Dario Amodei and Gavin Baker spent the weekend arguing about whether regulating AI concentrates power or protects against it. Whats it about?

Amodei rejects the choice itself: He argues that AI concentrates power on its own, because training frontier models depends on capital, chips and energy that few organizations can assemble, and that open weights do not fix this, since access to compute stays with the frontier labs and the large hardware providers. His answer is institutional: rules that fall harder on the biggest developers than on smaller ones, testing for models capable enough to matter whether they are open or closed, and something like the FINRA style oversight body Demis Hassabis has proposed.

Baker's problem is with who survives that: Compliance is a fixed cost, and fixed costs are a gift to whoever is already biggest, especially when the biggest are also in the room while the rules get drafted. His second point is the one I find harder to wave away. He thinks the constant catastrophe talk is doing political work whether Anthropic wants it to or not, feeding the local campaigns against data centers and against buildout generally, and that this makes the good version of all this less likely rather than safer. Amodei rejects that reading of his own record and points at Machines of Loving Grace. Fair enough, though that is not the essay people quote back at him.

I want a future with many models, many builders and real competition, where useful intelligence becomes abundant and, for a growing number of tasks, too cheap to meter. I want a future where we work together to combat the fear of AI. And I want a future where frontier AI becomes freely available to as many people as possible.