I have access to it. It is better, but not where most techies would care. It knows more, it writes better, it's more pleasant to talk to. I think they might hav

I have access to it. It is better, but not where most techies would care. It knows more, it writes better, it's more pleasant to talk to. I think they might have studied the traffic their hundreds of millions of users generate and realized where they need to improve, then did exactly that for their _non thinking_ model. They understand that a non-thinking model is not going to blow the doors off on coding no matter what they do, but it can do writing and "associative memory" tasks quite well, and having a lot more weights helps there. I also predict that they will fine tune their future distilled, thinking models for coding, based on the same logic, distilling from 4.5 this time. Those models have to be fast, and therefore they have to be smaller.