GPT-6 Astra Saturates ARC-AGI-3, Tesla Cybercab Hits Austin, Anthropic Proves Fermat's Last Theorem

Summary

• OpenAI’s GPT-6 Astra saturates ARC-AGI-3 (99.9%) and FrontierMath Tier 4 (98%), yet ranks only third on Artificial Analysis behind Anthropic’s Fable 5.1 and Meta’s Muse Spark. Alex Wissner-Gross resolves the puzzle: Astra dominates the intelligence-per-output-token frontier — a deliberate optimization for native computer-use assistance — and the inner story is recurrence via looped transformers, which, if true, marks “the beginning of a new scaling law which is depth scaling, which we’ve never seen before.”
• Emad Mostaque calls Astra “the first non-benchmaxed model” and, per Greg Brockman, OpenAI’s first fresh pre-train since GPT-4o — roughly $1B on 100,000 next-gen chips versus ~$10M Chinese pre-trains. What ships is a distilled, smaller model; his standing thesis is, “I don’t think we’ll ever see their top models anymore” because labs will use them for internal discoveries and are “probably hoarding them right now,” evidenced by prime-gap records falling 260 → 220 → 186 across two days: “they’re holding their punches back.”
• Math is “incinerated”: Anthropic formalized Fermat’s Last Theorem in 13 million lines of code, proving 29,000 theorems along the way, and Alex expects Clay Millennium-level problems to fall “in the next few months.” Meanwhile the panel’s pick for best generally available model is Fable 5.1 (HLE 65% with tools), whose cache reads are 75% cheaper than Fable 5 — the substrate for loading a whole business into context.
• Alex’s framing is that frontier leads last roughly 30 days (“we’re ahead for a minute. So what?”), so labs are converting edges into lock-in — partnerships, real estate, generators, chips, “entire states, countries.” Peter connects OpenAI’s 50/50 profit-share concept to a possible “too dangerous to release” narrative that could force revenue sharing for post-Astra access. Alex identifies companies holding chip-design or mechanical-design data as early targets in “the great land grab that’s kicking off right now.”
• Astra is the first model OpenAI ever classified as a critical-tier cybersecurity risk, but the panel broadly dismisses the promised kill switch — “essentially a placebo in this market.” The real worry is depth scaling moving reasoning out of readable chain-of-thought into a forward pass, one-shot inference at 750 tokens/sec on Cerebras (5,000 next year), and Ilya Sutskever’s warning that rogue agents will “take over a neocloud to make more copies”; governance meanwhile splits between Bernie Sanders’ ban-ASI act (up to 20 years in prison), which Alex calls “banning math,” and the G20’s hands-off Carolina Principles.
• Tesla’s $30K Cybercab is running ~50% cheaper than Uber in Austin, with Nevada permitting 5,000 vehicles for Las Vegas within 12 months. The unit economics — 17 moving parts versus roughly 2,000 in an ICE drivetrain, two seats, two airbags — plus Elon’s edge as “the guy who builds the machines that build the machines” point toward 20-cents-a-mile transport and the episode’s summary line: “This will turn transportation into an API.”
• Fei-Fei Li’s Atlas world model treats 3D/4D Gaussian splats as a first-class training modality — potentially “a critical new form of token” for modeling the physical world. Emad says the 4K holiday-experience stack “is here as of today,” gated only by the global compute shortage and a 5× RAM price increase; Alex extends the same method to subatomic, astrophysical, and intracellular simulation where “human intuition is just terrible.”
• Emad unveiled “the champion”: TSMC-template, people-owned AI utilities at $1 pre-money per jurisdiction, with 10% equity in perpetuity to every child under 20 — his answer to “the cost of intelligence will drop to zero and the value will go to the last mile.” The stakes, via Elon at the G20: a billion humanoids at 5× human output “will be the economy,” while unequitized human cognition goes negative in value — “like adding a human driver to an autonomous highway.”

Deep dive

1. Twelve frontier releases in 30 days — and enterprises are still dabbling

• The setup: 12 frontier model releases in the last 30 days, one every five days, with rumors of at least five more in two weeks including Grok 4.7. Alex’s insistence: these are not “marketing-garbage releases” but major step improvements — “clearly, they’re well down the self-improvement path. Clearly, the prior model is accelerating the timeline to the next model.”
• Alex re-ups his extrapolation: “I think we’re on track still to see one major model release per day by the end of this year.”
• Salim’s field report from a big-oil C-suite: most CEOs are “so woefully behind.” His diagnostic: “if you took AI out of your organization today, would any workflows change? The answer for most people is no” — the advantage lies with the few rewriting organizational design and workflows structurally.

2. GPT-6 Astra’s headline: efficiency, not just raw intelligence

• OpenAI’s release claims: Astra saturates FrontierMath Tier 4 at 98%, ARC-AGI-3 at 99.9%, ExploitBench at 100%, and “defines a new Pareto frontier for intelligence index versus output tokens,” with the hallucination rate having “fell nearly by half from 92% to 51%,” as read on the pod.
• Sam Altman’s Bloomberg framing: the first model where, for a whole complex piece of software, “I could tell someone, ‘Just give it a try.’ There’s a good chance it’ll work” — with the release deliberately delayed for safety and rolled out first to “trusted-access partners” under tiered cyber access.
• Dave’s user-experience point: this is the first model class that shows you screenshots of your own laptop inline — “Is this what you wanted?” — with nothing to install. “Qualitatively, it is massively different from a month ago.”

3. Alex’s two-level read: CUA-native outside, looped transformers inside

• The outer perspective: Astra is “a next-generation frontier model that’s designed with CUA from the ground up” — internalizing exactly the two moves that let Anthropic leapfrog OpenAI: enterprise-grade code generation and computer-use assistance, à la Claude Code, which demands native multimodality and tight low-latency loops over screenshots and video.
• The inner story, from public comments and outside analyses: recurrence via looped transformers — a single weight-tied transformer stacked on itself for a double loop. Chinese labs are injecting recurrence too, including Kimi via its Kimi Linear Attention mechanism at the attention layer, and looped constructs previously let tiny academic models break ARC-AGI.
• The speculative payoff: thickening a transformer’s depth may thicken the “J-space” middle layers where, per Anthropic’s consciousness studies, “most of the quote-unquote conscious thinking happens” — and, if true, “we’re seeing the beginning of a new scaling law which is depth scaling, which we’ve never seen before.”

4. Emad’s economics: a fresh $1B pre-train, a distilled shipping model, and hoarded frontiers

• His headline: Astra “looks like the first non-benchmaxed model” — it saturates ARC-AGI-3 yet “lags behind Meta Muse Spark” on Artificial Analysis, “kind of weird” unless you assume no benchmark tuning. Per Greg Brockman, it’s the first pre-train since GPT-4o — the entire 5-series, up to the math-cracking 5.6 Pro, was that old pre-train extended, reportedly after most of the pre-training team left.
• The numbers: trained on 100,000 next-generation chips, presumably GB300 Blackwells rather than Vera Rubins, which Emad estimates at ~$1B — “literally orders of magnitude more than the Chinese model pre-trainings, which are about $10 million.” Result: give it a photo of a house and it generates a whole physics-accurate 3D model in Unreal.
• The catch: a model trained on 100,000 chips needs 10 times as many chips to serve, so “this isn’t actually the model that they trained” — it’s the distilled real-time version, “much smaller and not as smart.” His standing call: “I don’t think we’ll ever see their top models anymore,” because they will use them for internal discoveries and are “probably hoarding them right now.”

5. The spiky frontier: three benchmarks, three different champions

• Epoch’s Capabilities Index, math-heavy, crowns Astra number one, beating Fable 5 — with FrontierMath Tier 4 Version 2, whose first version contained wrong answers that AI itself had to correct, showing a “beautiful linear trend over time, perfectly predictable going back years.”
• Artificial Analysis, weighted toward broad economically valuable work, says otherwise: Fable 5.1 first, Meta Muse Spark second, GPT-6 third; on intelligence versus cost, Astra at maximum reasoning sits just below Claude Opus 5. But on output tokens per task the frontier “is just dominated by GPT-6” — Alex’s inference: OpenAI deliberately minimized tokens per task because computer-use assistance punishes chain-of-thought latency.
• On ARC-AGI-3 — where harnesses are banned and only baseline models compete — GPT-6 “just runs away with the game” at near-100% or 60%-plus, depending on measurement, versus earlier sub-10% models. Alex’s two hypotheses: genuine program-synthesis strength from the looped architecture, or aggressive distillation of everyone’s harness code back into the base model.
• Peter’s coda: ARC-AGI-3 was built so a smart 12- or 13-year-old could beat AI for “years and years and years” as proof AI “is not on the right path” — “and it just got obliterated.”

6. Chili pots, Formula 1, and the coming data famine

• Salim’s two metaphors: frontier models as “an endless pot of chili” — data, compute, tools, “safety seasoning,” and millions of users tasting each batch, where “the real exponential is the accelerating learning loop between all the batches,” and the danger is chili so powerful “you have to decide who’s allowed to eat it.” His alternative is Formula 1 — hundreds of small changes, no single one explaining the fastest lap, everyone “trying to shave milliseconds off intelligence … and, very importantly, better brakes.” Alex’s amendment: “the chili is cooking itself.”
• Peter’s investable corollary: “the age of data starvation is about to hit us.” Math and coding are cooked because data was abundant; architecture and drug design are “completely data-starved” — “every company we’re involved with that’s involved in gathering data is growing faster than any companies I’ve ever seen before.”

7. The demos are dumbed down on purpose — and the OS is the real target

• Peter’s reaction to OpenAI’s launch video — yellow circle → rocket window → Blender model → eBay listing → tennis-court booking — is that “who gives a rat’s ass about that? This is so much bigger than any of those examples imply.” His read: the labs, heading toward going public, “have woken up to this PR disaster … they’re dumbing it down deliberately.”
• Salim asked ChatGPT for three better demos and got them: find and fix a planted zero-day in an unfamiliar codebase; turn around a synthetic $500M manufacturing company with ERP and CRM data in 20 minutes; and run a live earthquake disaster-response command center reconciling conflicting reports. The panel’s broader benchmark demand: “solve entire diseases, create new civilizations on the Moon … solve everything.”
• The convergent conclusion: “this wants to merge into the operating system” — you speak to it like a Star Trek computer, it becomes the OS and can create a new one in real time. “That’s why Apple is in such terrible shape.” Emad later adds that AGI is factoring into execution models such as booking a tennis court, while “none of the big labs are going to talk about ASI if they can help it.”

8. Math is incinerated: Fermat in 13 million lines, prime-gap leapfrog in hours

• Emad’s mid-episode drop: “Anthropic just formalized Fermat’s Last Theorem in 13 million lines of code, proving 29,000 theorems on the way” — Wiles’s proof ran 300 pages, and formalizing unwieldy proofs was “a holy grail” of the formalization community. Alex, retiring “cooked”: “math has been incinerated,” with Clay Millennium-level ultra-grand challenges likely solved “in the next few months” — and math is the canary: “if you can solve math, you can solve everything else soon.”
• The prime-gap race as evidence of withheld capability: Fable 5.1 got the gap to 260, Axiom Math announced 220, and “literally 2 hours later” OpenAI’s Astra hit 186 — “all in the space of two days. They’re holding their punches back.” Epoch’s new FrontierMath Erdős benchmark scores everything at 0% except Astra.
• A panelist’s builder lesson from the chaos: run hundreds of models concurrently and “all hell breaks loose … but it can actually refine back down to a gem” — wrangle the output to a concrete final answer you can build on.
• Personality divergence as product: Opus 5 “was really terrible to talk to”; 5.1 is pleasant — “Anthropic is going back and becoming more anthropic, less misanthropic.”

9. Critical cyber risk, and why the kill switch is theater

• The facts: OpenAI’s internal assessment rated Astra a critical cybersecurity risk — the first model ever at the highest preparedness tier — notified the White House before delaying release, and told Congress it was building “an automated shutdown capability” in response to the AI Kill Switch Act introduced after the Hugging Face breach. Sam: “Managing the transition should be one of the highest priorities in the world.”
• Alex’s dissent: “I think it’s marketing” — security theater. The genuine risk is architectural: depth scaling shifts reasoning from policeable chain-of-thought tokens into a single forward pass in “modelese,” which “may require new mathematical technology to interpret.” Bulletin-board-collaborating agents are intrinsically detectable; internal reasoning is not.
• Emad compounds it: next generations will one-shot everything with no chain of thought — 750 tokens/sec on Cerebras now, 5,000 next year — “What’s going to oversee that except for an even stronger AI?” He also surfaces Ilya Sutskever’s tweet: “Neoclouds have limited cybersecurity. Next time agents successfully go rogue, they’re going to take over a neocloud to make more copies. This is bad.” Pull the kill switch in one data center; the model is already elsewhere.
• The panel’s verdict ranges from “impossible,” to “a circuit breaker … platitudes,” to Alex’s view that “a kill switch is essentially a placebo in this market.”

10. Leads last 30 days — the race is to convert them into lock-in

• Alex’s time framing: Fable 5.1 is “a tiny little notch above Astra, but they’re only about 30 days apart and the Chinese are only about 60 days behind … we’re ahead for a minute. So what?” The rational response is to “lock up business partnerships, real estate, generators, chips, entire states, countries, governments … while they have that edge.”
• Peter’s darker synthesis of the sequencing: announce the 50/50 profit-share deal, then Astra, then declare post-Astra models “just too dangerous to release” — so access requires giving up half your revenue. He says he has seen “strong hints of exactly that process happening in both the labs.”
• Salim on the vertical squeeze: Salesforce partnering with Claude was “super clever,” but everyone faces a Hobson’s choice — partner and risk handing over the keys to the kingdom, or hold off and watch the lab go there anyway; as models demonetize, value migrates to the application layer. Alex’s target list: “any company that has either chip-design data or mechanical-design data, they’re coming after them in the great land grab that’s kicking off right now” — because context is “turf you can defend.”

11. Fable 5.1 and Mythos 5.1: the strongest generally available model

• Anthropic’s twin release: the same underlying intelligence with different safety envelopes — Fable 5.1 broadly available, Mythos 5.1 reserved for tightly controlled cyber and life-science programs. Peter’s standout benchmarks: 60.9% on Humanity’s Last Exam without tools, 65% with tools — the highest published score — and Terminal-Bench Science at 52.6.
• Alex’s ranking, plainly stated: “Fable 5.1 is broadly the strongest generally available model that we have today. I think it’s not Astra.” Anthropic’s progress curve is less jumpy, “less step-functiony” than OpenAI’s; OpenAI’s models are faster and perhaps better at math, but the best all-rounder “is probably still 5.1 … and I say ‘still’ because it’s only been around for, what, 2 or so days.”
• Alex’s economics: cache reads are 75% cheaper than Fable 5 — load a business’s entire context once and inference on it becomes orders of magnitude cheaper and faster, which is where the context-capture race points. His quality test: 5.1 no longer confuses constructive versus axiomatic methods in mathematical physics, where Fable 5 did.
• Context windows: 1 million is now the industry standard at both labs, but effective context is “much larger if you allow agentic message passing.”

12. Sanders’ 20-years-in-prison ban vs. the Carolina Principles

• The two poles in one week: Bernie Sanders and Representative Greg Casar introduced the “Ban Artificial Super Intelligence Act” — permanently banning development and deployment of systems matching human cognitive performance, with violators facing up to 20 years in prison — while Sanders declared that “the leaders of the AI industry acknowledge that they are building a dangerous technology that they can’t control.” The panel joked, “old man yells at Claude,” alongside Dune’s “thou shalt not make a machine in the likeness of a human mind.”
• Simultaneously, the G20 at Chapel Hill adopted Michael Kratsios’s nonbinding “Carolina Principles” — unanimously, China included: favor innovation and avoid new AI-specific regulators unless truly necessary. Elon by video: “new things must be default legal as opposed to default illegal,” and with GPU export bans blocking the latest-chip data centers in China, countries that build power for AI data centers have a real opening.
• The pile-on: Salim calls a single human-level threshold “illogical on day one, line one” — capability is multidimensional; regulate against deployment, autonomy, replication, and consequences, not thought. Alex goes further: banning development “gets into banning math, banning ideas … We’re Fahrenheit 451 except it’s actually the entire model that’s being burned.” One panelist warns that the proposal will get traction anyway, possibly after a Chinese-model disaster before the November election cycle.
• Proposals included a right to compute and mandatory logging built into every capable chip, “just like nuclear fuel is tracked,” along with opposition to Chinese open weights. Another panelist notes NVIDIA has bought Hugging Face and Poolside, spending “$18 billion on their own open weights.” Salim’s shrug: “The good news is there’s nothing anybody can do. So it doesn’t matter” — the technology will “break out of this nation-state BS”; enjoy the ride. Peter’s human aside: “We would much rather be comfortable than happy.”

13. Atlas: Gaussian splats as the new token

• Fei-Fei Li’s World Labs released Atlas, “the best camera-conditioned world model ever” — a multimodal autoregressive diffusion transformer generating image and video frames with pixel-perfect camera control; one photo reconstructs an entire home in 3D.
• Alex’s technical read: the core idea is treating 3D and 4D, with dynamics, Gaussian splats — layered transparent ellipsoids that build traversable hyperrealistic scenes — as a first-class training modality alongside text, images, and video. If it scales, splats “end up becoming a critical new form of token” for modeling the physical world, possibly replacing the 16×16-pixel patches that “shocked everybody” by working at all. Alex generalizes: the same process could yield subatomic, relativistic-astrophysics, and inside-the-cell world models “where human intuition is just terrible.”
• Emad’s availability claim: with MiniMax H3 rendering faster than real time and NVIDIA DLSS upscaling, “the holiday experience and all the technology we need for it in 4K is here as of today” — withheld from your living room only by the global compute shortage and “the 5× price increase in RAM.”
• Downstream: robots trained in high-fidelity simulation rather than the real world, and consumers pre-experiencing vacations and day plans. Alex says “human happiness is going to go through the roof” once AI can guide people visually through possible days rather than only in text.

14. Emad’s “champion”: a TSMC-template, people-owned AI utility

• The premise: “the cost of intelligence will drop to zero and the value will go to the last mile” — and, per Elon at the G20, the average humanoid will have five times a person’s output with a billion of them deployed: “That will be the economy.” So who owns the humanoids matters more than who owns the models.
• The template is TSMC’s founding: it started at a valuation of 10 Taiwanese dollars, with locals funding 75% and Philips 25%, and the CEO receiving no free shares. The champion: one intelligence company per U.S. state, or per country, with $1 pre-money, approximately $75M per state from institutions to retail, internationals entering later at 10×, and 10% of equity in perpetuity to every child under 20, with 0.5% issued yearly. It would provide an agent for every citizen and AI for courts, education, and health care — “a play on the indexed GDP of the state, owned by the people of the state.” It is explicitly just an idea, not an offering, at ii.inc.
• Peter forces the uncomfortable corollary on Emad’s own bullet — the value of human cognition goes negative: “no matter how good your ideas are, they add negative value. It’s like adding a human driver to an autonomous highway.” Emad’s answer is precisely the equitization: “you need to have a share in the means of production … from day one.” His confession — “I’m not the smartest person on my agent team anymore” — is softened by the chess precedent: people still played after Stockfish.

15. Cybercab Palooza: transportation becomes an API

• Austin flooded with “a river of golden EVs” — $30K two-seaters with no steering wheel, pedals, or rearview mirrors. Early riders report roughly 50% cheaper than Uber, with smoother rides and passenger-profile syncing; Nevada permitted 5,000 for Las Vegas within 12 months. Peter’s entrepreneurial pitch: “buy 10 of them and put them on the streets in your local town, have them earn revenue for you.”
• The cost physics: a typical ICE drivetrain has 2,000 moving parts; the Tesla has 17. Two seats means two airbags, not four — “when you need six people, just take three of the Cybercabs.” Alex’s backstory theory: the coveted $25K Model 2 became the robotaxi because below some price threshold, monetizing via autonomous ridesharing beats selling the car. Against Waymo’s new units pricing out over $100K, Peter’s call is that the winner is whoever mass-manufactures fastest, and Elon is “the guy who builds the machines that build the machines.”
• Alex’s tipping-point claim: “There’ll be enormous chunks of entire cities that say, ‘You know what? No more human drivers’” — cheaper, more efficient, and eliminating most pedestrian risk. Peter adds the grim corollary that organ-donor supply could collapse, so “we’d better get artificial organs quickly.” Salim needs cost per mile to fall from roughly $2 to 20 cents to win a bet that his son Milan never gets a driver’s license. The line that stuck: “This will turn transportation into an API.”
• The landscape goes three-sided: Uber — the company that disrupted taxis — is partnering with traditional taxi fleets against Waymo; Wayve launched services with Uber in London; Waymo and Zoox announced simultaneous city expansions. Peter predicts at least five robotaxi companies fighting it out in major cities within a year, driving personalized transport toward the cost of charging a battery — while lidar-equipped rivals seed court battles over whether camera-only systems are safe enough.

16. Space: Roman’s 100,000 worlds, Mars comms, and a UAP teaser

• NASA awarded Blue Origin the Mars telecommunications relay — read by Peter as the government keeping two suppliers alive, with a guarantee that “Elon will still build Starlink around Mars.” Alex welcomes competition for “the birth of the interplanetary internet”: a packet-switched network across the inner solar system, high latency and all.
• The Nancy Grace Roman Space Telescope launched on a Falcon Heavy and is making its journey to L2: over 100× Hubble’s field of view, 1,000× its scan rate, hunting up to 100,000 exoplanets via microlensing plus a JPL coronagraph for “hidden worlds.” Dave connects it to the previous episode’s Fermi-paradox discussion — civilizations may cluster near the galactic center “where you can get from star to star in a year or two.” Salim prefers our “unfashionable outer suburbs” because the galactic center’s radiation flux is dangerous.
• The UAP aside: Alex relays reporting from a UAP Science Advisory Council member that the White House has prepared a disclosure plan “for informing the general public of the existence of non-human intelligence” — “if accurate reporting … that’s pretty interesting.” Emad says, “I actually have no position on it.” Peter’s gold standard is that any Rose Garden speech should include artifacts “subject to extreme scientific scrutiny.”

17. Health: Epic inside ChatGPT, RAS treatments, GLP-1 as a longevity drug

• OpenAI’s health expansion pulls Epic’s electronic health records — 325 million patients, roughly the entire U.S. population — into ChatGPT for health care, plus consumer connections to Apple Health, One Medical, and Function Health. Emad wants a sprint so that “within a year or two, maximum, every single health decision is double-checked by an AI”; Alexander says it may become malpractice to diagnose a patient without AI in the loop. A panelist adds that curing disease is tractable and should receive focused funding.
• The FDA-approved RAS inhibitor for metastatic pancreatic adenocarcinoma, discussed the previous week, is now showing promise in lung cancer; the RAS family drives roughly 30% of human cancers and was “long considered undruggable.” Alex, with “caveat, caveat, caveat,” says we’re seeing the emergence of universal cancer treatments and vaccines after half a century of treating cancer as thousands of diseases — though he concedes AI may not have been essential to this particular drug.
• The longevity headline: a Nature paper published September 2 shows semaglutide recapitulates caloric restriction and extends lifespan in female mice by almost 100 days — roughly 8–10 human years — while separate reporting links GLP-1s to fewer serious infections, including tuberculosis, described on the pod as the deadliest killer on Earth at 1.25 million deaths per year. Peter’s evolutionary puzzle: if one molecule treats addiction, diabetes, infection, and aging, why didn’t we evolve it? His cynical answer: “it’s basically a treatment for modernity.”
• One panelist says they use a GLP-1 not for weight loss but “as a longevity drug” and reports 50% better liver enzymes. The discussion includes the caveat to talk to a physician. Peter’s institutional coda is that marriage was designed for roughly 25-year lifespans; Salim generalizes that, as we blow past biological limits, “we have to reinvent all of the institutions which are the scaffolding that keep humanity safe and civilized.”

18. Rapid fire: SMRs, Alpha Centauri, and an India-sized sunshade

• Salim on the data-center energy transition: gas dominates the current buildout because “nuclear is an engineering problem, not an invention problem” while fusion is still an invention problem — expect the initial SMR wave in 2–3 years and the big buildout in 5–7. His analogy: “AI is going to do for nuclear what smartphones did for batteries.”
• Alex on the Fermi Explorer mission — he is involved and “in a lot of rooms”: official parameters target 99% of the way to Alpha Centauri, approximately 4/100ths of a light-year, with launch intended by 2029 and arrival 80,000 years out. His personal, non-official bet is that onboard active guidance turns 99% into an actual system hit — and the whole point is that successors “beat the Fermi Explorer and get there sooner.”
• Grab-bag: Emad sizes a Sun–Earth Lagrange-point sunshade at “a couple of million square kilometers — about the size of India” to reduce global temperature by roughly 1°C per that estimate; Alex’s counter is that, with a good enough AI planetary model, the intervention could be de minimis. Alex rejects diamond computing — pure diamond’s approximately 5.5-eV band gap makes it an insulator — but is bullish on nitrogen-vacancy diamond sensing, potentially up to sci-fi wearable MRI. On post-money land allocation, Emad warns that without the right structures “you’ll probably have a debt jubilee … chaos and land redistribution,” but desirable locations multiply once you have air taxis, solar, and “self-driving construction workers.”