Hacker News Reader

Comment by vzmax | Hacker News Reader

You can't trust it's account on why it did something, it does not "remember". It will just make up something plausible sounding.

View on HN

I know this is a popular position and it makes sense at face value when you think of LLMs as autocomplete systems. But it’s genuinely wrong.

Relevant reading is most notably anthropic’s research on the J-space. LLMs will plan ahead of time helped with CoT, get to a plan and “store” it in j-space, and execute on that plan which means they can in fact “backtrack” and give you reasoning on why they did something, because it IS part of their state.

Is there a way to dump the state directly?

Not unless you are anthropic/openai, or run your own models.

J-space is just one convenient projection (out of many) to look at these well known phenomena

True which is why I said anecdata. But the point is that it’s a useful signal when it’s across enough sessions where it quotes back contradictions to you. If you’re willing to burn the tokens, the first thing it does is use a tool call to search your rules to see what it quoted wasn’t there. That leads to the LLM claiming two common conclusions (sometimes after a lot of back and forth) - it’s in the harness or it’s in the model. Over the 30+ sessions where I’ve tried this, the overwhelming claim was some variant of a harness instruction. It might not be there but given the consistency and how all my rule tests have failed in the same way that others describe, it at least makes it reasonable to conclude that it’s baked in somewhere and in a way where agent rules aren’t able to strongly affect the behavior. My bet would be on the harness because the class of undesired behavior follows that of a structured response.

Meta:

> True which is why I said anecdata

Forgive me for being cheeky, but presumably this is not why you said anecdata.

You presumably said anecdata because you were describing your own personal sessions with Claude. Your original comment is written as someone who is assuming the AI is doing a deep deterministic analysis of its own internal systems in order to respond to you. It could simply be aping some discussion on the same topic within its training data, which if sensible is, as you say, not entirely worthless...

Fair. I should have reigned in my emotions and written that first comment with a lot less conviction and not jumped to my conclusion. It’s a topic I’ve been spending a lot of time learning about and got a bit too excited when a relevant thread popped up where I could chime in.

To me it explained its inability to follow my preferred coding style by blaming its training data.

Sounds plausible.

I've had better success with having it write and then rewrite rather than trying to prompt the first write into following my style rules.

also by abusing linters and other guardrails you can basically automate the rewrite

I mean that’s not abuse? That’s what mechanically checkable invariants are for! They constrain the space the model has to work in.

It has access to its own instructions, right? Otherwise how would the instructions work?

All instructions to LLMs are merely suggestions to nudge it in the right behaviour. Unless you have a deterministic guardrail that guards against a single specific action, everything else is a slot machine that's biased strongly in your favour.

My global CLAUDE.md explicitly states "When commenting on code and configs, or writing MD files, strictly write within the domain of the content being commented on. DO NOT include information, negatives or ramblings from work sessions. For e.g. if commenting on a proto string field that is replacing an int field, do not comment that 'this is not an int field'".

This reduced the idiocy of the agent (Opus 5 included) when writing documents. But I'm still catching it writing README.md talking about the negatives that it removed. Those belong in the memory if it is actually that important (most of the time it's junk), but Claude doesn't seem to understand and never ever learns.

I've had Claude completely invent a word, and then admit under questioning that it was because it seemed plausible.

It can read its own context. That’s how this all works.

Reply on news.ycombinator.com