News for hackers. Stuff that matters.
Re:ChatGPT Now Losing Users by YeGoblynQueenne on Thursday July 13, 2023 @03:33PM (#36714821)
>> Why is it so hard to believe that they have lessened the capability of the chat model?

One obvious question is: how would they do it? How does one nerf a language model? Train it again with less data, or different hyperparameters, especially chosen to make it worse? Given the costs of training LLMs that sounds like it would need a very strong motivation.

Fine-tune it, or RLHF it so it's doing worse? That's not cheap either, and what would be the benefit justifying the expense? Nerf a model, to achieve what?

Besides I think you're assuming a degree of fine control on LLM training that just isn't there. If it was so easy to control performance, it would also be much easier to train (both pre-train and fine-tune) LLMs, and OpenAI would not be in the dominant position they are right now.

There are no comments on this comment yet.