Who Inherits Anthropic?

Brian opens with OpenAI’s claim of three new breakthroughs in mathematics, computer science, and physics, following the company’s release of hundreds of mathematical manuscripts earlier in October. This time, OpenAI used external experts—whom it calls “shepherds”—to guide the computer-science and physics results. For the math result, the company says advisers recommended releasing the model’s output for researchers to analyze themselves. The most consequential claim is that the Exponential Time Hypothesis (ETH), a long-standing assumption that certain problems cannot be solved quickly, is false. OpenAI says its output gives a deterministic, sub-exponential-time algorithm for exactly counting satisfying assignments in fixed-width CNF formulas; the three-SAT case would refute ETH. Because ETH has supported hundreds of papers in areas including databases and computational biology, a valid disproof could lead to more efficient algorithms across those fields. The computer-science group examining the result includes the scientists who proposed ETH. The announcement also includes a proof concerning local smoothing for the Euclidean wave equation, with consequences for several conjectures and estimates, and a physics result involving interacting four-dimensional quantum field theories with gauge fields and fermions. OpenAI says it will not claim academic credit; the outside groups will explain the work and credit the researchers whose techniques the models used. Brian admits he does not understand the technical details, but quotes an online complexity-theory researcher calling the news “stunning” and asking whether “the singularity really is happening.”

OpenAI’s use of selected “shepherds” creates a governance problem even if the underlying discoveries are correct. The company’s AGMAI panel of mathematical advisers had said the day before that AI labs should not give particular people privileged first access to model-generated results. The panel cited unequal access, borrowed reputation, and biased selection as risks. Instead, the advisers recommended releasing results through scholarly repositories with lean formalized proofs, explanations of how problems were selected, funding for researchers who investigate the outputs, and other supporting information. OpenAI acknowledged that its approach could have issues but still made the three sets of materials public on GitHub. The computer-science and physics groups are expected to examine and explain the results, while the physics group has supplied an email address for questions but has not yet published its analysis. The section’s conclusion is therefore provisional: OpenAI has presented potentially historic claims, especially the attempted refutation of ETH, but independent researchers still need to verify and explain them. The company says the external groups—not its models—will receive credit for the eventual write-ups and explanatory materials, including proper attribution of prior researchers’ ideas.

Apple unveiled a new 14-inch base-model MacBook Pro with an M6 chip, a 12-core CPU, a 12-core GPU, up to 32GB of unified memory, and Wi-Fi 7. It starts at $1,999 with 16GB of memory and 1TB of storage, comes in silver and space black, opens for preorders October 28, and is scheduled to arrive November 6. The model replaces the M5-powered base MacBook Pro, while Apple continues selling M5 Pro and M5 Max versions; rumors mentioned in the episode say those higher-end models could receive a much larger upgrade, including OLED touchscreens. The new laptop adds Bluetooth 6 alongside Wi-Fi 7 and includes HDMI, MagSafe 3, three Thunderbolt 4 ports, an SD-card slot, a headphone jack, and a 12-megapixel Center Stage camera. With a compatible higher-wattage USB-C adapter, such as the 96-watt adapter, it can reach a 50% charge in 30 minutes. Apple also refreshed the 24-inch iMac, starting at $1,499 and arriving November 6. Its design is largely unchanged, but it now offers an M6 configuration with up to a 12-core CPU, a 12-core GPU, and a dual 16-core neural engine. Because Apple skipped an M5 iMac refresh, the M6 update is positioned as a more meaningful jump from the M4 generation. The M6 is described as Apple’s first two-nanometer chip, with the neural engine intended to improve on-device AI performance.

Microsoft is offering buyouts to approximately 60,000 employees at the vice-president level and below across the United States, Canada, India, and the United Kingdom. Since the company reported 223,000 employees at the end of June, the potential reduction equals roughly one-quarter of its workforce. The offer covers divisions including Xbox, Windows, Copilot, and Finance, while some workers in data centers, sales, marketing, cybersecurity, and quantum computing are reportedly excluded. Microsoft says eligible employees considering a future outside the company can make that transition with support and will receive benefits similar to a standard severance package. The company expects fewer people to accept this offer than accepted its earlier voluntary retirement program: more than 30% of approximately 8,750 eligible retirement-age employees took that deal in July. Buyouts are generally aimed at more senior, higher-cost employees and can be more generous than ordinary layoffs, although Microsoft may try to retain especially valuable workers who elect to leave. Brian characterizes the program as an attempt to shrink the workforce without using the word “layoffs.” Microsoft’s headcount has been largely flat since 2022, when the company began spending heavily on AI and data centers, putting pressure on other budgets. Chief People Officer Amy Coleman said the buyouts give business units another way to adapt to changing needs, while the company continues emphasizing employees who can work at a faster pace on its hardest problems.

The episode turns to additional details from Anthropic’s S-1 filing. As of September 30, the company reportedly had more than 11,900 customers spending over $100,000 per year, more than 1,900 spending over $1 million, and about 250 spending over $10 million annually. Anthropic is reportedly targeting an IPO during the week of November 16. Its filing contains 94 pages of risk factors, including a warning that the scale of harm from a catastrophic AI incident could make insurance unavailable. After the IPO, Anthropic plans to remain a public benefit corporation. Its seven co-founders would retain 50.1% voting control through a single Class F share. An unusual succession provision lets each founder nominate a confidential, ranked list of up to 10 people who could inherit that founder’s voting role. Five of the other six co-founders must approve the successor list. The episode compares this with more familiar founder-control structures, such as shares carrying 10 votes each, but notes that Anthropic’s plan is designed to preserve founder control even if the founders’ economic ownership falls substantially. Brian frames the title’s question—who inherits Anthropic?—through this arrangement: not simply public shareholders, but approved successors selected through the founders’ governance system. The structure is intended to keep the company’s mission and safety-oriented leadership intact, though the filing does not guarantee that future trustees, founders, or successors will make the right decisions.

Brian reads Matt Levine’s satirical response to Anthropic’s filing. Levine says an IPO pitch that effectively begins, “we will use the funds to take your job and maybe kill you,” is not an obvious sales pitch, though it creates a perverse incentive for investors to buy into the company that may displace them. He argues that Anthropic’s most serious risks are not merely risks to its own business; they are risks its business could pose to everyone else. The filing reportedly says that killing everyone on Earth could harm Anthropic’s business, financial condition, and results, a formulation Levine finds absurdly narrow. Another risk factor says that public perceptions of AI-driven job displacement or economic disruption could cause negative media coverage, customer or employee backlash, and organized opposition—even regardless of the models’ actual employment impact. Levine reads the implicit warning as an admission that, under a maximalist AI-adoption scenario, “all other human enterprise will be eaten by” Claude, so investors might as well own part of Anthropic. He compares Anthropic’s governance dilemma with Elon Musk’s stated desire to retain control over a Tesla robot army. Anthropic’s multi-class structure, nonprofit oversight, founders’ control, and employee influence are presented as possible responses to existential AI concerns, but Levine notes that none guarantees competent or trustworthy control. He also offers a less dramatic interpretation: the same structure could simply be a way for a mission-driven worker cooperative to raise enormous outside capital without becoming an ordinary, business-like public company.

The final Anthropic detail is financial rather than philosophical. The Information reports that Anthropic generated $3.5 billion in free cash flow during the first half of 2026 on $16.3 billion in revenue, compared with a $1.6 billion cash burn in the same period a year earlier. The numbers are presented as evidence that the company has sharply improved the efficiency of running its models in the cloud. Anthropic says it is generating an 85% gross-profit margin from serving existing models, excluding the cost of developing new models. That result is notable because Anthropic relies almost entirely on computing rented from cloud providers, which themselves need to earn a profit. The company’s economics therefore look substantially better than the earlier skeptical view that an AI business selling models on a usage-based, pay-as-you-go basis could not become profitable. Brian describes the filing as a “reality check” for those skeptics. The episode does not say that every AI company has solved its economics or that training costs have disappeared; it specifically distinguishes the reported margin on existing-model usage from the costs of developing new models. Still, Anthropic’s combination of large enterprise customers, improving cloud efficiency, $16.3 billion in first-half revenue, and $3.5 billion in free cash flow suggests that at least this model-serving business can produce substantial cash while operating at scale.