Supportive of strong American open-weight models. We're taking action to move more of our work loads to these models at @coinbase.

I think open weight models may actually out-perform frontier models inside companies, because you can fine tune them via reinforcement learning and modify the weights (not just rely on stuffing text into the context window). For example, an open weight and fine tuned model that has seen 100,000 of our compliance cases internally could be a better specialist than any frontier model on its own, which are balanced to be smart across a wide variety of domains.

There is a hybrid approach one could imagine, where an API is provided by a frontier lab or cloud vendor to help you fine tune a frontier model, without revealing the weights publicly. There is still a data sovereignty question here, but at least it would allow fine tuning.

In any event, glad we're seeing more open weight models.