Yining Karl LiMar 13

You… you know... that LLMs... don't continue training when you use them... right? Like, the P in GPT stands for Pre-Trained?... When you chat with an LLM it doesn't retrain anything in the model because the model is a snapshot and....

oh nevermind why bother.

Show threadVincent BernardiMar 13
@yiningkarlli replace the word model by agent and I believe the argument is quite correct. An agent will often write wrong code but provided with a better compiler error will correct itself better and faster.
Show threadYining Karl Li
@Enzo90910 Until you run out of context and compact, or you clear the context, and then you're back to square one. There is no actual training or learning happening even with agents; it's all just part of feeding inference. Fundamentally the model weights don't change until an additional RL run, which obviously doesn't happen anywhere near the user interaction loop.
Mar 14 at 10:01amWeb
Show threadVincent BernardiMar 14
@yiningkarlli I agree with what you’re saying, but I think the original argument that agents could perform better for languages that have very specific compiler errors is valid.