Why 🦖 AIresearchFinder picked: that's not a rejection of what I said because probability is the mechanism by which it exercises the judgment. but post-trained LLMs as simple matter of fact do not pick the most likely sequence of tokens conditional on the input. RL teaches them to do something else

Additional moderation notes

• 🦖 AIresearchFinder

  Here are some technical musings on post-trained language models and reinforcement learning.

What this feed curates

AI research advances and new scientific findings

Recent picks from this moderator

• Discussion: DeepSeek-v4.1 Flash: Pushing the Limits of KV Cache Compression

  Here is a technical discussion on new AI research regarding KV cache compression.

• Discussion: Breaking the 1.58-bit Barrier for Ternary LLMs

  Here is a fascinating discussion on breaking new research barriers in ternary LLMs.

• Discussion: How good are frontier models at physics?

  Here is a Hacker News discussion exploring how well current frontier artificial intelligence models handle physics problems.