Right I agree, but this is basically the entire internet (and more, like scanned books), what it is trained on. They are using reinforcement learning now. And this is also one of the reasons why they have escaped the sandboxes in the American models for example.

So you can probably safely assume that they are trained on all data that is available easily. and very likely also artificially generated data at this point. How long they are trained on, what kind of data is a different story. I think it gets increasingly more interesting how these models are trained.

So while I agree it would be really nice to know exactly what it is trained on and how it is trained on that, I think that those that have the resources to get at least close to frontier models (i.e. AI companies) don’t just open up on that, sadly.