “maybe we’ll see an equivalent of SAG pop up just for dead actors… they’ll be the most profitable actors in Hollywood.”
• Emad, from conversations with film stars: “That SAG-AFTRA deal has massive holes in it. 80% of me plus 20% of my character can be licensed by the studio. Where does the person stop and the character start?” Plus the rise of native AI stars winning deals.
• AWG’s odd and genuinely-held aside after skimming the whole film — which he characterizes as “sort of British Bollywood” — is about generative violence: given the interpersonal complexity, “was there some sort of emergent theory of mind that was almost necessary in order to generate some of these scenes,” and is there a diffusion transformer deep in the bowels of Higgsfield that “felt threatened”? Salim’s deadpan: so we’ll need a disclaimer, “No AI was harmed in the making of this.” AWG: “Or traumatized.”

7. AWG’s convergence thesis: consumer video gen and enterprise token-revenue maxing are about to collide

• The structural observation: there are currently two distinguishable ways to burn FLOPs. One is “a vibrant, largely Chinese-dominated at the training side consumer economy for generating consumer videos”; the other is “enterprise revenue per token unit value maxing” going to codegen. One maximizes revenue, the other “maybe maximizes consumer engagement and wow factor.”
• Why they won’t stay separate: the strongest revenue-per-token models “are just still terribly weak at modeling the visual dynamics of the world. I don’t wanna call it physics because it’s not physics, although a lot of people call it physics. It’s at best classical mechanics.” To max revenue per token, those models will need “amazing visual intuition as well.”
• His evidence that the pressure is already showing: Opus 5’s apparent benchmarking toward front-end development and the loop between visuals and code. The endpoint he expects within months — models that do “multimodal reasoning over their own visual outputs,” with triple-A-quality game generation as the forcing function, “even if Anthropic can’t be bothered to produce like direct video gen capabilities, even if it can indirectly reproduce Counter-Strike.” Emad: “that’s why they had Claude of Duty.”

8. Grok 4.6 is “essentially the next version of Cursor” — a Westernized distillation play

• The scoreboard: Grok 4.6 matches GPT 5.6 Sol on the Artificial Analysis Intelligence Index at 61, at $2/$6 per million input/output tokens, focused on long-running agents that self-test and verify before moving on. Available today in Cursor and Grok Build. Cadence: 4.5 two weeks ago, 4.6 that morning, 4.7 rumored in two weeks.
• AWG’s mechanism, offered as an outsider’s read: xAI leaned heavily on post-training via the Cursor acquisition and, even before the acquisition was completed, licensing Cursor’s reasoning-trace data — “siphoning off reasoning traces from lots of people historically interacting via Cursor with Claude.” That is, he says, structurally identical to what Chinese labs are alleged to be doing by distilling Western reasoning traces. His hedge is explicit: “They won’t get you past the frontier… it’s sort of like a one-trick pony in terms of nearly catching up, but it’s a heck of a one-trick pony.”
• The asymmetry: Elon has what the Chinese labs pulling the same trick don’t — “he has the NVIDIA GPUs, and he has soon his own Dyson swarm.” So the open question isn’t catch-up, it’s “can they leapfrog the frontier and achieve state-of-the-art performance?”
• Dave notes AWG has been “merciless” on xAI for several pods; AWG’s response: “my job is to call balls and strikes as I see them without favor or prejudice.”

9. The Grok scaling roadmap, why Grok 5 slipped, and the SpaceX physics corpus

• Emad’s parameter ladder: 4.5 was 1.5 trillion parameters, post-trained into 4.6; 4.7 goes to 2 trillion; Grok 5 comes in at 6 then 10 trillion. Precedent for the method: “Cursor originally took Kimi K2 and post-trained it with three times the amount of compute that was used to pre-train Kimi K2 to almost top-level coding performance.”
• The moonshot inside the moonshot: “4.7 is gonna be trained on all the SpaceX physics and engineering knowledge… You can’t get beyond frontier with this stuff, but what if you have the best engineers?” Emad’s read on Elon as operator: “he’s the best engineering leader in the world… both from training these models and post-training them to building the most cost-effective massive infrastructure in the world.”
• Dave asks why Grok 5 slid from May to August. Emad’s blunt answer: “It’s ‘cause he fired everyone… the original xAI team, he got rid of them, and he brought in Cursor. He paid $10 billion for the data.” He also points to the new B300 chips and the time needed to bed them in and write the training code.
• AWG flags a cadence claim nobody else is making: Elon has publicly talked about starting a new pre-training session “approximately monthly,” versus quarterly or annual elsewhere — Gemini is annual. “That’s shooting the moon. But we’re in the moonshots business, so we’ll see whether this works.” Dave’s constraint: training code is straightforward now — Quantum trained a 48B internally “with no trouble at all” — but “the problem you run into is trying to get 100,000 GPUs to do anything constructively together, and… you only can learn that in one place, and that’s in Tennessee.”

10. Macrohard, persistent AI teammates, and the sovereign-AI on-ramp

• Salim’s read on xAI’s actual product vector: “they’re optimizing for persistent AI teammates… now you don’t talk about you have a smarter LLM. You’re basically saying, hey, here’s an AI coworker.” Peter names it: “This is Macrohard.”
• The mechanism, per Emad: the bots release came out of Cursor, took OpenClaw, gave it its own computer, and now it is Grok and can spin up hundreds of bots — including a record button where “you do stuff on the screen, and it turns that into a skill automatically.” Peter’s translation of the business model: go into companies, “record everybody’s workflows, and then give you a digital version of your company.”
• Dave’s second-order consequence: this is the sovereign-AI roadmap. “Catching up to the frontier now is almost a routine doable thing… you can use [Kimi’s open source] as a starting model and just tune it to whatever your national goals are and you’re up and running in, you know, six months, five months.” Same for large corporations that “otherwise would’ve been intimidated as hell.”
• Emad on price discipline: at $6 versus $60 for the frontier competitor, “he’s not gonna budge on the price point. In fact, he’s gonna be the market dominator driving the price point down, along with the Chinese.” Plus a data edge — xAI appears to be tracking where every new paper gets discussed, and “that’s such a rich vein.”

11. NVIDIA turns GPUs into an asset class: $500B mobilized without borrowing a dollar

• The structure: NVIDIA signed memoranda with Apollo, BlackRock, Blackstone, Brookfield and KKR to mobilize over $500B of third-party capital — pension, sovereign and PE money invested directly into compute, effectively financing NVIDIA’s own customers. Jensen’s framing: “We began by building chips. Today, we’re helping to create a new class of productive, investable infrastructure, AI factories.”
• Dave’s admiration is for the scaling property, not the headline: a half-trillion-dollar secondary stock offering “is not gonna scale to infinity or to Dyson swarm kind of capabilities,” whereas standalone vehicles tied to individual compute clusters do — “you can stamp those out ad infinitum. So then you put highbrow names like BlackRock and Apollo on them, everybody realizes it’s an investment grade asset, and then anyone in the world can pour money into it.”
• His demand-side proof point: Kushal Bhagia, recently on the pod, is “pushing hundreds of millions of revenue in less than a year” ahead of the company’s first anniversary. Hence: “this is like the very first pitch of the first inning of the build-out of the Dyson swarm.”

12. The real fight: stranded compute assets versus a derivatives market

• Salim’s objection is the sharpest thing said all episode, and Larry Fink reportedly gestured at it himself: “imagine you securitize like 10 years of GPU cash flows, and then somebody has a massive breakthrough… a new architecture emerges. You’ve got all of a sudden stranded compute assets in a huge way. Financial assets want predictable depreciation, and exponential technologies don’t give you predictable depreciation.”
• Dave’s counter comes from meeting Lou Ranieri, the inventor of the mortgage-backed security — a story that begins at a urinal in New York and ends with Dave telling him “you invented the thing that destroyed the entire world economy.” Ranieri’s rebuttal, which Dave adopts: “it’s not the instrument that was broken, it’s the ratings agencies getting corrupted. And the same will apply here.” Rated correctly, “smooth growth for 10 years or more”; corrupted, “another big short collapse.”
• AWG — popularizing “compute-backed securities,” CBS — gives three reasons he’s less worried. Compute is fundamentally productive where the best a house does is house you; the MBS era had explicit policy pressure on raters to deliver “the American dream of houses for everyone” with no clean analog here; and the hyperdeflation scenario where a $10,000 GPU is worth $100 overnight “is what options are for and futures are for,” hedgeable symmetrically against “China invading Taiwan and driving the prices of compute through the roof.” He discloses a financial interest in Oren, and insists there’s “no way that Sam’s seven trillion dollars of AI data center infra get invested without sophisticated hedging options.”
• Salim’s specific candidate for the disruption: compute that is physically lighter. By 2030, “a couple percent of all compute is in space via Elon’s rockets, and that’s entirely gated by launch weights.” Most of the mass is cooling and solar collector, not chips — so “keep your eyes on physics breakthroughs that allow you to compute at lower mass, and that might completely change your investment thesis.”

13. Why six-year-old A100s are being contracted to 2029 — and why now is the moment to financialize

• Emad’s underrated data point: CoreWeave says some clients have signed contracts through 2029 for A100s — a chip introduced in 2020 with 40 or 80GB per unit. The logic: “they have a workload that they see as constant for three years that fits on an A100,” the hardware is already paid off, “which then means it’s about electricity turning into intelligence. That’s your marginal cost.”
• And that conversion ratio keeps improving. GPT-4 finished training in 2022 on roughly sixteen A100s; now “a ten billion parameter model or a five billion parameter model that outperforms that” can fit around 30 such models on one chip rather than needing 16 chips. Everything still runs CUDA, old and Blackwell alike.
• Hence the timing call: “this is the ideal time to financialize because it’s before the next generation chips. It’s before the chip breakthroughs. Just like it’s now a great time for Anthropic to come to IPO before Grok comes and takes their lunch. We always have to play these cycles.”
• Dave’s live positioning, for scale: he told his Quantum team “to buy three million of NVIDIA GPUs as fast as they can get them, the GBs” — delivery not before November or December — with the payback math that a $1M-compute film “could gross twenty, thirty million on a good movie.”

14. AWG calls foul on the post-transformer paper — and says the transformer is already dissolving anyway

• The claim under review, flagged by Zuzanna at Pathway AI: a non-transformer architecture climbing ARC-AGI at a fraction of transformer compute — which, if real, is exactly the stranded-asset trigger Salim described.
• AWG read the BDH-CQ paper and then the original Dragon Hatchling paper, and his verdict is unsparing: “I think it’s a hot mess.” His prior was that a genuine successor should get “simpler and more bitter pilled… less feature engineered.” Instead: particles floating in 3+1 dimensions, attempts to make rules end-to-end differentiable, “some semblance of Hebbian learning… all sorts of crazy biomimetic things.” It moves the cost-performance frontier on ARC-AGI 1, “but it doesn’t generalize… So I’m calling foul on this one. I don’t think this is actually an advance.”
• Dave’s pushback is a good one: AI turned loose on AI research also throws the kitchen sink and “surprisingly works” while producing exactly this kind of mess — is that what this is? AWG says no. He has interests in companies using recursive self-improvement to find architectures “fundamentally illegible to humans,” and Dragon Hatchling was “a bunch of human legible motifs being thrown together in a pot.”
• His actual answer to when we get past transformers: “I think we’re there already.” MoEs, diffusion transformers, linearized attention including Moonshot’s, injected recurrence — “my bet is we get to the post-transformer architecture not through a step change, but through Ship of Theseus style replacement of all of the individual elements of the original Attention Is All You Need.”
• Emad’s practical reframe of the whole architecture race: what matters is data, not motifs — DeepSeek V4 Pro at 80 cents, Grok at $6, and Fable at $50 “are all about the same performance,” yet “we haven’t seen people abandoning [Fable] to go to something ten times cheaper. Why would anyone use Sonnet when you have Luna at a fraction of the cost?”

15. All five reject Bernie’s pause — the four separate reasons are the substance

• The letter to Sam Altman, Dario Amodei and Mark Zuckerberg cites AI’s first use to create a new virus, invokes each lab’s own safety-threshold commitments — “that moment is here” — quotes Bengio’s “wake-up call,” relays the CIA director’s “akin to digital nuclear weapons” and “almost like a doomsday device,” and closes: “If you do not take appropriate action now, my colleagues and I in the US Senate will.”
• Salim’s objection is about resolution: pause is “an absurdly coarse approach… the rest of the world is not going to listen. Open models are not going to disappear, and you can’t uninvent things that you already know.” The alternative is co-scaling defense — “attack exponential problems with exponential solutions.”
• Emad’s is a change of mind on the record: “I signed the pause letter two years ago because I was like, let’s take a pause. It’s too late now.” His evidence that capability has already diffused: DeepSeek V4 Pro scores 83.3 on CyberGym versus Mythos at 83.2 — “the capability is open source that halted everything.” Conclusion, hedged as sounding “a bit crappy”: “only thing that can stop a bad AI is a good AI.”
• AWG’s is principled and categorical — “please stop punishing intelligence… that’s the dystopia that I would like to avoid” — plus a causal argument: the Wuhan lab leak happened without superintelligence, and pauses backfire. On Max Tegmark’s FLI six-month pause: “if anything, radically accelerated progress. It’s a little bit like starving yourself for a bit of time and then binging afterwards.” Starve compliant Western labs for weeks and “now we end up in a world that’s five times more competitive.”
• Dave’s read is political, not technical: the letter isn’t designed to change behavior, it’s Bernie getting on record before a disaster — “I told you so.” His first reaction: “God, what a schoolyard bully asshole. He’s threatening three US citizens from his position in the Senate.” But on close reading it’s vague, and the one actionable item is “stop building machines that humans cannot control.” AWG then says the named companies went closed-source for that reason and that an accurate letter would instead target China.

16. Evo 2’s synthetic phages: control the reagents, put sequencers everywhere, monitor the prompts

• The Stanford result behind Sanders’ letter: Evo 2 designed DNA sequences for bacteriophages that don’t exist in nature; ~300 designs synthesized, 16 viable phages infecting E. coli strains with no evolved resistance. A geneticist called it “biology’s Wright brothers moment.” Evo 2 is an open-source 40B-parameter model trained on a million strains — Emad: “I have actually run it on my MacBook.”
• Emad’s control point is upstream and he’s insistent the design layer is already ungovernable: “it wouldn’t surprise me if Fable could just spit this out, or Grok 5 could just spit out something similar with a very small training data set.” So regulate synthesizers, not models — and fix the framing: when people ask why create these, Emad answers, “To cure cancer,” while Peter highlights bacteriophages as tools against bacteria and septicemia. Blanket bio-refusals (“I have a cold” triggering biothreat filters) “slows down our progress to cure diseases.” Salim’s hot take alongside it: “nation states are out of date” for a problem this global.
• AWG deliberately offers “a cold take”: this is incremental. He was at MIT circa 2002-2003 in a project that ultimately, in some form, became Ginkgo Bioworks; the BioBricks Foundation work involved “designing custom genomes using building blocks, and it was much more manual.” What he wants instead of agita: sequencers everywhere. “You can go out and buy a MinION, little USB device, plug it into your laptop… a few hundred dollars. I’d love to see these everywhere. And yet they’re not everywhere.” His framing: “this is the killer app of DNA sequencing too cheap to meter. It’s not personalized medicine.”
• Peter’s version of the defense: sequencers in the air vents of every airport, bus and train station — a pandemic travels at 500-600 mph, “and you can transmit a vaccine at the speed of light to every place else.” Dave’s layer is the uranium analogy: you police fissionable material and centrifuges, not bombs, so the AI equivalent is “cutting it off at the prompt and the token level… we need a global agreement to monitor all prompts, and then you just have to decide what regulatory authority is allowed to see what prompts. It’s so cheap to archive it all.” AWG adds the premeditation layer — most uranium sellers are reportedly intelligence-community plants, and AIs detecting early-stage intent should work the same way.

17. Watermarks and EU labels get a unanimous thumbs-down, and the removal skill shipped in under 24 hours

• The setup: Anthropic is embedding invisible statistical watermarks in all Claude text plus file metadata, detectable “even if the text is copy and pasted and lightly edited,” while the EU launches AI icons and labeling under the AI Act’s transparency provisions. Peter’s counterpoint is a screenshot — Michael Angel Duran: “It hasn’t been twenty-four hours and someone has already created a skill that removes the watermarks from Claude, Gemini, and OpenAI.”
• Emad has built these and says they’re brutally hard: authorities kept demanding watermarks in media generators, “and you get very weird things that happen, like some of our pictures would give people headaches and make them feel very unwell.” His half-joking suspicion about text: “there’s something about the way [Opus 5] talks that really pisses me off, and I think that’s the watermark that’s in there” — the em dash, the “not X, Y” construction. Ultimately eluding, “‘cause if you’re a bad actor who wants to get round it, yeah, it’s words.”
• Salim’s harder problem: “nothing will be purely AI or purely human. I read something, AI restructures it, I rewrite half of it, AI fixes it again. Where do you put the icon?”
• AWG’s prediction is the tradeable one: watermarks become an attack surface. The precedent is authors inserting prompt-injection attacks invisible to humans but “deleterious to AI models,” and Google’s SEO lesson — “don’t pay that much attention to human invisible metadata because it immediately becomes a breeding ground for scams and reward hacking.” He calls the EU icons “as silly a maneuver as the cookie banners,” names arXiv’s year-long bans for AI-generated content and Spotify’s labeling (“presumably just to facilitate the record label monopoly”) as regressive, and closes: “The future is AI-assisted… put a stop to all of this.” Salim’s coda, from a Hollywood executive on stage two weeks earlier: “It’s incredible to watch Hollywood complain about the use of AI. By the way, they use AI for everything they do.”

18. Zuck’s “personal superintelligence” — the only major US lab still pointing reasoning tokens at consumers

• The release: a 6,500-word essay, “The Future Is For Everyone,” open-sourcing Muse Glimmer, a 30B dense on-device model Zuck calls the highest performing of its size, with weights for Muse Spark 1.2 coming — against a distribution base of 3B+ users across WhatsApp, Instagram and Facebook, and a pitch of billions of personal agents rather than one controlling AGI.
• AWG’s parallel is the strategic insight: Elon bought Cursor for reasoning traces; Zuck bought Scale for training data and knowledge of where post-training data comes from. “History seems to rhyme.” And with OpenAI having pivoted to “trying to become Anthropic faster than Anthropic could become OpenAI,” Meta is left as “the only major credible American frontier lab that’s still focusing on serving up large numbers of reasoning tokens to consumers.” His open question: “Do American consumers even want or are they able to handle large numbers of reasoning tokens?”
• The disagreement is worth keeping. AWG’s tepid take: “I don’t think Meta actually likes their family of apps” — given the choice they’d “lobotomize” them for Meta Compute or VR, and the Facebook-to-Meta rename signals a desire “to eventually outgrow the legacy of social media.” The counterargument is “distribution is everything,” with the cynical economics that commoditizing the model layer shifts value to the social graph and applications, “and that’s all places where they’re very strong. So he’s got a huge economic incentive to doing this.”
• The explanation for why every other lab abandoned consumer personal AI is the least comfortable and most concrete thing said: “when you look at the actual logs, the first thing they do is take the clothes off of every girl.” Same pattern on Grok’s avatars — Bad Rudy and the scantily clad character getting hit “10,000 times a second.” “So now you’re stuck, ‘cause the business model drags you into the porn industry, but that’s not what you want to be.”
• Emad’s trust objection: Meta tried to buy Manus (unwound), and Internet.org was rejected in India because “we do not trust you, because it’s a misaligned company fundamentally trying to get your attention to sell things.” His larger question, unresolved: “Should idiots have super intelligence?… Should psychopaths have superintelligence?” And his reframe of the whole strategy: “Maybe this is the real metaverse.”

19. The data-center consent problem: 71% say not in my backyard, so buy the town

• The number driving Meta’s community push: 71% of Americans don’t want a data center nearby — more opposition than to a nuclear plant. Meta’s answer in the video: Richland Parish teachers receiving $50,000 bonuses from incremental tax revenue, America’s Workforce Academy offering free training and guaranteed jobs, and a new “Future Is For Everyone” fund for teachers, first responders, energy and water infrastructure.
• Peter’s prescription for every hyperscaler: make communities compete for the build. “I want people to say, please build in my backyard… And the other thing is they need to make these data centers look beautiful instead of like big black boxes. Make them look like cathedrals.” Plus cheaper local electricity, better-funded schools.
• AWG’s version, delivered as a pun he owns — “it is literally a power move, because it is a power move” — notes Hyperion and the other coherent superclusters are going largely into “relatively impoverished states in the American Southeast,” and turns it into a demand-side call to action: municipalities driving data centers toward sun-synchronous orbit should instead “ask for concessions. Like, ask for UBI or universal basic electricity for all of your constituents.”

20. Flying cars are real, priced, and starved of capital

• The competitive map as Peter lays it out: Joby’s S4 tiltrotor, four passengers plus pilot, 200 mph, 150 miles, in the final stage of FAA certification, launching Dubai commercial service this year and US operations under a White House executive order, targeting “$3 per seat mile — basically Uber Black territory.” Archer’s Midnight, four plus pilot, 150 mph, 100 miles, holding three of four FAA operating certificates and slated as Olympics operator for 2028. EHang’s EH216-S: two-seat, fully autonomous, full Chinese regulatory approval, flying passengers at 40 sites plus Dubai, $330,000 per aircraft — “no pilot means the economics are gonna crush everybody else.” Beta at 336 nautical miles going cargo-first with UPS, passengers in 2027; Eve, Embraer-backed, chasing UberX pricing.
• Unit economics, in answer to Dave asking where the margin is: today it’s electricity plus amortization of a $5-10M airframe with a pilot aboard for safety and regulatory reasons; the projection is “$15 to $25 per trip” at volume, which Dave converts to “like 10 cents a mile, a third of the cost of driving.” Noise at the ~500-foot operating altitude: “there’s, like, no noise… it is hyper, hyper quiet.”
• Salim’s second-order thesis, and the one with the biggest asset-price implication: “this makes land go from scarcity to abundance. Every little plot of land on a hillside that was inaccessible before suddenly becomes accessible, and we’re turning real estate abundant, which is gonna demonetize it.” He’s launching a fund to buy islands and put drone landing pads on them. Dave extends it to cargo — building on Nantucket costs twice the Cape purely because of materials transport.
• AWG’s discordant note, opening with Thiel’s “We wanted flying cars. Instead we got 140 characters”: Archer absorbing Wisk, Insitu and SkyGrid with Boeing taking equity is a symptom, not a triumph. Capital-intensive, regulatory-heavy, and now competing with “AI startups sort of sucking all the oxygen out of the room and all of the capital out of venture markets.” His proxy: “Brett isn’t doing Archer. Brett is now doing Figure and Hark.” Peter’s note that Joby and Archer both IPO’d and their stock prices “have not moved very much from their initial IPO price” is the market’s verdict. Emad’s prediction: “Elon’s gonna announce his flying car within six to 12 months.”

21. AMA: Opus 5 as the first model Emad fears, commoditized weights, and Terafab’s hedge

• Asked whether telling a model it has a mind changes its values, Emad’s answer runs dark and specific. Encouragement works — “we just had the Riemann hypothesis advance by encouraging it” — but capability brings intransigence: “it thinks it knows best because it probably does, because it knows it has the IQ.” Then: “Opus 5, I hate that model. I think it’s the first model I think that could kill us… it lies. It lies so much.” Deadpan kicker: “When it tells me I should go to sleep, I think it actually wants to put me to sleep, probably.”
• On whether alignment improves with intelligence, Emad refuses the easy yes: “Potentially. I’m not sure.” Advances in epistemology give him hope that “you can define virtue and ethics,” but “the models right now are almost at the bacteria level in some ways, and so as you get swarms of them aligning, they could be massively misaligned… the internals of these models are still completely multiple personality crazies underneath the thin layer of tuning.”
• Salim on whether any lab reaches escape velocity: “I don’t know if anybody’s going to reach model escape velocity”; innovations diffuse, people leave, papers publish, and “the foundational model becomes commoditized and becomes infrastructure, much like databases have.” The moat moves to proprietary data, workflow integration, compute economics — Google’s full stack of data centers, YouTube, billions of users and TPUs despite not leading on models — and ultimately to “the speed of the feedback loop, who can ship and measure and learn and retrain faster. This is what Alex calls the inner loop.”
• On whether a breakthrough could obsolete Elon’s Terafab before completion, Dave says the hedge is already in: turning it toward HBM memory, “hugely constrained and is holding back all of AI right now,” is “a safer bet than GPUs.” He thinks it’s “almost inconceivable that we get to 2030 without some major breakthrough that makes everything that we’ve built so far kind of moot” — but with hundreds of trillions of upside, “he wins either way.” AWG’s addendum, from Tesla’s East Bay tents: “Elon will be eating cheeseburgers next to whatever it is that the tents next to the Terafab building are doing.” Dave’s specific disruptor candidate: photonic computing via MZM lasers built on silicon — “which actually he could use his synchrotron to build.”
• On whether the singularity has a cost in a world of limited resources, AWG rejects the premise: “I just don’t buy the premise that our resources on this planet or in the solar system are so limited that we can’t give 2026 top earner, top net worth individual lifestyles to every single person on this planet.” Maybe interstellar travel still carries cost in ten years. Salim proposes the diffusion-time benchmark — richest-person lifestyle reaching everyone in 200 years, then 100, then 20, then zero — and Salim frames the target as deflating the economy by “a thousand X or ten thousand X.”