Post

haha, I’ve stumbled upon that blog at almost the same time as your post

Just further proof that currently, if you are an expert, AI can boost your skills. If not, maybe less so... Takeaway. Become an expert and outsource to AI everything you can.

I wonder how important the particular prompt is. Does it matter that you're asking for *better* code? What happens if you say something like "Please rewrite this code"? It would still give the model the benefit of having existing code to start with.

Yes, unless it has "long-term memory" which then becomes a rabbit hole of semantics.

Yeah it's one of my favorite unintuitive details about LLMs - "chat" is just autocomple against a little screenplay simonwillison.net/2024/Jul/14/...

Funny, whenever I tell myself to write better code, it never happens.

As an aside, I find what makes your linklogging work so well on Bluesky is that you can have multiple link previews in the same post -- the original and yours. Works very well, and don't think I've seen that on the other socials

Which helps explain why "do better" is so effective - each prompt becomes "here's is a solution, suggest a better one"

Great post. I’ve found a similar thing in my interactions with Claude writing Terraform configuration (to deploy web services and supporting infrastructure)… I start out with a prompt stating what I want, get something decent back, then iterate with more prompts and better results.

I think this is why some agentic workflows produce seemingly novel results - they’re designed to iteratively explore until a condition is met. We don’t necessarily do that when chatting with an llm.

Oh interesting, I hadn't considered that. It kind of is? You're effectively prompting it to think step-by-step through repeated iterations of "now think some more"

And now I see that's what the original article starts with, which I only read after I read yours :-)