yeah its not fair to call it a finetune because finetune carries connotation of "there wasnt that much extra compute and data added". RLFT has a lot more added to it as Sasha alluded in his talk https://www.youtube.com/watch?v=md8D8eNj5JM - the x axis is log scale, think about that

the framing here is more about "why would you start from random weights when perfectly good starting weights exist" https://www.latent.space/p/fastai

It's more than that. They have both their own completion model and now agentic one. It's not a basic fine-tune, because it's faster than anything else available out there, so there's something interesting in the architecture itself.

As a user, I don't care.

Composer-1 is very good for routine code edits.

Claude and Gemini get pulled in for hard problems and architecture.