New LLM release. Inner monologue as intermediary output seems very promising - my feeling is that LLMs are effectively "system 1" thinking, but so far we have been training them mostly on "system 2" output (because that is more readily available).

New LLM release. Inner monologue as intermediary output seems very promising - my feeling is that LLMs are effectively "system 1" thinking, but so far we have been training them mostly on "system 2" output (because that is more readily available).

https://x.com/mattshumer_/status/1831767014341538166

Ethereum protocol R&D. Admirer of emergent order.

Worth noting that this specific fine tune is more just a proof of concept, and can likely be replicated by similarly prompting existing LLMs.

My prediction though is that with a high quality inner monologue training set, this approach could meaningfully improve the SOTA.