alt Hacker NewsI suppose this is as good a place as any to mention this. I've now met two different devs who complained about the weird responses from their LLM of choice, and it turned out they were using a single session for everything. From recipes for the night, presents for the wife and then into programming issues the next day.
Don't do that. The whole context is sent on queries to the LLM, so start a new chat for each topic. Or you'll start being told what your wife thinks about global variables and how to cook your Go.
I realise this sounds obvious to many people but it clearly wasn't to those guys so maybe it's not!
Replies
Yeah I think a lot of us are taking knowing how LLMs work for granted. I did the fast.ai course a while back and then went off and played with VLLM and various LLMs optimizing execution, tweaking params etc. Then moved on and started being a user. But knowing how they work has been a game changer for my team and I. And context window is so obvious, but if you don't know what it is you're going to think AI sucks. Which now has me wondering: Is this why everyone thinks AI sucks? Maybe Simon Willison should write about this. Simon?
> I realise this sounds obvious to many people but it clearly wasn't to those guys so maybe it's not!
It's worse: Gemini (and ChatGPT, but to a lesser extent) have started suggesting random follow-up topics when they conclude that a chat in a session has exhausted a topic. Well, when I say random, I mean that they seem to be pulling it from the 'memory' of our other chats.
For a naive user without preconceived notions of how to use these tools, this guidance from the tools themselves would serve as a pretty big hint that they should intermingle their sessions.
Problem is that by default ChatGPT has the “Reference chat history” option enabled in the Memory options. This causes any previous conversation to leak into the current one. Just creating a new conversation is not enough, you also need to disable that option.
In my recent explorations [1] I noticed it got really stuck on the first thing I said in the chat, obsessively returning to it as a lens through which every new message had to be interpreted. Starting new sessions was very useful to get a fresh perspective. Like a human, an AI that works on a writing piece with you is too close to the work to see any flaw.
I constantly switch out, even when it's on the same topic. It starts forming its own 'beliefs and assumptions', gets myopic. I also make use of the big three services in turn to attack ideas from multiple directions
It's not at all obvious where to drop the context, though. Maybe it helps to have similar tasks in the context, maybe not. It did really, shockingly well on a historical HTR task I gave it, so I gave it another one, in some ways an easier one... Thought it wouldn't hurt to have text in a similar style in the context. But then it suddenly did very poorly.
Incidentally, one of the reasons I haven't gotten much into subscribing to these services, is that I always feel like they're triaging how many reasoning tokens to give me, or AB testing a different model... I never feel I can trust that I interact with the same model.
I was listening to a podcast about people becoming obsessed and "in love" with an LLM like ChatGPT. Spouses were interviewed describing how mentally damaging it is to their partner and how their marriage/relationship is seriously at risk because of it. I couldn't believe no one has told these people to just goto the LLM and reset the context, that reverts the LLM back to a complete stranger. Granted that would be pretty devastating to the person in "the relationship" with the LLM since it wouldn't know them at all after that.
That is interesting. I already knew about that idea that you’re not supposed to let the conversation drag on too much because its problem solving performance might take a big hit, but then it kind of makes me think that over time, people got away with still using a single conversation for many different topics because of the big context windows.
Now I kind of wonder if I’m missing out by not continuing the conversation too much, or by not trying to use memory features.
Thing is, context management is NOT obvious to most users of these tools. I use agentic coding tools on a daily basis now and still struggle with keeping context focused and useful, usually relying on patterns such as memory banks and task tracking documents to try to keep a log of things as I pop in and out of different agent contexts. Yet still, one false move and I've blown the window leading to a "compression" which is utterly useless.
The tools need to figure out how to manage context for us. This isn't something we have to deal with when working with other humans - we reliably trust that other humans (for the most part) retain what they are told. Agentic use now is like training a team mate to do one thing, then taking it out back to shoot it in the head before starting to train another one. It's inefficient and taxing on the user.
This is why I love that ChatGPT added branching. Sometimes I end up going some random direction in a thread about some code and then I can go back and start a new branch from the part where the chat was still somewhat clean.
Also works really well when some of my questions may not have been worded correctly and ChatGPT has gone in a direction I don't want it to go. Branch, word my question better and get a better answer.
It is annoying though, when you start a new chat for each topic you tend to have to re-write context a lot. I use Gemini 3, which I understand doesn’t have as good of a memory system as OpenAI. Even on single-file programming stuff, after a few rounds of iteration I tend to get to its context limit (the thinking model). Either because the answers degrade or it just throws the “oops something went wrong” error. Ok, time to restart from scratch and paste in the latest iteration.
I don’t understand how agentic IDEs handle this either. Or maybe it’s easier - it just resends the entire codebase every time. But where to cut the chat history? It feels to me like every time you re-prompt a convo, it should first tell itself to summarize the existing context as bullets as its internal prompt rather than re-sending the entire context.
How are these devs employed or trusted with anything..
I know I sound like a snob but I’ve had many moments with Gen AI tools over the years that made me wonder: I wonder what these tools are like for someone who doesn’t know how LLMs work under the hood? It’s probably completely bizarre? Apps like Cursor or ChatGPT would be incomprehensible to me as a user, I feel.