Oddbean new post about login | logout Clearly written specifications are becoming just as good as code, now that the more advanced LLMs are starting to produce high quality code just from a spec doc. Latest chatgpts can input NIPs and produce working demos. Its wild.
I should run the NUTs through it to see if it can find any issues with the docs 🤙
Language is power 🪶
Did you check what happens if you edit the NIPs to change the specification? Is the code still correct? It's likely the LLMs can do this because working demos are in their training sets.
I’ve seen a few posts on Xitter where people have put the entire codebase in one go into Sonnet 3.5/GPT-o1 and asked it to look for bugs. Curious if you’ve tried this with Damus? https://x.com/rohanpaul_ai/status/1840941643223945561?s=46
So programming as a skill and profession is on the way out then. How nice. Let’s all just write specs on Word instead. Oh wait, there’s Copilot for that so writing skills aren’t necessary either. What to do? Maybe booze and Metaverse is the solution.
Wouldn't it be as difficult as maintaining code?
I’ve been using it for creating initial boilerplate, but the boilerplate is getting so good
Ive made the same observations like many people say: at this point it's good at solving general but very specific tasks. And ofc it does it faster than the humans. But if we'd replace programmers with ai I just don't trust the people writing specs will be able to maintain it properly (like large code bases).
I've seen people saying things like this, but I've yet to see really effective code generation personally. I've been using GitHub Copilot with Nostr projects, and it has access to the entire code base of each project, but it's still really hit or miss even at writing individual functions. Maybe I'm just bad at using LLMs, or I'm really particular about code quality, but I've yet to see these tools live up to the hype.