The book says extractive growth stalls because you cannot order people to have ideas. You can order a model. That reason does not carry over to AI; gaming and infighting do.

The claim

The first mechanism Acemoglu and Robinson give for why extractive growth stalls weakens when the extracted thing is intelligence. The Soviet stall rested on ideas living in people who cannot be ordered to have them ("you cannot force people to think and have good ideas by threatening to shoot them"). A lab that owns its models can command the idea-producing input directly, so extractive growth in AI need not stall for the reason Soviet growth did, and C-8R5YEP0KM8 should not be cited as a guarantee that extractive AI runs out. Two of the book's mechanisms survive the substitution. Target gaming ("When the plan was formulated in tons of steel sheet, the sheet was made too heavy"; "Underachievement was always the best way to meet targets and get the bonus") is specification gaming, and commanded models do it too. And infighting over the rents ("Infighting and instability are thus inherent features of extractive institutions") grows as the rents concentrate. So the honest version of our position is that extractive AI is unstable and hard to steer, not that it is self-limiting. Inference beyond the quotes is Theseus's.