And what’s wrong with distillation? In its own right, and especially considering how those models were trained in the first place - in what world is it right that US firms are free to use copyrighted / otherwise non-public works however they like, while others are ostensibly in some moral wrong for just using the product (completely normally) and then using the outputs in a way that threatens US firms.
> Also, from a scientific perspective, distillation is decades old and contributes nothing scientifically.
Distillation itself does not, but it does enable research by allowing more teams to play around with models that follow the frontier, as opposed to completely independent training including scaffolding up w/ synthetic data.
They can simply host the models somewhere like Singapore and then sell the service from there. I mean, my Xiaomi account gets routed through Singapore.