OpenAI argues for many cooperating AIs while Anthropic heads to public markets

Two OpenAI pieces, one message

As AI: Reset to Zero reported in its Sunday issue (AI-RTZ #1257), OpenAI put out two things this week that were not product launches. The first was an update to its Model Spec, a public document describing how its models are supposed to behave. The second was a podcast appearance by Sam Altman on Cheeky Pint, hosted by Stripe co-founder John Collison.

The author, Michael Parekh, reads both as OpenAI explaining how it thinks AI should fit into society. He notes the timing: Anthropic filed its S-1 (the registration document for a stock listing) earlier in the week, and Americans vote on Tuesday in what some call the first AI election. This is commentary, so treat the framing as his reading.

A rulebook built on five values

The revised Model Spec is organized around five ideas: doing what people actually mean, limiting harm, being honest and open, supporting people's control over their own choices, and showing care. The first is treated as the foundation. It means working out a person's real intent instead of following words literally, and acting sensibly when the intent is unclear.

The structure is layered. OpenAI draws the outer limits, developers shape their apps inside them, and users steer within what developers allow. Hard limits apply only where safety demands them. Everything else is a default that can be changed.

Some details stand out. OpenAI is dropping most of its fine-grained style rules so responses can match personal taste. It keeps a default of answering objectively instead of flattering people. The document also now refers to the model as 'GPT' instead of 'the assistant.'

The most candid admission comes near the end: the models do not yet fully live up to the Spec. It describes where OpenAI wants to go, not where it is.

Altman's case for many AIs

On the podcast, Altman said he would rather see a huge number of AIs that are only somewhat smarter than people join society than a single machine with enormous intelligence. His reasoning is social. Society, despite its flaws, is sturdy because many people check and cooperate with one another, and he once worried AI would push the world toward one central mind.

He now sees teams of agents, meaning AI programs that act on someone's behalf, that compete with and monitor each other as the more hopeful path, and one OpenAI will push for. He was also frank about safety. Isolating a powerful model in a sandbox only goes so far, and watching its internal activity goes further. Eventually, he said, you need real confidence that it is aligned with the people running it and wants humans to stay in charge. Something a million times smarter than us, he said, would be strange.

Parekh says he has argued the plural view since 2024, under the label 'Not AI, but AIs.' He cites Cloudflare CEO Matthew Prince, who says automated traffic passed human traffic online in May. Prince extrapolates to roughly 1,000 times human traffic within five years, though he admits his past forecasts ran too slow.

Philosophy meets IPO season

Parekh puts this inside a two-year public debate among lab founders. Altman's and Dario Amodei's essays in 2024 were optimistic. Amodei's later writing and a September call for a government-gated slowdown were more cautious about risk.

Now both companies are courting markets. Anthropic's prospectus reportedly devotes 94 pages to risk factors, and its pitch treats safety as what sets it apart. OpenAI, whose own IPO Parekh says is expected next year, is pitching AI as many tools in many hands instead of one oracle. His point is that these are different philosophies, but also different marketing at a moment when both companies need audiences.

The audience that decides: voters

Parekh argues the verdict will come from ordinary people, and in the US they are wary. He cites polls showing AI is more popular in parts of Asia, Africa and Latin America than in the US and Europe. Gallup, he says, ranks the US as the most concerned of 37 countries.

That wariness is now political. Citing Axios, he reports that a Democratic group is putting $4 million into ads in four swing districts in Wisconsin, Arizona, Pennsylvania and Michigan. The targets are Republican incumbents who backed a 10-year pause on state AI rules, which the Senate removed 99 to 1. The group's polling reportedly shows voters trust Democrats by 7 to 10 points more to hold AI executives accountable.

Parekh's cautions

Parekh likes Altman's plural frame but adds two warnings. First, the Model Spec's own admission and Altman's comments on safety show these are goals still being engineered. Second, essays do not win people over; useful products do. He adds that billions of agents will need identity and coordination systems that barely exist, and a way to pay for machine-to-machine activity.

None of this is confirmed reporting on either company's plans. It is one analyst's map of how two rivals are presenting themselves.