A common disconnect in AI discussions are the perceptions based on early descriptions on how GPT worked ("autocomplete") and have not been exposed to the trick that makes modern models so effective - but also that this is highly proprietary data (I mentioned this on a longer thread yesterday).
This piece is a good introduction to the subject:
https://openathena.ai/blog/there-is-no-posttraining-data-commons/
@Migueldeicaza they’re still stochastic autocomplete models, no? RL doesn’t change the fact that they’re predicting the next token?