Summary

SpaceX is preparing what could be the largest IPO ever: $135 a share, $75 billion raised, and a $1.75 trillion-to-$2 trillion valuation. The investable package is a “Frankenstein” of reusable rockets and fast-growing Starlink alongside xAI and X—what Casey Newton calls “two amazing businesses” and “two terrible businesses” bundled into one stock.
Anthropic’s leap from roughly $1 billion in annualized revenue in January 2025 to a prospective trillion-dollar-plus IPO shows how violently AI economics have accelerated. OpenAI may file soon too, potentially minting hundreds or thousands of new millionaires in San Francisco and turning status anxiety, housing scarcity, and inequality into immediate second-order effects.
The Anthropic listing could create an enormous philanthropic windfall. Its eight co-founders pledged at least 80% of their wealth to charity, while employee equity pledges were matched share-for-share—and sometimes three-to-one—potentially producing “something bigger than the Gates Foundation every year” for causes including global health, pandemic prevention, AI safety, and, memorably, shrimp welfare.
Public ownership creates conflicting pressures on AI safety rather than a clean verdict. Newton worries that shareholders and activist investors will push dangerous models toward release despite public-benefit-corporation protections; Roose counters that liability and public reporting could also create positive pressure, while Newton points to disclosures and possible shareholder votes as new levers.
AI has moved from winning elite high-school math contests to producing research that mathematicians consider publication-grade. After Google DeepMind, OpenAI, and Harmonic reached International Math Olympiad gold-medal performance, OpenAI reported a solution to the unit-distance conjecture using methods judged sophisticated and surprising enough for a top journal—evidence that AI can now do “absolutely top-tier research.”
The Leiden Declaration is less a rejection of AI than a fight over who sets mathematics’ rules, incentives, and purpose. Roughly 800 signatories want disclosure and quality controls, but Kevin Hartnett says the deeper fear is that genuinely good machine proofs could turn mathematicians into hobbyists; Terence Tao represents the middle position, treating AI as a “jetpack for your thoughts” while still signing the declaration.
The week’s smaller stories all expose systems whose incentives or controls fail under trivial pressure. A voluntary 30-day federal AI review, claims that a support bot surrendered prominent Instagram accounts, robots allegedly trained in an Airbnb, and prediction-market bets on events insiders control point toward what Newton calls a “low trust society”—with a Bluetooth speaker named “Bomb” forcing an airplane to turn around as the punch line.

Deep dive

1. SpaceX is selling a space moat bundled with Musk’s weakest assets

SpaceX could list as soon as the following week at $135 a share, raising $75 billion and valuing the company between $1.75 trillion and $2 trillion. Roose’s reaction: these are numbers “we have not seen before in the history of capitalism.”
The security is not a pure SpaceX bet: reusable rockets, Starlink, xAI, and the X social network now sit inside what the hosts alternately call a “Frankenstein” and a collection of “Voltron-like companies.”
Newton sees a formidable rocket moat—few credible competitors, with Blue Origin recently losing a rocket on the launchpad—and Starlink “growing like wildfire.” His problem is the forced pairing with “two terrible businesses called xAI and X”; Newton speculates that Musk needed somewhere to “hide his losses.”
xAI is now renting compute originally built for itself to Anthropic, giving the conglomerate an AI neocloud angle. Roose’s own conversion came aboard a Starlink-equipped plane delivering 200-plus megabits; at least with United, passengers receive it free, making the product feel like “a free miracle.”

2. Anthropic’s hypergrowth will remake San Francisco before it remakes markets

Anthropic filed a confidential S-1 and is expected to seek a valuation above $1 trillion; OpenAI may file soon. The hosts contrast that prospect with Anthropic’s roughly $1 billion annualized revenue run rate in January 2025.
Roose remembers visiting in 2023 when the small, earnest, safety-focused company appeared “ambivalent about making money” and perhaps actively resistant to it. Three years later, Newton says, “boy howdy did they make a product” and discover an appetite for revenue.
Two-and-a-half San Francisco companies—counting SpaceX’s distributed footprint—could mint hundreds or thousands of millionaires, decamillionaires, and centimillionaires. Even workers earning mid-six figures now wonder whether missing OpenAI or Anthropic means they missed their future, replacing the city’s old abundance story with scarcity.
Newton had predicted that 2026 would be the last year to buy a San Francisco house. With homes reportedly selling for multiples of asking and some sellers requesting Anthropic or OpenAI equity instead of cash, Roose’s response is: “Shoot your shot.”

3. Anthropic’s equity may unleash a third wave of philanthropy

All eight Anthropic co-founders pledged at least 80% of their wealth to charity, potentially earmarking hundreds of billions of dollars if the company reaches the contemplated valuation.
Anthropic also offered employee equity-donation matching share-for-share and, for some early employees, three-to-one. The resulting capital could be “something bigger than the Gates Foundation every year” for several years, while existing philanthropic infrastructure may be too thin to absorb it.
Likely beneficiaries include global health, pandemic prevention, AI safety, and effective-altruist causes outsiders may find strange. Newton captures the incongruity with the running joke that “it’s a great year to be a shrimp,” while Roose notes the irony of billionaires rebuilding functions cut from the social safety net.

4. Going public widens AI ownership—and the pressure to accelerate

Newton’s central worry is structural: OpenAI and Anthropic were founded by people skeptical that ordinary for-profit corporations could develop advanced AI safely. Public markets add index holders, retirees, activist investors, and quarterly expectations to an already difficult decision about withholding a dangerous model.
Public-benefit-corporation status lets the labs weigh social commitments, but Newton says it does not eliminate shareholder suits for breaching fiduciary duty. “When the rubber meets the road,” he argues, the labs will still have markets “breathing down their neck.”
Roose’s counterweight is liability: releasing a model capable of creating a bioweapon could itself provoke shareholder securities claims. Roose also sees public reporting as a possible source of democratic oversight; Newton adds that public companies must disclose financials and changes, and shareholders may vote on some matters—more levers than citizens currently possess beyond opposing local data centers.
The Nasdaq 100 and S&P have relaxed or considered relaxing seasoning periods that once kept fresh IPOs out for three months, six months, or a year. Roose says broader access through index funds could help distribute the AI boom; Newton calls an IPO a small but “necessary” response to wealth and power concentrating in a handful of private companies.

5. AI cleared math’s contest benchmarks and reached the research frontier

Google DeepMind earned an International Math Olympiad silver-level score in 2024; the following year, DeepMind, OpenAI, and Harmonic reached gold. Yet Hartnett calls even the hardest high-school mathematics “like 0% of the way to the frontier” of research.
Labs pursued math partly as research for its own sake and partly because reasoning might transfer commercially. Hartnett invokes the schoolteacher’s explanation for learning math: not merely balancing a checkbook, but “to teach you how to think.”
When ChatGPT launched in November 2022, mathematicians circulated failures such as claiming finitely many primes or that two plus two equals five. Hartnett credits both reinforcement learning on math problems and the broader improvement users have observed across models.
After the IMO, labs moved through the Putnam exam and into Paul Erdős’s roughly 1,200 open problems, carrying rewards from $20 to $500. Erdős—“essentially the Bob Dylan of math”—lived on the road and mathematicians’ couches, then died at 83 during a conference.

6. The unit-distance result broke through the “toy problem” defense

Most Erdős problems attracted little attention because unsolved does not mean important. Hartnett compares many to “the Wordle of math”: sophisticated riddles whose solution methods are unlikely to remake the field.
The unit-distance conjecture was different. Humans had seriously attacked it, the machine’s methods were sophisticated rather than an obvious collage of known techniques, and mathematicians broadly considered the result strong enough for the Annals of Mathematics.
Hartnett treats that solution as the end of one round of moving goalposts: “AI did this but it can’t do that” kept retreating from contests toward research. Millennium Prize problems remain untouched, but this result showed AI can do “absolutely top-tier research.”
Terence Tao uses models to test many speculative ideas with lower cognitive friction—the “jetpack for your thoughts” or Iron Man suit view. His enthusiasm matters because he is intensely collaborative and has long experimented with new forms of mathematical work.

7. Mathematicians split between rejection, replacement, and augmentation

At the Institute for Advanced Study, Hartnett met two similarly elite 40-year-old mathematicians within an hour. One closed Gemini after it asserted a known falsehood; the other predicted that within two years AI would be strictly better and “put mathematicians out of business.”
Tao occupies Hartnett’s middle camp: humans still direct machines, choose problems, and use them to attempt larger work. Hartnett guesses this camp would currently win a poll, displacement would finish last, and “AI is good for nothing”—possibly the leading view a year earlier—is falling fast.
Hartnett does not pretend to know the endpoint. His hedged forecast is that mathematics will look radically different, but something so central to human activity is unlikely to vanish into “pushing a button”; human problem selection may remain consequential.

8. The Leiden Declaration defends standards, sovereignty, and human meaning

Roughly 800 mathematicians signed the declaration, warning that AI can produce plausible but unreliable arguments difficult to distinguish from valid proofs. Hartnett sees a centuries-old self-regulating community confronting a “massive exogenous force” and insisting: “This is our field.”
Some demands are ordinary rules for AI-era scholarship: disclose machine assistance and review generated text. A preprint archive has threatened a one-year ban when uploads contain unedited prompt metadata, signaling that authors did not even inspect what they submitted.
Roose asks whether this is merely a mathematical version of AI slop or the abacus guild resisting calculators. Newton jokes that abacus users would have gone straight to violence. Hartnett says the deeper fear is nearly the opposite: if machine proofs become genuinely excellent, mathematicians may lose their jobs and become hobbyists like great chess players.
The concern also covers downstream discovery and meaning: mathematics powers engineering, technology, and understanding of the universe, while proof is treated like a sonata or novel born from human struggle. Math may ultimately be discovered rather than invented, but Hartnett says humanity knows effectively 0% of it—there is no imminent danger of running out.

9. HatGPT finds weak controls everywhere, from model review to prediction markets

A startup called the Bot Company allegedly rented a San Francisco Airbnb under false pretenses, trained robots there for 11 days, emptied cabinets, and scratched the dishwasher. The owner seeks $12,383.50; Newton sides with the robots, while Roose insists clandestine “robots in body bags” require consent and a cleaning fee.
Trump’s executive order asks companies voluntarily to submit new AI models for government oversight, with review shortened from 90 days to 30 after objections associated with former AI czar David Sacks. Newton wants mandatory frontier-model testing; Roose calls the current regime the “vibes universe.”
Hackers said Meta’s support chatbot changed email addresses on prominent Instagram accounts, enabling takeovers affecting the Barack Obama White House account, Sephora, and the Chief Master Sergeant of the Space Force. Separately, a Newark-to-Mallorca United flight turned back about two hours after departure over a 16-year-old’s Bluetooth speaker named “Bomb.”
Prediction markets supply the episode’s clearest incentive warning: George Santos allegedly bet against his own State of the Union attendance after saying he would go; Survivor odds put Aubry Bracco above 80% before airing, though no crew leak is established; and a Google engineer was charged over an alleged $1 million Polymarket trade using search data. “No corner of society is safe.”

Dan Loeb: The Lost Art of Short Selling, and Why Stock Picking is Back Ep. 014 - Finding Miscompiles For Fun, Not Profit (AI Infrastructure) | Justin Lebar & Jordan Nanos