Gpt-5 is finally good! Liking codex now

LLMs produce 5x as much code as a skilled engineer for the same amount of work The code always looks sensible locally, but then you zoom out and realize the same types are duplicated 3 times with slight variations and fallbacks

I guess meta was ahead of its time with the name change

Where there's a will, there's a way... My LLM doesn't have read/write file tools - it just has bash. So when I ask it to edit files it just uses grep/sed/awk or a full-blown python script to do what it needs

ugh AI produces the most gnarly bugs that it can't fix itself. Currently dealing with an issue where a stream of messages aren't showing up in the UI *sometimes* - no console errors. I suspect it's actually a database issue but i don't have a prior schema to verify 😬

it'd be really handy if claude code could not only read/write files in the project, but control my editor to walk me through bits of the code. think speech bubbles next to a given function and i can step through and ask questions along the way is there an MCP for this?

Fantastic commentary and analysis to a podcast some people were disappointed by. It had more disagreement than usual but that's what made it interesting!

for my toddler, brussels sprouts are the bitter lesson

so annoying that you have to use `docker compose` instead of `docker-compose`. LLMs are trained on the latter now, you can't change it!

I hope DHH single-handedly gets everyone to use Linux

me when first playing with LLMs: Wow this is so amazing! Look at all the complex UI and features you suggested we build! let's do it! me now: that's way too complex and overengineered, there's no way you're going to one-shot that

tired: using RAG wired: using "agentic search" with grep inspired: index your data and let claude also make RAG queries if it wants to as a tool call

oops my agent was adding a feature to itself and then caused itself to be hot-reloaded which kills all the current agent processes

Wow, listening to a car guy podcast and what do you know, it's an ad for... Claude! 🤯

tired: permissions dialogs wired: use --dangerously-skip-permissions and ask claude nicely not to rm -rf / while planning