I’ve been saying this all started on consumer hardware, and brute force at scale isn’t necessary… and I still doubt this.

The researchers began with a base language model,

Ah, okay. They’ve bolted on cutting-edge features for a pittance. That’s still interesting - but it’s not nearly as exciting.

If it’s just for this one task, it’s borderline “homemade megaphone” territory. Like saying anyone who added one new anime character “replicated Stable Diffusion.”