Live data from Hacker News

A Bitter Lesson for Memory

personal-website-3bed.onrender.com

1–4 of 4 posts

Re: A Bitter Lesson for Memory

#3
I've always been amazed at how terrible most frontier LLMs are at compaction given how embarrassingly easy it is to come up with half a dozen different RL training evals which would teach models to generate useful context summaries. Heck, you could bolt it onto any existing RL eval by just forcing a compaction every three turns.

Re: A Bitter Lesson for Memory

#4
post #3

I've always been amazed at how terrible most frontier LLMs are at compaction given how embarrassingly easy it is to come up with half a dozen different RL training evals which would teach models to generate useful context summaries. Heck, you could bolt it onto any existing RL eval by just forcing a compaction every three turns.

Yep. Or even better, compact after a random number of turns. The model must then learn to preserve useful context at arbitrary context lengths.