OpenAI pivots toward a future of trillions of diverse AIs

A new rulebook for AI behavior

OpenAI recently updated its Model Spec, which serves as the public rulebook for how its models should behave. This document reorganizes the company's goals around five primary values: faithfulness to human intent, minimizing harm, honesty and transparency, supporting human agency, and care. These values are intended to guide how the models interact with the world as they become more capable.

Faithfulness is the core foundation of this new framework. It requires models to go beyond literal instruction following to understand what a person actually values and intends. When a user's intent is unclear, the model is expected to act sensibly based on that underlying understanding rather than just guessing.

The update also includes specific changes to how models handle style and information. OpenAI is retiring most detailed style rules so that the AI can adapt to individual preferences while keeping a default of objectivity. The company also noted that its models do not yet fully reflect these specifications, describing the Spec as a destination rather than a current state.

The case against one superintelligence

In a recent conversation with John Collison, Sam Altman argued against the development of a single superintelligence. He stated that he would much rather have "a trillion AIs of slightly superhuman intelligence" that can join society. This perspective suggests that a pluralistic approach is safer and more robust than a centralized mind.

Altman's reasoning is rooted in the way human society functions through collaboration and mutual checks. He expressed concern that AI might trend toward a single centralized mind, which could undermine democratic and pluralistic values. Instead, he sees a future where teams of agents compete and check one another, creating a more balanced ecosystem.

This vision aligns with the idea that plurality is itself a safety feature. If there are billions of agents working for different people, the possibility of any one AI taking over the world becomes as remote as any one human doing so. Altman noted that while sandboxing and monitoring work to a point, true safety requires confidence that models want people to stay in control of the future.

Marketing and the IPO landscape

These philosophical shifts come at a time when both OpenAI and Anthropic are positioning themselves for public markets. Anthropic is currently marketing its mega-AI IPO into the Thanksgiving season, using its 94-page prospectus to highlight safety as a key differentiator. OpenAI is expected to follow with its own IPO next year, pricing itself against Anthropic's margins.

OpenAI is now pitching a different philosophy to investors and the public. Rather than presenting AI as a single oracle, they are framing it as many tools in many hands. This move seeks to differentiate OpenAI from Anthropic's safety-first branding by emphasizing a democratic, pluralistic future of many diverse agents.

The competition between these two labs is not just about technical capabilities but about which narrative will win over the public. While Anthropic leans into the risks of powerful AI, OpenAI is leaning into a vision of human agency. Both companies are trying to establish a trustworthy brand as they prepare for the scrutiny of public trading.

Political and social hurdles

While the labs debate philosophy, the general public remains divided on the impact of AI. Recent global polls show that while AI is popular in parts of Asia, Africa, and Latin America, it is least liked in the US and Europe. In the US, Gallup ranks the country as the most concerned of 37 surveyed nations regarding AI's impact on humanity.

This mainstream resistance is beginning to influence the political landscape. Reports indicate that Democratic groups are spending millions on ads in swing districts to test the backlash against AI. These ads target Republican incumbents who previously voted for a moratorium on state AI regulation, focusing on holding AI executives accountable for harm.

Voters are increasingly expressing concern over data centers and the costs they impose on utilities. The anxiety is bipartisan, showing that the