Discussion: create more entropy

[pinnerwt]

Hi, I’m pretty sure that if I post this on X, no one will see it—so I’m posting here instead.

I’ve been experimenting with something similar to autoresearch on OpenAI’s parameter golf challenge, and I feel that what autoresearch is lacking is insight, rather than implementation. Insight could be introduced by allowing the agent to search the internet or draw from its existing knowledge. However, this still lacks what I’d call “auto-inspiration” or true innovation—where the LLM generates a genuinely new idea it hasn’t encountered before under these conditions.

Here’s what I’m thinking might help improve “auto-inspiration”: we could feed in a random sequence of tokens as a source of “innovation.” Sometimes, an agent just needs a small spark of insight to retrieve or apply its knowledge—it simply can’t find that spark because it’s optimized for a global loss. With this setup, users could inject random token sequences, hoping they trigger useful insights.

More concretely, I’m imagining that something like “asdf;lkmxzcv” might nudge the LLM toward thinking about stochastic gradient descent, while “dfm,./zcv” might lead it to consider Adam’s optimizer. More interestingly, this approach might even allow insights from entirely different disciplines to emerge and help solve the problem at hand.

Open to discussion.

[balala8]

Interesting idea. I think a completely random and meaningless word would be ignored. Perhaps we could give a relevant or irrelevant word, like a domain name or a noun? For example, when doing LLM iterations, we could suddenly be given an ant, or perhaps a human brain?

[morozow]

Your idea seems like well aligned with neural mirroring in a general sense. Check the discussion at #234, and follow the forked repo branch to find valuable touch points for recognizing thoughts related either to this repo's goals or to your own thoughts machine next to the original instance.