Altman Wants a Group of AIs to Check and Balance Each Other; Commenters: Every PR Is Approved by the Person Who Wrote It

In the Cheeky Pint interview, Sam Altman offered an AI safety vision. In the world he described, there is no single, god-level AI. A large number of AIs only slightly smarter than humans check and balance one another, like the multi-party contest in a healthy society.

Rohan Paul summarized this on X: Altman believes human society is robust enough because many players keep one another in check. He wants AI to follow the same model and avoid putting all intelligence in one place. The full video is on the Cheeky Pint YouTube channel; Rohan Paul's original post has a link in the replies.

This sounds like separation of powers and checks and balances. Models review one another, like code review. The problem lies exactly here.

One commenter grumbled in a reply: AIs checking and balancing each other, so every pull request is approved by the same colleague who wrote it. The analogy hits on a reality: checks and balances depend on independent reviewers. If AIs share training data, objective functions, and reasoning paths, their independence may be only superficial. The review chain may look complete, while blind spots are shared together.

Another commenter, Code With Jerry, was more optimistic. He believes that a group of models slightly beyond human level checking and balancing one another may be safer than concentrating intelligence in one AGI. Decentralizing power is sometimes the best safeguard. This judgment makes sense, but it has many preconditions. Models must have real differences; reviews must be traceable to external evidence; authority and objectives must not be concentrated in the hands of the same controller; and when something goes wrong, someone must be able to shut it down, roll it back, and hold it accountable.

Giga Gerard's reply was more direct: Yeah robust. Tech people are so naive. He also attached an image of two men speaking behind a podium. The image itself does not explain much, but under the sarcastic line “robust,” its meaning is clear enough.

Altman's statement is more of a design direction than a proof of safety. A healthy society not only has multiple players, but also laws, media, ballots, conflicts of interest, and external accountability. If a group of AIs lacks these institutional layers, mutual checking can easily degrade into a multi-agent version of code review: the process is complete, responsibility is dispersed, and in the end no one can say clearly who should be responsible.

So the key is not the number of AIs. The key is whether the checks and balances are truly independent. Decentralized intelligence can reduce the risk of a single point losing control, but it can also spread the risk thinly across every node, making it look as if there is no problem anywhere.

Source: https://x.com/rohanpaul_ai/status/2116283505919643984