[josefresco]

This resonated with me based on my recent experience using Claude to help me code. I almost gave up, but re-phrased the initial request (after 7-10 failed tries) and it finally nailed it.

> 3. Performance improves with multiple attempts Allowing the o1 model 7 attempts instead of 1 nearly tripled its success rate, going from 16.5% to 46.5%. This hints that current models may have the knowledge to solve many more problems but struggle with execution on the first try.

https://newsletter.getdx.com/i/160797867/performance-improve...

[Suppafly]

I haven't really messed with Claude or other programming AIs much, but when using chatgpt for random stuff, it seems like the safety rails end up blocking a lot of stuff and rephrasing to get around them is necessary. I wonder if some of these programming AIs would be more useful if it some of the context that causes them to produce invalid results was more obvious to the users.

[runlaszlorun]

> safety rails end up blocking a lot of stuff

curious if you had any examples. i'm fairly meh on llm coding myself but have a pet theory on safety rails. i've certainly hit plenty myself but not with coding with llm's.

[Suppafly]

With chatgpt, for noncoding things, it has rails to avoid things like copyrighted art, violence, adult topics, etc. For coding LLMs, I suspect they have things like preferences for certain data structures, avoiding directly returning training data (even if that training data might be the only feasible way to do something), preferences for certain languages and APIs, etc.

If you knew what some of those preferences and rails were ahead of time, it'd be easier to design your request and also know why it making some odd or unworkable suggestions.