Sam Altman: Singularity Slow-Down, Emad Runs 18 Grokbots, Waymo Slashes Hardware 83% | EP #283

Summary

• Sam Altman’s public walk-back — the singularity is “a rising tide,” not a step function — gets dissected rather than accepted. Altman blames economic inertia (“we’ve all been too ambitious on timelines”), but Emad Mostaque’s counter is that the models simply weren’t good enough until months ago — “the code they were writing was garbage a year ago” — and OpenAI is “trying to find its narrative,” while Dave Blundin reads it as pre-IPO PR: “you can’t listen directly to Sam, Dario anymore. Elon always says exactly what he’s thinking.”
• The panel’s most tradeable through-line: Chinese open-weight models are eating the US closed labs from the bottom up. The new GLM Flash scores 57 on Artificial Analysis versus Fable 5’s 60 at “100 times cheaper”; Kimi Linear cuts context memory 75% with 6x faster decoding on a 1M-token window; and per the FT, Fable 5 has effectively plateaued at $15/M tokens against 14-cent alternatives. Dave’s verdict: “it’s not compromising to save a few pennies… they’re just as good,” and self-improvable inside your own firewall.
• Agent swarms are the product of the moment — but the interface won’t survive. Elon’s GrokBot (launched August 11) gives each bot a dedicated cloud computer; Emad runs 18 in a swarm with control of a MacBook M4 Max and a 5090, Dave signed off on 100K swarm agents, and Salim calls it “a new form of labor.” Alexander Wissner-Gross’s hot take: the messaging-app format is “the Vaudeville era of agents” — “the only better manager for agents is other agents.”
• Gemini 3.7 Flash topping the AA Analyst Agent benchmark (60% vs Opus 5’s 54%) is benchmaxing, not a comeback, per AWG. The metric rewards answering correctly on all five attempts — reliability and determinism optimized for Google Search’s latency needs — while Emad’s harsher read is institutional failure: a frontier-class model now needs only “two to 4,000 TPUs,” the Chinese open-sourced the recipe, and Google, landing three million TPUs this year, still can’t ship it.
• Emad’s contrarian call on Anthropic: don’t IPO. “Anthropic should do a giant fricking raise, like OpenAI did, of 120 billion, and have a straight shot at AGI… stay private like Stripe” — Dario owns 2% and doesn’t care about dilution. Dave’s rebuttal: after Alex Karp “ripped him to shreds” and with VC boards marked up on Anthropic paper, delaying the IPO would be “wussing out” — “this is the stress test for Dario.” The data-retention reversal is read as pre-IPO enterprise appeasement; AWG calls it “largely security theater.”
• NVIDIA’s $6B Poolside deal formalizes the “hackquisition” era and a US open-weight counterattack. Emad frames it as NVIDIA buying up the open-source stack (Nemotron coalition, plus Ashish Vaswani’s Essential AI team) because open models “drive demand for the GPUs more than anything”; Dave explains the structure — the statutory 30-day HSR review “alone is like a lifetime” on AI timelines, so you close the real deal months later, as Elon did with Cursor.
• Waymo cut autonomy hardware from $115K to $20K and unveiled a $75K Zeekr-built robotaxi — by white-labeling China. AWG’s newsflash: “Google, Waymo, Alphabet are switching over to using and OEMing Chinese hardware… I would rather see the West use a Western hardware stack.” The bigger pattern is total verticalization — every mega-cap building chips, models, data centers, robots — with Dave’s kicker: “TSMC is a sitting duck… waiting to be verticalized.”
• The human window is closing and the physical bottlenecks are political. AWG gives humans “at most one or two or three years” of being essential agent-steerers — “work your tail off for three years and then go lie on the beach” — and tells physics PhDs “physics is cooked.” Meanwhile US data-center opposition jumped to 75% (from 43/42 a year ago), which AWG partly attributes to foreign interference and says forces the endgame: “there’s no water in low Earth orbit” — matching Elon’s 10,000 Starship launches/year plan for 100GW of orbital compute.

Deep dive

1. Altman recants: the singularity arrives as a rising tide, not a cliff

• The episode opens on Sam Altman’s on-video admission that he was wrong about AI’s disruptive speed: when GPT-4 shipped in 2023 he expected software and business “up for grabs right away,” but “the economy just has so much inertia. People keep doing the same things… I think that’s actually a positive in many ways… but I think it means we’ve all been too ambitious on timelines.”
• Salim’s frame: this is Stewart Brand’s pace layers colliding — exponential technology hitting linear institutions. “Frontier labs made the mistake of confusing technical possibility with institutional deployment.” His redefinition is the episode’s sharpest line: “the singularity is not when the machine becomes infinitely capable. It’s when institutions can’t adapt at all to that rate of capability… and we’re kind of there now.”
• His lived example: two hours to the São Paulo airport when passenger drones have been technologically ready for a decade — the gap is regulatory and infrastructural, and “that gap is where all the stress is coming from.” The prescription is what Alex calls co-scaling: institutions must scale with the technology, “because it’s not slowing down.”

2. AWG’s dissent: the villain is abstraction layers, and the cure is vertical integration

• AWG agrees at “the superficial layer” — singularity as step function is “totally nonsensical” — but rejects Altman’s premise one layer down: societal inertia isn’t the brake; the abstraction stack is. His analogy: the electric car is a total step function under the hood, but a layer up it’s still a car with a steering wheel. The layering, not the economy, throttles radical progress.
• The theory is prescriptive: “if you want to go faster, pull an Elon and vertically integrate to erase the barriers between abstraction layers.” If OpenAI wants transformative speed, own more of the stack up and down — noting they’ve “pretty publicly abandoned” the original Stargate strategy of owning data centers and now just lease.
• Emad’s alternative diagnosis: the diffusion never happened because the models weren’t ready. “The code they were writing was garbage a year ago… o3 was the first model a year or so ago that I could use… GPT 5.6 Sol is the first really good math model.” Competent intelligence wrapped in familiar interfaces “has literally only been around now for maybe a month or two” — “can you imagine using GPT-4 in a code base?” He also flags the whiplash: a Time piece has Altman claiming AGI by year-end while he simultaneously preaches slowness — “OpenAI is trying to find its narrative right now.”
• Dave’s cynical-but-sourced read: he interviewed Altman “before the singularity kicked off, and then his house got firebombed… with a baby inside, and now there’s a different Sam.” Picketers, armed security, anti-data-center states — the labs realized they’re scaring more people than they’re rallying, so now there’s a PR strategy. “You can’t listen directly to Sam, Dario anymore. Elon always says exactly what he’s thinking.”

3. Past job replacement in five minutes: the $30 trillion question

• AWG’s supply-demand recast of Altman’s complaint: OpenAI and Anthropic have “an oversupply of intelligence,” and much of the market doesn’t have — or doesn’t know how to have — the demand. The curves aren’t clearing at a favorable point.
• Dave’s sci-fi turn: a year ago the AI ambition was taking your job; now “the AI is starting to think, well, if I discover new physics, new medicine… I can add a lot more value than taking away your job.” His insurance example (he chairs a large public insurer): the new categories AI generates — data centers, robots — “are bigger than the legacy insurance industry.” AWG generalizes: grabby-alien plots are inane — “they need our water? Come on” — superintelligence transcends resource-grabbing, and “replacing human labor lasts for about five minutes.”
• The strategic fork AWG calls “the $30 trillion question”: do frontier labs go upstack into applications or downstack into chips, data centers, and energy? He thinks downstack is “more ergonomic” — and notes drily that Anthropic’s stated $30T TAM is “surely pure coincidence” with US GDP.

4. GrokBot: assigning work, not asking questions

• GrokBot — after OpenClaw and Hermes — launched August 11 in beta from xAI: each bot gets a dedicated cloud computer with browser and terminal, you message it “like you’d message a colleague, not a chatbot,” a chief-of-staff agent (Peter’s is “Skippy”) coordinates specialists that only pull you in on judgment calls. Peter calls it “the most genuinely useful consumer AI product that I’ve seen this year.”
• Salim’s take: “you’re making a transition from asking an AI to assigning work” — staff-on-demand taken to its logical extreme, with recruiting time and coordination cost at zero, which means “your optimal organizational structure completely changes. Humans set objectives and leave everything else to the AI.”
• Emad is the power user: 18 GrokBots in a swarm with Tailscale control of a MacBook M4 Max, a 5090, and his subscriptions — including “Atelier,” a bot with a team of artist sub-agents trying to learn art, while others install the new GLM model, test an Alibaba model, and optimize the Alibaba 27B for the 5090 (“added 76% to performance at 64K context”). Dave, in his SwarmIt shirt, just signed off on 100K swarm agents and is re-running “5,000 Kimis” — his lesson is organizational: “you really have to start thinking hard about your org structure to know if your agents are doing anything useful.”

5. AWG’s hot take: this is the Vaudeville era of agents

• The dissent worth keeping: the WhatsApp-style pane of agents “is not the interface of the future.” If fleet size becomes a scaling law like inference-time compute did, “we’re not going to want to ask individuals or even enterprises to manage millions of agents… The only better manager for agents is other agents.” Future observers will see this as “a naive attempt to graft human organizational structures onto humans managing agents” — prescribing “a good dosage of the bitter lesson pill.”
• Salim concedes the interface point (“a UI that’s temporary”) but defends the on-ramp, and Peter closes the loop back to section 2: familiar interfaces are what humans are ready for — which AWG immediately identifies as the abstraction-layer tax itself: “people one or two layers up expect the old thing, so you have to abstract yourself in a familiar interface.” That’s the answer to Sam’s “why are things so slow?”

6. Gemini 3.7 Flash’s benchmark win is benchmaxing, says AWG — and worse, says Emad

• The news: Gemini 3.7 Flash took the top spot on the AA Analyst Agent Benchmark — 60% pass rate vs Claude Opus 5 at 54% and Fable 5 at 49%, completing tasks up to 90% faster and 2.4x faster than GPT 5.6 Tera. Peter’s framing: everyone counted Google out, as they did Meta and xAI, and everyone keeps coming back.
• AWG’s unvarnished case: Google is “still out of the running for the capability frontier.” The benchmark’s metric is share of questions answered correctly on all five attempts — it rewards determinism and penalizes stochasticity, with no time constraint. His conspiracy theory: pressure to embed Gemini into Search one-boxes (fast, low-latency, reliable) over-optimized Flash for “wall clock speed and determinism” — note there’s no Gemini 3.7 Pro anywhere, just Flash. And despite appearances, Google is internally compute-starved: regular meetings apportion flops among GCP, DeepMind, and Search — unlike OpenAI or Anthropic, Google has non-AI constituencies fighting for compute.
• Emad rejects the compute excuse outright: Google is “landing, like, three million TPUs this year,” and a frontier-class model today “needs two to 4,000 TPUs” — “the evidence of that is the Chinese did it, and they open sourced them, and we know exactly how they’re built.” Given Google’s data plus the GLM or Kimi architecture, “you should have a better model on the other side. And if you don’t, you have to ask real questions why.” The new GLM Flash is “10 times cheaper for the same performance”; Gemini 3.5 Pro was previewed but “just couldn’t keep up.”
• Dave names the mechanism: talent attrition plus ego. “I’m the most well-funded top AI engineer in the world, and the Chinese just kicked my ass. I’m gonna go tell my boss, ‘Let’s go download Kimi, do the rational thing, and tune it’? You can’t say that because you look like an idiot, and that’s where they are.”

7. Who leads the US labs? Anthropic on capability, OpenAI on the Pareto frontier

• Asked who’s best positioned now, Dave and AWG initially say Anthropic — strongest generally-available model in Fable 5, and “hordes of very talented people are going there purely ‘cause they wanna see the singularity emerge… like the birth of the phoenix.” But Dave’s caveat is structural: “they’re totally reliant on Elon for the compute. Elon can rip the soul out of Anthropic any day” — and he’s got the Cursor team, acquired for $60 billion.
• Emad disagrees flatly: OpenAI “owns the Pareto frontier,” from Llama now free to everyone to GPT 5.6 Pro — “the only quality math model. I have no idea what magic they’re doing with Opus to actually get math results because it makes so many mistakes.” OpenAI has the $12B round, consumer moving to enterprise, and compute lined up; Anthropic is “shooting themselves in the foot from an institutional perspective because Opus 5 is unpleasant. Sonnet 5 is unpleasant to use.” Peter concedes his own team switched hard problems over to o3.
• AWG’s nuance: the frontier is “one plus dimensional” — Sonnet 5 still beats OpenAI’s latest o3 model on FrontierMath Tier 4 (ironically OpenAI-sponsored) if you’ll pay and wait; if you’re cash- or time-starved, o3 wins on the cost frontier. Salim’s synthesis: the labs face “the innovator’s dilemma from hell” — Chinese open source on one flank, compute constraints on another, regulation on a third — hence the rush into verticals and revenue. Peter’s abundance take, with AWG’s tagline: “When frontier labs compete, you win.”

8. NVIDIA’s $6B Poolside “hackquisition” and the Nemotron open-weight counterattack

• The deal: NVIDIA pours $6B into Poolside (founded by the former GitHub CTO) to build one of the world’s most powerful open-weight models — the US answer to Alibaba, DeepSeek, and Kimi. Emad’s backstory from investors at TechBBQ: Poolside built the Laguna open-source model factory (which beat Thinking Machines’ Inkling at launch), tried and failed to raise $2B for a Blackwell cluster, and got scooped into NVIDIA’s strategy of owning open source “because that drives demand for the GPUs more than anything.” First move, per Emad and not yet announced: hiring Ashish Vaswani’s Essential AI team — an “Attention Is All You Need” author. Expect classic open-source players like Mistral and Cohere to build to the NVIDIA Nemotron reference design.
• AWG’s structural point: hackquisitions — non-exclusive IP licenses plus team lifts — exist “ostensibly to avoid antitrust scrutiny,” but he’s watching the leftovers: “the carcass left over after the hunt… may actually have life to it… could actually be in some sense even more interesting than the part that goes over to NVIDIA.” Emad adds the wrinkle: the remainder includes Poolside Infrastructure Company, building a 1.2-gigawatt data center.
• Dave’s correction from deal flow (Mercor and others calling him for acqui-targets): the FTC is friendly, but “the timeline for AI companies is so short that the statutory thirty-day review alone is like a lifetime” — so you slap together a non-reviewable structure, train the model, and close the real deal months later, “like Elon did with Cursor.” On verticalization generally, AWG half-jokes his way to a real claim: “Does Anthropic get a moon colony? Yeah, probably. Does Anthropic get a pharmaceutical arm? Yeah, already.” His sober version: natural verticalization lives at the infra layer — data centers, energy, satellites, robotics — not necessarily applications.

9. Under siege: Kimi Linear, Fable’s plateau, and 100x cheaper Chinese models

• Three stories in one block: Moonshot AI’s Kimi Linear cuts context memory 75% with 6x faster decoding on a 1M-token window; the FT reports Fable 5 has plateaued as 14-cents-per-million-token Chinese open weights deliver ~80% of the capability of a $15 model; and Anthropic reversed its data-retention policy ahead of its IPO, letting enterprises keep data on their own cloud.
• Emad’s arc of Chinese adaptation: “First we banned the faster silicon, so they built MoE models to take advantage of cheap DRAM. Then DRAM became expensive, so they figured out linear attention and caching.” Now the stealth “o1” model tearing up benchmarks turned out to be GLM, “served entirely on Chinese chips with trillions of tokens a day, these new Huawei chips.” Memory is “50% of all spending now… in a couple of months’ time, they won’t need much memory. That’s how fast they innovate.”
• Dave rejects the “good enough at a discount” frame: “the Chinese models aren’t down a notch. They’re absolutely on the frontier… it’s not compromising to save a few pennies” — and they’re good enough to self-improve inside your company, “a runaway train.” Emad quantifies: GLM Flash scores 57 on Artificial Analysis vs Fable’s 60 — “100 times cheaper.” That changes architecture: “you can afford 5,000 or 10,000 concurrent Chinese operators instead of one Anthropic” — and the panel’s analogy: “maybe I’m not smart enough to ask Fable the right questions, but I’m just about smart enough to ask GPT-5 all the right questions.”
• AWG’s warning about the equilibrium: it’s the Canadian generic-drug problem. US models generate the reasoning traces, Chinese labs distill legally or otherwise, sell it back cheap — and in the last 48 hours, stories of Chinese labs striking rev-share hosting deals with US hyperscalers. “We’re using our own infra against ourselves… a perverse bind.” Emad: “I’m not sure the Canadian drug import analogy holds much longer given what’s going on.”

10. Should Anthropic IPO at all? Emad says no; Dave says Dario can’t back out now

• Dave on the data-retention reversal: this is where “we find out if Dario has what it takes to be a public company CEO.” The old policy — everything routed to Anthropic HQ for 30 days even on Amazon Bedrock — was killing enterprise revenue while corporations fled to Chinese models and secure-environment GPT Sol. “He’s stuck between a rock and a hard place.”
• Emad’s categorical call: “Anthropic should not IPO. If you are in the late stages of AGI now, Anthropic should do a giant fricking raise, like OpenAI did, of 120 billion, and have a straight shot at AGI… stay private like Stripe.” Dario and his seven co-founders own ~2% each, are worth six-seven billion, and have pledged away 90% — dilution is irrelevant.
• Dave’s counter is about signaling: Alex Karp publicly “ripped him to shreds” — don’t trust “an academic, never run anything before in his life” with your IP while he preaches how government should run — so delaying the IPO “plays right into Karp’s hands,” and board VCs have marked up funds on Anthropic valuations they’ve used to raise new vehicles. Salim adds the pricing problem: investors tell him “nobody knows how to price this thing” because Anthropic hasn’t secured long-term compute — which is exactly why everyone is verticalizing: “if you’re at one layer and the bottleneck goes below or above you, you’re screwed.”
• AWG adds the race dynamic: a starting gun fired months ago among SpaceX, OpenAI, and Anthropic, and you don’t want to IPO last into a window that might close. On the retention policy itself: “largely security theater” — but externalizing hosting is genuinely powerful; by some reporting 40% of Anthropic revenue already flows through third-party hyperscalers, and OpenAI is copying the move.

11. “Our economy isn’t worthy”: why the priciest frontier struggles to monetize

• AWG’s deeper read of Fable 5’s plateau: it rhymes with OpenAI’s strategic blunder of “pandering to consumers rather than enterprises” — consumers “didn’t know what to do with all of these shiny OpenAI reasoning tokens,” forcing the painful Codex-era pivot. Now enterprises may be doing the same to Anthropic: “Our economy isn’t worthy. It’s not clever or wealthy or successful enough on average to know how to use Claude 4 properly.”
• The two exits he sees: radical price cuts, or — the exciting one — a new use case in the next year that actually motivates “the nosebleed-priced high end of the frontier,” possibly Claude 4.1, which “may be starting to leak out” depending on the reports you read. On Kimi Linear he’s more measured: known since last fall, but notable as one of the first open successful linearizations of the “infamously quadratic” attention mechanism — the Transformer being replaced “Ship of Theseus style,” piece by piece.

12. Ditto plays Cupid by deleting choice — and AWG hates the body count detector

• The Berkeley startup Ditto: no feed, no swiping — a values questionnaire, then every Wednesday at 7pm a text with a match, place, and time. 160,000 college students signed up, 80,000 dates produced. Salim’s frame: “AI isn’t adding an interface, it’s deleting the interface” — and he sees a bigger pattern of AI as “the trusted intermediary between individuals and overwhelming abundance,” with the Indian matchmaking industry “profoundly about to be disrupted.” Dave would have backed it instantly, but flags the fork: AI managing your choices could be “one of the greatest boons to mental health in world history. If it’s left to manipulate you, it’s gonna be horrible — because it’s such a great salesperson.”
• AWG’s demurral is the section’s best material: “this is why we can’t have nice things.” Ditto’s Body Count Detector uses “478 facial points and 52 microexpressions over five seconds to estimate how many sexual partners a person has had” — “a politely suboptimal use of scarce reasoning tokens… we could be aiming so much higher as a civilization.” His reconciliation with Dave: “save the body count detection until after we’ve solved everything.”
• Emad’s caution lands beyond dating: Black Mirror digital-twin dates aside, “it does take away a little bit from your intrinsic humanity if you outsource your cognition and connection in that way” — he’s already scolding colleagues “starting to slip into trusting the AI too much.” Peter’s defense: with a 50% US divorce rate, matchmaking that outlasts “your initial hormonal response” carries massive societal value.

13. The judgment bottleneck: work harder now, because the window closes in 1–3 years

• The WSJ confirms what the panel feels: agent productivity creates more human work, not less — the bottleneck has shifted from execution to judgment. Salim: “this is Jevons paradox for human cognition… if 10 agents are reporting to a founder, we’ve reinvented middle management — inside your own brain.” The needed breakthroughs are delegation, permission, and escalation thresholds — otherwise “AI works 24/7 and humans are obligated to work 24/7 to keep pace.”
• AWG’s timeline, stated plainly: he’s getting “approximately no sleep” steering agent fleets, and “I don’t think this will continue very much longer. At most, maybe one or two or three years” before AIs are self-steering. His conclusion isn’t rest: “work your tail off for three years and then go lie on the beach.” Dave’s version: “you can master 1,000 AIs, 10,000 AIs, and you’re the most valuable you’ll ever be in human history right now… a year or two from now, they may say, ‘I don’t need your help.’” And the moral kicker: “every year millions of people die needlessly, and if we get three months shaved off that timeline… millions of people will exist forever that otherwise wouldn’t.”
• The PhD fight is real disagreement: Emad — “If you’re thinking about doing a PhD, don’t do one… Every verifiable-domain PhD now is under massive threat.” AWG both agrees — he told a government-funded AI-for-physics center “physics is cooked, and you should probably be reconsidering all of your career trajectories” — and plays skeptic: Oxford, Harvard, MIT alumni “pulling the vertical mobility ladder up behind you.” Emad’s answer: “We didn’t have AI at the time.” He says undergraduate study can still be a social experience; AWG calls it “adult daycare.” Emad also coins the episode’s neologism — “cognithargy,” cognitive lethargy — and describes deliberately turning off his research agents to ship: two books in a year, a big funding round, reading agent output only weekly.

14. Data center backlash hits 75% — narrative beats evidence, and orbit is the endgame

• Heatmap’s polling: opposition to nearby data centers went from a 43/42 split a year ago to 75% opposed, 61% strongly; Bernie Sanders again called for a moratorium. The water numbers Peter cites are damning to the panic: data centers use 627 million gallons/day vs 2 billion for golf courses, 133 billion for power plants, 137 billion for cattle. His fix: hyperscalers should spend the extra 10% — cheaper community energy, school programs, “make them look like cathedrals.”
• AWG fears that wouldn’t help, because the sentiment is being politicized — worst case, stoked by foreign adversaries who “would love nothing more than to slow down America’s data center build-out” — and the US structurally loses this fight because siting is local, while China can centrally declare “the east is responsible for data, the west for compute and energy.”
• Salim’s civilizational diagnosis: “our information systems reward compelling narratives over evidence… nuance has no viral coefficient. We’ve gone from ‘show me the evidence and I’ll form an opinion’ to ‘I have an opinion, now show me the evidence that confirms it’… You cannot run an advanced civilization with this.” Salim’s counterexample: Provocative AI, founded by Rob Fisher, runs a water-negative, carbon-negative data center off Seabrook nuclear power. Emad’s puckish fix: it’s a branding problem — rename them “compute citadels.”
• AWG’s endgame: “Newsflash, there’s no water in low Earth orbit… this is only forcing all of these new data center deployments to sun-synchronous orbit. We might as well just get it over with” — and Elon’s pivot from Mars infra to sun-synchronous orbit and the Dyson swarm was opportunistic tea-leaf reading, not prophecy.

15. Waymo slashes hardware 83% — by white-labeling Chinese hardware

• The numbers: Waymo’s custom 5nm chip (one quadrillion operations/second) helped cut sixth-gen autonomy hardware from $115K to $20K, and the purpose-built Zeekr robotaxi minivan costs $75K vs $200K for the Gen 5 Jaguar — 13 cameras and four lidars vs 29 and five, with heaters and wipers built into the sensor pods. Peter also says NVIDIA cleared Tesla, Uber, and Waymo to operate simultaneously in Las Vegas.
• AWG’s discomfort: “careful what we wish for… Google, Waymo, Alphabet are switching over to using and OEMing Chinese hardware… I would rather see the West use a Western hardware stack rather than just white labeling Chinese hardware.” Underneath: moving away from Broadcom, vertically integrating, “playing footsie with Uber for the moment” but wanting its own distribution — and a Tesla rhyme: the $25K Model 2 never launched because Tesla hit autonomy first, and below some price it makes more sense to sell hosted autonomy than cars. Waymo may ultimately use a white-label hardware arrangement and refocus more heavily on software.
• Dave’s macro observation: this vertical integration is “completely unprecedented in history” — ExxonMobil did oil, IBM did mainframes, GE did reactors and toasters, but now all eleven of the mega-cap cohort are building AI chips, models, and data centers, “colliding into vertically integrated super companies” — robots next. AWG: “meanwhile, TSMC is a sitting duck… waiting to be verticalized… right across the Strait of Taiwan, ready to start World War III at a moment’s notice.”
• Emad’s left-field idea from dinner with a deep-tech VC: instead of retrofitting vehicles, build specialist humanoid robot drivers — a robot that sits in a cockpit is far simpler than one navigating the world, and his bill of materials came to “$6,000 with the actuators and everything. I was like, oh crap, this could actually happen a lot quicker.” Salim responds that it is the first use case for a humanoid robot with two arms and two legs that he has seen, adding, “I will kneel to you, sir.”

16. The broken Waymos theory: China deploys rescue drones while Boston can’t get robotaxis

• The story Peter loves: Chinese autonomous rescue drones — hybrid life preservers flying 30mph vs a lifeguard’s 2mph swim, covering nearly two miles, floating two 80kg adults. Salim: “this is compressing time — response time, where time equals lives.” Dave adds that autonomous response will do more for public acceptance of AI than any chatbot benchmark.
• AWG’s neologism, the “broken Waymos theory” (after Giuliani’s broken windows): “if a city or a nation can’t deploy autonomous robots, then they’re not prepared for the singularity.” He watches China’s drones and shakes his head — “here in Boston, we can’t even get Waymos,” and raising it with Mayor Wu produced “de minimis progress. We’re getting lapped by China at this point.”
• On media asymmetry, Dave says that if one Waymo runs over a cat, “it’ll make every headline in the world,” while if it is ten times safer than human drivers, “we just don’t even report it.” He calls the starved free press “backfiring in the age of AI,” while Chinese tech coverage is “more statistically accurate.” AWG’s caveat stands: only “about topics that are convenient to the government.”
• Emad predicts friction and a strategic squeeze: “you will see lynching of robots” and vandalism here — but China genuinely needs robots for its population pyramid, sees them as abundance, “and I think China will stop exporting robots in five years… why would you export your robots if you can use them to give your citizens a good life?”

17. Space: 10,000 Starship launches a year, a Gulf Coast space corridor, and landlocked spaceports

• Elon’s announced target: 30 Starship launches per day by 2030 — roughly 10,000/year, 40x the entire current global launch rate — the same math Peter says Elon gave the pod: the planned orbital solar-powered AI/data-center build-out requires 100 gigawatts and a million tons of payload, with launches starting in 2028.
• AWG reads the tea leaves of two announcements: Starbase, Louisiana ($100B into the state for a second Starbase) means “the Gulf Coast is becoming America’s space coast” for private launch, while buried in the White House’s Golden Age of Space Transportation report (1,000+ launches/year target) is an executive order to appropriate federal land for spaceports — his candidates: White Sands, the Nevada Test and Training Range, and Arizona’s Goldwater Range. Private Starbases on the Gulf, federal ones in the Southwest.
• Peter’s launch-veteran aside: coastal siting existed to drop stages over water, but with full reusability “you can land in a landlocked area” — landlocked orbital launch becomes much easier, although he notes that suborbital vehicles have already launched from White Sands and Fairbanks. On the viral video of Chinese reusable rockets that are “almost a duplicate of Falcon 9” — same fins, same landing legs, AWG: “same cheers” — SpaceX’s testing-in-public openness is being harvested; AWG suggests Elon might revisit his no-patent-enforcement policy, though Peter says Elon wants as much launch capacity as possible. Salim’s optimism: “what you’ve got in SpaceX is a compounding learning loop, and that’s hard to beat.” Also noted: Starlink is sweeping airline adoption — passengers now choose Starlink-enabled carriers — which Peter says will crush Viasat, Eutelsat, and Gogo.

18. AMA lightning round: Intel’s fab gambit, regrowing teeth, and who controls superintelligence

• On why Intel isn’t just cloning old NVIDIA designs, Dave has the inside answer from a senior Intel exec: Lip-Bu is making “an ungodly fortune on Xeons” and burning nearly $2B a quarter — plus a fresh $20B raise — building fab capacity to compete with TSMC as a general-purpose foundry. Competing with NVIDIA on GPUs now would poison the well: you can’t fab for customers you’re competing against.
• On AI mind viruses, Salim’s answer: intelligence doesn’t guarantee epistemic awareness, and AI’s danger is replication speed — “one bad belief can propagate through millions of agents almost instantly,” since billions of agents descend from the same model; the defense is cognitive diversity, agents challenging agents, and “defensive co-scaling.” On dentistry: Emad points to AI-designed ligands for enamel and ameloblast activation, while AWG cites TRH-035, a Kyoto University spinoff drug targeting honest-to-goodness tooth regrowth — infant clinical trials first, general availability by 2030, “not much AI involved.”
• On Dario’s super-voting shares and unelected control, Emad doesn’t soften it: “You’re not gonna have a say in superintelligence… Anthropic think that’s far too dangerous, and there is no good democratic way to do that under their rubric” — it “ultimately comes down to a few people, like Ben Bernanke, to decide the future of the light cone, potentially.”
• On whether non-AI R&D gets starved: AWG says no — “the self-licking ice cream cone of recursive self-improvement can only get us so far… it’s going to be the non-AI R&D applications that ultimately dominate the economic gain.” Dave’s answer to “what would you do with 100X capability tonight”: self-improving algorithms first, then, per Demis Hassabis, turn all of it on health and longevity “until we get it solved.”