Should Every Country Build a Frontier AI Model?

Demis Hassabis would not have every country build its own frontier AI model. He expects half a dozen offerings at or near the frontier, six to eight if Chinese companies are included, leapfrogging by maybe three to six months and staying relatively close. An enterprise can choose among them, because one model will be stronger at multimodal understanding and another at coding, and leading on every measure at once is very hard.

Open source sits in that set and maybe only six months to a year behind, which is why he says Western open-source options are needed. The open problem is how to restrict bad actors while still letting scientists and other good actors use the tools. A country or a company that needs its data kept private, or that wants to fine-tune further on sensitive or sovereign data, can do that on top of an existing model. Big sectors such as healthcare can treat the models almost like commodities and build the value on top.

Training a frontier model in every country does not make sense to him. The shortcut is some other path. Reid Hoffman suggests arrangements with the multinational companies that train the models, and Hassabis agrees.