@nedbat One thing I'll be interested to see is how developers respond to the inevitable rug-pull, because access to the models are currently being offered at unsustainable prices (and I mean that even in the context of "pushing electrons around is very cheap;" training the models and evaluating hundreds of thousands of queries per day is not too-cheap-to-meter). So when the inevitable price hike comes, some companies and users will choose to pay and some will fall off.

Of those who fall off... I wonder if any will roll their own replacements? Executing the model is cheaper than training it, and I don't think software patterns change so quickly that an infrequently-trained model would fall out of use. So it might only take one weights exfiltration for a lot of users to be able to spin up their own "good enough Claude" in a server farm...