You can't trust it's account on why it did something, it does not "remember". It will just make up something plausible sounding.
You can't trust it's account on why it did something, it does not "remember". It will just make up something plausible sounding.
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.
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.
J-space is just one convenient projection (out of many) to look at these well known phenomena
Is there a way to dump the state directly?
Not unless you are anthropic/openai, or run your own models.
To me it explained its inability to follow my preferred coding style by blaming its training data.
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.
Sounds plausible.
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.
[flagged]
100% sentences like i convinced it or made it admit are just weird But how else to describe the situation
there is an established term for this behaviour: hallucination.
lets me rephrase what i said before as on retrospection the point i was trying to make didnt come across:
Asking the model such as question is a pointless endeavor. It will pretty much always answer in a plausible sounding manner to whatever question you gave it.
Its like another top HN story - an opus5 parody site right now, https://opusfived.dev/
pretty much the whole dialog is plausible given the prompts the user is required to make to illustrate the point the author of the site wanted to make... but the thing thats ironic about the site is that ... if youre prompting the model badly, it outputs garbage. though i'm pretty sure the author didn't want to make that point, to me its a perfect illustration of this.
And the same applies to asking the model to explain themselves and the related question. Asking for a source of the term is fine, and it will either respond with one and its cleared up -- or not and you know its a hallucination. nothing special about it, as they constantly hallucinating. its usually just self-correcting during the agentic loop.
"Asking the model such as question is a pointless endeavor."
It used a medical term that is not an actual term. I asked it to explain how and why it would do this, having a rudimentary understanding of the technology. I learned the technical reason. How is that a pointless endeavour to learn more about the tool I was using?
Using a term like "admit" assumes the reader (on a technical site) understands the context and what is being suggested. A bit like someone saying, "The landscape spoke to me." I don't reply to the landscape architect and say "Mate, the landscape can't actually speak" because I understand the sentiment and getting pedantic is pointless in that moment.
The word "admit" conveys a context of "fault". If you're going to say a similar thing about the landscape admitting to something then I'd be callimg you crazy, yeah.
The ai psychosis in this case it's specifically because of the used word. Same story as with words like "love" etc
> Confess or acknowledge: Agreeing that a fact, mistake, or crime is real.
And I can see were really barreling down the road of stupidity on this comment chain with one person being more idiotic then the next
Your HN comment is not useful. If you think it is, then you're suffering from HN psychosis.
>you're suffering from HN psychosis. No. They must be hallucinating.
And mind how 'they' so often select anthropomorphic terms, eg "hallucination" themselves, then ooze from the woodwork to behead someone who does the same, as if they own English and context itself. AI Psychosis, I think, simply means Double Standards and insulation from accountability in design.
There is a very interesting Plausible Deniability angle here. But I'll leave that for the future.
[flagged]
Ah yes. It's always the ones you most expect who come up with these retorts.
[flagged]
It can read its own context. That’s how this all works.