[tsunamifury]

LLM technology specifically beam-searches manifolds (or latent space) of lingustics that are closely related to the original prompt (and the pre-prompting rules of the chatbot) which it then limits its reasoning inside of. Its just the basic outcome of weights being the primary function of how it generates reasonable answers.

This is the core problem with LLM tech that several researchers have been trying to figure out with things like 'teleportation' and 'tunneling' aka searching related, but lingusitically distant manifolds

So when you pre-prompt a bot to be friendly, it limits its manifold on many dimensions to friedly linguistics, then reasons inside of that space, which may eliminate the "this is incorrect" manifold answer.

Reasoning is difficult and frankly I see this as a sort of human problem too (our cognative windows are limited to our langauge and even spaces inside them).

[nomel]

This is why I only use chat clients that allow me to modify both my previous messages AND the AI's previous messages. If the AI gets something wrong, and you correct it, you're now in a latent space with an AI that gets things wrong! It's very easy for context to get poisoned this way. I also see all the pre-amble of many chat clients as a type of poison for the context, so use the raw, blank, API if I need best problem solving results.

[astrange]

This is one of the benefits of using subagents inside Claude Code, they have cleaner context. Unfortunately it's not the best at writing new context for them.

[afpx]

What you're saying sounds pretty cool but can you give some examples? Is this what you're talking about?

https://chatgpt.com/share/69f246e5-e0e8-83ea-aa88-6d0024b915...

[tsunamifury]

yea this is a good example, its the nature of sort of how you salt the prompt -- regardless of any baseline truth it will search the various manifolds mathematically closest to the order and type of words you put in. It will do that always and willingly. Thats what the technology does.