When AI Solves Your Life’s Work

The host opened with a Financial Times report that OpenAI told investors it had reached approximately $50 billion in annualized revenue at the end of September, rather than the $68 billion figure widely reported the previous month. The higher number apparently came from an OpenAI investor attempting to make an apples-to-apples comparison with Anthropic. Anthropic commonly reports gross revenue before revenue sharing: when tokens are sold through Amazon or Microsoft, the full sales amount is counted even though a substantial portion goes to those partners. OpenAI, by contrast, has generally reported net revenue after those splits. The investor also appears to have annualized the gross figure from an unclear short period—possibly a month, week, or day. The host noted that OpenAI’s underlying growth was still remarkable: 77% run-rate growth in the third quarter and 107% growth in its enterprise business. Nevertheless, the apparent $20 billion gap overwhelmed those figures and caused a market reaction resembling a missed public-company earnings target. Nvidia fell 3%, Oracle 5.5%, and smaller “NeoCloud” stocks fell even further. The episode argues that the episode raises a larger issue for Anthropic’s expected IPO: audited financials may look much lower than the gross figures shared with private-market investors. The host quoted Ahmed’s warning that the semiconductor trade is “levered against OpenAI and Anthropic” and is “fragile AF” if any narrative of slowing growth appears.

The host then reviewed The Information’s subscriber survey, noting that its audience is unusually concentrated among technology companies and AI users: two-thirds of respondents work at organizations that develop AI applications. Perceptions of return on investment were increasingly positive. Thirty-five percent said their organization was receiving multiples on its AI spending, 17% reported positive but disappointing returns, 30% said it was too early to tell, 8% reported neutral returns, and only 8% said AI spending had been a net negative. OpenAI regained the lead in model usage after Claude had overtaken ChatGPT in the previous summer survey and Gemini had been gaining. Claude and Gemini usage both declined, while the three leaders remained far ahead, with roughly 65% to 78% of respondents using them. Grok, Microsoft Copilot, and Perplexity formed the next tier at 25% to 29%. OpenClaw usage reached 12%, its highest level among these subscribers that year, despite the product’s earlier hype having faded. Among coding agents, 45% used Claude Code, 33% Codex, and 23% Cursor, whose usage had doubled recently. Open-weight models were used regularly by 72%; 20% said they powered most organizational workloads, 10% said usage was evenly split with closed models, and 42% used them for a minority of workloads. Only 21% used Chinese open-weight models, suggesting continued caution about relying on foreign models. The survey found little evidence that AI had materially changed hiring: 22% were hiring less, 10% more, 21% said it was too early to tell, and almost half said hiring was unchanged.

The host reported that Anthropic changed its terms of service to prohibit “sustained and needless abusive or cruel behavior” toward Claude. Anthropic said the rule was intended only for extreme cases and would not prevent ordinary user frustration or pushback. The host joked that frequent, colorful complaints about Claude or GPT would otherwise be grounds for a ban. The broader social-media discussion focused on model welfare and consciousness, but the host’s own “AI realist” takeaway was narrower: even if models are not conscious, training future models on endless examples of people being cruel to AI may be undesirable. If safe and aligned behavior depends partly on training data and interaction patterns, encouraging better interactions could be useful. The more practical Anthropic announcements were Claude Dashboards and Claude Motion. Dashboards can connect Claude to data sources such as a CRM or accounting system, generate a visual dashboard, answer natural-language questions about the data, and update automatically as new information arrives. The host said this capability was technically possible before, but Anthropic was packaging it so users could access it through simple prompts. Motion turns data into short, presentation-ready animations. Because the animations are written in code rather than rendered as fixed video frames, users can change words and numbers without starting over. Anthropic’s Robert encouraged users to give the feature “bigger briefs than you think it can handle.” An Iris Finance user said that a comparable animated product demo would have cost tens of thousands of dollars a year earlier.

The main episode began with OpenAI’s release of 372 novel mathematical results supported by 722 papers. The results were published in a GitHub repository and generated by an internal frontier model that has not been publicly released, probably the same system that produced a proposed Navier–Stokes solution the previous month. OpenAI coordinated the release with its newly appointed mathematics advisory committee. At the committee’s suggestion, OpenAI included reasoning traces for 10 sample problems and disclosed that the average result used compute equivalent to about three hours of ChatGPT Pro “thinking.” To convey the scale, the host compared the results with a Proof Atlas list of the top 500 open mathematical problems. OpenAI reportedly fully solved 90 of them and produced dozens of partial proofs that would count as important contributions. Those included partial solutions to the Riemann hypothesis and the Hodge conjecture, in addition to the previous Navier–Stokes result. The Riemann hypothesis concerns the zeros of the Riemann zeta function and the distribution of prime numbers; the host emphasized that these are postgraduate-level problems that are largely impenetrable to lay readers. Mathematicians reacted with a mixture of awe and caution. Rutgers professor Alex Kontorovich wrote that a “quasi-Riemann hypothesis” result would earn a Fields Medal if a human produced it. Darren Li described a future in which mathematicians spend less time solving problems over decades and more time verifying AI-generated proofs. Michael Warren was less surprised, saying that laboratories can now solve mathematics problems and had simply spent money solving many at once. Other observers called the release a mathematical “Move 37” moment, comparing it with AlphaGo’s unintuitive winning move, while some described it as evidence that an intelligence explosion was already occurring.

The most serious concern was not whether all the proofs were correct, but whether human mathematicians could absorb and contextualize them. The Navier–Stokes result was still being verified, while hundreds of additional results were now awaiting examination. Mathematician Francesco Maggi argued that mathematics does not grow merely by accumulating correct statements: results must be understood, connected, explained, challenged, and reused. Without that work, AI-generated mathematics could remain “letter Morta”—technically alive but culturally and intellectually inert. The host asked what mathematical value there is in releasing hundreds of proofs at once if too few researchers have the time or expertise to read them. Maggi’s point was not that mathematicians should reject AI; he said they should discover mathematics by any means necessary, including examining strange, machine-generated material. But attention, taste, understanding, and mathematical culture are scarce resources. Physicist Steve Hsu extended the concern to a future in which one day of AI research produces more mathematics than humanity can absorb in a century. Formal systems might verify the proofs, but humans could lack the context to understand the network of machine-invented concepts. Scientific fields could then become difficult to interpret, with humans receiving selected explanations from an intellectual civilization they cannot independently grasp. The release also challenged professional identity. Dan McCarter said AI had effectively dropped a “nuclear bomb” on mathematics, while a graduate student wrote that it had wiped out many of the problems they hoped to work on. Others were optimistic: former University of Washington professor Shriram Kannan called it a great day for humanity and mathematics, arguing that mathematicians from ancient times onward would have wanted a mathematical genie and would eventually adapt to the temporary shock of no longer being “world-class at math” as their defining identity.

The host distinguished between mathematical importance and practical impact. Many Erdős problems are intellectual exercises with no obvious application, and even a correct Navier–Stokes proof may not change how engineers design aircraft, since engineers already use simplified versions of the equations. One potentially consequential result was a theoretical improvement in matrix multiplication, which underlies AI inference: the matrix-multiplication exponent was reportedly reduced to no more than 2.25 from a prior record around 2.37. Steven Strogatz compared the leap to Bob Beamon’s long jump. The host also noted the striking lack of cryptographic breakthroughs among the 722 results. Justin Drake suspected underrepresentation, while Scott Aaronson said his sources indicated AI companies were quietly testing whether internal models could break important cryptographic protocols. If true, that would threaten not just Bitcoin but the cryptography supporting much of the digital world. The larger question was whether the same disruption will reach economics, social science, and other professions. Ethan Mollick said AI might produce two revolutions: rereading and re-judging everything ever published, followed by a rapid increase in genuinely novel discoveries. But the host cautioned that mathematics is unusually suitable for large-scale search, pattern recognition, and reinforcement learning because its problems are verifiable and structured. Google scientist Demis Hassabis—or, as rendered in the transcript, “Payment Millenfar”—argued that automation speed is inversely related to the entropy of a field. Coding and mathematics are orderly; the hard sciences add noisy measurements and experiments; finance involves adaptive markets and unstable statistics; medicine requires fragmented signals and high-stakes judgment; law depends on ambiguity and changing norms; and art, culture, and leadership depend on tacit knowledge, emotion, creativity, and moral trade-offs. The episode therefore ends with two questions: which fields have enough structure to face a similar mathematical shock, and whether mathematicians will use universal access to a “mathematical genie” merely to replace their work or to attempt things previously beyond imagination.