there is clearly overhype. Given the influx of people building here (and i am not aware if it happened previously too), in a bid to differentiate from other sta

there is clearly overhype. Given the influx of people building here (and i am not aware if it happened previously too), in a bid to differentiate from other startups building the same thing, many just stripped the nuance out of any technical idea, and made it a simple marketing term. As an exec where ten startups are promising you "training on your data" to provide the best chatbot, it's hard to tell the difference between who would be actually training and who would be just tweaking prompts. This has happened to other concepts too (anecdotally, most famous is how everyone is offering deep research). yes, it helps growth, but comes at the cost of trust. There is this startup which promised "experience based learning" when all they were doing is adding memory to the prompt to get it to perform better. (you can look it up, recenty raised series A). This does not mean ideas are not working. I personally think pretraining has done its job. We did not know what the job previously was, but now we do given the way RL works. Pretraining and test time compute enables models to develop a generalized prior they can use to solve any given problem (much like how humans solve such problems). Sometimes priors are lacking so you need to train more using RLVR, and still early days, but directionally I think we have another scaling curve here.