Troed Sångberg on Nostr: I get 80% of the "boring work" done (scanning the codebase to figure out where X ...

I get 80% of the "boring work" done (scanning the codebase to figure out where X happens, routine library and API setup stuff etc) so that I can concentrate on the 20% which are tricky and requires insight. The effect of that is that I'm now much more inclined to jump into new projects since there's less effort into learning _yet another thing_ before I can be productive.

That's for development. The rest of what I do is in cybersecurity, where LLMs connected using MCP to a reverse engineering tool (Ghidra, in my case) are crazy effective. The LLM walks through the while unknown binary and maps up data structures, functions, including renaming them if it recognizes them, all the parameters etc. Not something I can't do myself - but the model does the same work in 30 minutes that takes me 10 hours.

I never try "oneshotting" solutions with LLMs though. I use local models and actively follow the "thought"-process and interject when I see it's needed.

Erik Jonker (npub1jnv…zpfx) Some Guy (npub1duw…xq6f) Max Leibman (npub1388…sxrr)