Sam Altman: OpenAI’s new internal model solved multiple academic problems over the weekend and left him feeling melancholy
OpenAI CEO Sam Altman said in the latest episode of the Cheeky Pint podcast hosted by Stripe co-founder John Collison that the company once obtained a new internal model and set up a channel called “Model Name-Hold My Beer” to test it. Researchers put in the hardest research questions from their own academic careers, and the model successively solved these problems across multiple fields. Altman said that as a result, he was so excited for two consecutive nights that he could hardly sleep, but afterward he also felt sadness, melancholy, and a sense that “something was ending” that was far beyond expectations.
In the interview, Altman pointed out that experiments related to proving the Navier-Stokes equations involved roughly $10 million in computing power, possibly a little more, and at the time they knew almost nothing about the scaling laws for agent teams. He said that with smarter models and better methods, similar work might be redone a year later at about one percent of the cost. He made it clear that he does not like the term “Agent Swarms,” and prefers “Agent Teams,” and he noted that frontier users can already use chain-of-thought to carry out tens of thousands of parallel attacks. On safety, he admitted that he is pessimistic about whether sandboxes can continue to constrain superintelligence far beyond humans, saying that supervision and true alignment are needed, and said he would rather have one trillion slightly superhuman AIs balance each other as they enter society than see a single superintelligence with an IQ of one trillion.
Altman also criticized the current public narrative about AI in the industry as “terrible marketing”: on the one hand promising to solve cancer and lower prices, while on the other hand depicting a world in which ordinary people lose their jobs, give up control, and rely on surveillance and basic income. He stressed that one reason OpenAI insists on iterative deployment is to avoid a small elite in San Francisco making decisions for billions of people around the world. The interview also touched on PHP vulnerability discovery, compute supply constraints, a preference for deflationary technologies, and how personal agents will make traditional web form filling “below wage grade.” The episode ran for about 53 minutes, was released around October 29, 2026, took place in a bar setting, and Altman drank his first Guinness.
Original sources:
Title: Why Sam Altman is against a single superintelligence Publisher: Cheeky Pint (hosted by John Collison) Publication date: 2026-10-29 URL: https://cheekypint.substack.com/p/why-sam-altman-is-against-a-single Nature of source: original interview transcript Verification status: this reading used the full available transcript (including the PHP, math, agent, safety, and compute sections), verified on 2026-10-30 Supporting content: description of the Hold My Beer channel, two nights of insomnia and feelings of sadness/melancholy, about $10 million in compute for Navier-Stokes, preference for Agent Teams and ten thousand-way parallelism, pessimism about sandboxes and the one-trillion-AI balancing view, terrible marketing and reasons for iterative deployment, and the Irish horse analogy (raised by Collison)
Title: OpenAI’s @sama joined me at the pub (and had his first Guinness) to discuss the PHP exploit, math weekend at OpenAI, agent swarms, and how he thinks AI leads to deflation. Publisher: John Collison (@collision) Publication date: 2026-10-29 URL: https://x.com/collision/status/2115809881399922849 Nature of source: post by the party involved Verification status: this reading used the full post and accompanying video information, verified on 2026-10-30 Supporting content: the interview took place, topics covered the PHP exploit, math weekend, agent swarms, AI and deflation, and the first Guinness
ABAB AI Insight
The “Hold My Beer” weekend disclosed by Sam Altman in this pub interview directly corresponds to OpenAI’s Navier-Stokes-related work that the company publicly discussed in September 2026. At that time, the company said an internal model and about 10,000 agents produced a proof within 88 hours, sparking debate in the mathematics community, including questions about sourcing and verification. In the interview, Altman acknowledged that they knew nothing about the scaling laws for agent teams at the time of the experiment, and predicted that costs could fall 100-fold quickly, which is consistent with his long-standing emphasis on test-time compute scaling. He had previously described, in settings such as Stanford, how mathematicians moved from “we’re done for” to accepting a changed role; here, he explicitly framed his own feelings as melancholy after excitement, rather than as a simple technical victory.
On the capital and compute path, the roughly $10 million investment in a single experiment reflects OpenAI’s practice of giving frontier researchers broad compute budgets internally. Altman said the company does not reward mere high token consumption, but important work can be supported; he also noted that memory suppliers such as DRAM makers are conservative in expanding production because of historical boom-bust cycles, which has created current supply constraints. This contrasts with OpenAI’s continued expansion of data centers and partnerships, such as with Oracle: in the short term it relies on expensive compute to achieve capability breakthroughs, while in the long term it is betting on falling costs and deflationary technology. The interview did not disclose the exact total cost of that weekend experiment or the internal model name.
Compared with historical cases, the rapid breakthrough in mathematics resembles the impact of port automation on dockworkers — Altman used the reaction of dockworkers’ unions as an analogy in the interview. The key similarity is that technological progress irreversibly changes the value and identity of certain intellectual labor; the key difference is that this time a single company’s internal model completed in a few days work that previously required extensive human research, and did so amid public controversy and verification needs, rather than through gradual industrial migration. Collison’s analogy involving the end of horse cultivation in 1950s Ireland provided a frame from the observer’s perspective.
In specific terms, this disclosure strengthens OpenAI’s leading position in test-time compute and multi-agent collaboration, shifting frontier math problems from a “human one-on-one fight” to an engineering problem that can be tackled in parallel with budget. What changes is the entry point and cost structure of mathematical research: actors with large compute budgets can quickly verify or generate proofs, while independent researchers’ work of verification and understanding remains necessary. Altman’s preference for “one trillion slightly superhuman AIs balancing each other” rather than a single behemoth, combined with his iterative deployment stance, points to his attempt to reduce concentrated risk through distributed deployment, though he still regards sandbox effectiveness as an unresolved problem. The indicators to watch are the speed of cost declines in subsequent similar experiments and the mathematics community’s acceptance of their verification.
ABAB News · Cognitive Law
Compute budgets turn frontier mathematics from a human one-on-one contest into an engineering problem that can be tackled in parallel.
Better to have one trillion slightly superhuman AIs balancing each other than a single tyrant with an IQ of one trillion.
Technological breakthroughs are often accompanied by the irreversible sealing away of old intellectual labor temples, with excitement and melancholy coexisting.