I think part of the idea here is that we’re not talking about putting GPT-4o in charge of a company, we’re talking about GPT-7a (for “agent”). By the time we get to that turn of the game, we may not have as many issues with hallucinations and context size will be immense. At a certain point the AI will be able to consume and work with far more information that the human CEO who “employs” it, to the point that the human CEO essentially becomes a rubber stamp, as interactions like the following play out over and over again:

AI: I am proposing an organizational restructuring of the company to improve efficiency.

CEO: What sort of broad philosophy are you using to guide this reoorg?

AI: None. This week I interviewed every employee and manager for thirty minutes to assemble a detailed picture of the company’s workings. I have the names of the 5272 employees who have been overpromoted relative to their skill, the 3652 who are underpromoted or are on the wrong teams, 2351 who need to be fired in the next year. Would you like me to execute on this plan or read you all the rationales?

CEO (presumably after the AI has been right about many things before): Yeah OK just go ahead and execute.

Like, we’re talking about a world where CEOs are no longer making high level “the ship turns slowly” decisions based on heuristics, but a world where CEO AIs can make millions of highly informed micro-decisions that it would normally be irresponsible for a CEO to focus on. All while maintaining a focus on a handful of core tenets.