Amodei's "pace the frontier" calls also serve lab finances

The article's argument is that the safety message and the business message are pulling in the same direction. As Anthropic and OpenAI prepare to go public, they must convince investors they can one day be massively profitable — and an open-ended race to build the biggest, most capable models is an expensive way to make that case. The author notes that solving a Millennium Prize Problem is great for bragging rights, but it doesn't necessarily translate into better reliability on mundane business tasks.

That claim rests on an observation about where reliability actually comes from: not from the models themselves, but from all the software built around them. The article points to AI pioneer Richard Sutton's work at Oak Lab as a possible different direction — a capable model that runs on 20 watts of power and continuously updates its weights, like the human brain. The author also flags the tension this creates: AI safety experts might worry more about such continually learning models than today's big ones, since they would be harder to contain.

More efficient AI could lower the labs' costs, but it could also undercut their business models if customers could get comparable capabilities without relying on their data centers. That tension — between cheaper, more independent AI and the infrastructure the labs have already funded — is what the outcome of the public listings likely hinges on, and it is the question investors and enterprise buyers will have to weigh.