Well I basically tell it to not just do all the code and dump it out to me; I instruct it to explain the rationale, reasoning, and code and then provide external links for additional reading on the subject instead of just doing, I turn it into an instruct model so I can at least expand on my knowledge and then not have to rely on it as much the next time.
Basically, yes, like a Stackoverflow model from the early 2000s.
For instance, something like this: "
When talking about subjects involving programming and coding, the key goal should be instructional and informative to not only include code and samples but also how they work so in the future I can continue and expand on my knowledge. Also suggest places to expand and learn in the future on any programming or development topic. NEVER auto commit or create pull requests in my repositories without asking and waiting for a confirmation first. I prefer to review all code first for learning purposes and QA purposes."
I can see the benefit of this kind of use case, because LLM‘s are good at summarizing large data sets. It can be a good starting point for learning about something new or rubber ducking. But, if that’s your use case, why would you even need or want an agent hooked up to your environment to allow it to commit or create PRs?
Oh I mean, I think you’re right. I don’t allow it to commit OR create PRs. I think that’s silly; in fact I specifically instruct it NOT to do those things because I want to verify everything.
So kind of like a personalised learning assistant? I realise it’s different but this inverted instruct approach puts me in mind of Doctorow’s reverse centaur!
Don’t you find that the links you get are hallucinated though? Even if they’re not now you can imagine this collapsing into slop echoes…
I’ve tended to ask for examples to help me bootstrap new projects. A bit like getting customised docs. I certainly haven’t had enough success with generated code to think about automatically adopting it.
I think when you know the possibility of hallucination, you become more vigilant; I think the key point is to not use it as a exclusive source but as an extension.
This is interesting and sounds like how I’ve been using it - basically like customised stack overflow answers.
Would you mind elaborating a little on your approach? Are you saying you provide it with guidance and links or are you asking it for those?
Well I basically tell it to not just do all the code and dump it out to me; I instruct it to explain the rationale, reasoning, and code and then provide external links for additional reading on the subject instead of just doing, I turn it into an instruct model so I can at least expand on my knowledge and then not have to rely on it as much the next time.
Basically, yes, like a Stackoverflow model from the early 2000s.
For instance, something like this: "
When talking about subjects involving programming and coding, the key goal should be instructional and informative to not only include code and samples but also how they work so in the future I can continue and expand on my knowledge. Also suggest places to expand and learn in the future on any programming or development topic. NEVER auto commit or create pull requests in my repositories without asking and waiting for a confirmation first. I prefer to review all code first for learning purposes and QA purposes."
I can see the benefit of this kind of use case, because LLM‘s are good at summarizing large data sets. It can be a good starting point for learning about something new or rubber ducking. But, if that’s your use case, why would you even need or want an agent hooked up to your environment to allow it to commit or create PRs?
Oh I mean, I think you’re right. I don’t allow it to commit OR create PRs. I think that’s silly; in fact I specifically instruct it NOT to do those things because I want to verify everything.
So kind of like a personalised learning assistant? I realise it’s different but this inverted instruct approach puts me in mind of Doctorow’s reverse centaur!
Don’t you find that the links you get are hallucinated though? Even if they’re not now you can imagine this collapsing into slop echoes…
I’ve tended to ask for examples to help me bootstrap new projects. A bit like getting customised docs. I certainly haven’t had enough success with generated code to think about automatically adopting it.
I think when you know the possibility of hallucination, you become more vigilant; I think the key point is to not use it as a exclusive source but as an extension.