Hacker News new | ask | show | jobs
by xnx 431 days ago
Can't speak for anyone else, but my own AI chat history has low/no relevance to the quality of response to the next question I ask. This is not a moat any more than search history is.

My email and work documents are obviously important if I'm querying for information about them, but that is self evident and also not a moat (I could grant another tool access to these things).

Computational efficiency is a moat. If Google can provide an AI response for $0.05 of infrastructure and electricity, but it takes OpenAI $0.57, that's bad news for OpenAI.

1 comments

Does that mean you never engage in multi-question dialogues exploring / trying to solve a particular problem? In other words, every query you have is unrelated to all others?

If that’s the case, it mostly just seems like you’re not working on sufficiently complex problems to find the AI useful. Or are just keeping that complexity in your head, and just bringing in the AI “as a consultant,” as it were.

If that’s the case, I recommend trying to organize your project with the AI from the start. I’ve had a lot of productive benefits from treating ChatGPT folders as ongoing conversations about a particular project, questions I have on it, random ideas, etc. Memory is absolutely crucial for my use case.

> Does that mean you never engage in multi-question dialogues exploring / trying to solve a particular problem?

No. Iterative interrogation is the main way these tools are used, hence "Chat" GPT. It is rare that I'm revising queries from a week ago.

More useful AI context comes from permanent (and portable) artifacts like a code repo. Having a 2 million token context window is much more useful than being able to continue a chat session from a week or more ago.

Also that memory (and your conversations, your interactions) are the actual moat. There's plenty of code out there, but there's not a lot of "how does a developer interact with a code base" outside of commits.

The interaction data is the actual interesting but, but there's no guarantee that its the refinement that's best needed.

I run into this all the time with chatgpt, we hit a roadblock on A, go down path B, we solve B which solves A, now it cant rewind to keep solving A so i have to start over.

OR it keeps telling me one thing or another

"X didnt work, heres the output"

OK try "X"

ok buddy.