@sunnygg@mattpocockuk Same here.. my company gives me a claude sub, so I use claude code for my daily job and pi with gpt for my personal stuf. Pi gets the job done way faster
My company gives me a Claude plan, so i have to use Claude Code. When I max it out, I switch to @pidotdev with my personal GPT plan.
And every time, Pi feels way, way faster.
The difference is honestly pretty striking.
Good morning from Vienna People of Pi 🌞
Sunday meditations from @badlogicgames and @mitsuhiko
- On memory: code is the truth
- Bash is all you need
- Build context efficient tools
I saw @dexhorthy interviewed by @GergelyOrosz recently - great chat, worth checking out
He talked about how AI coding sessions have a "trajectory"
I.e. the agent tends to pick up habits during the session. If it verifies the app by cURL once, it'll probably do it when you request the next change.
Sometimes you need to clear the context just to change trajectory.
I've even seen this when you have state that's retained across sessions. I've got a personal podcast generated from my feed. All previous podcast transcripts are available for it to view. So it picks up recurring catchphrases, and annoying traits. It's started over-focusing on meaningless metrics (wow, 400 bookmarks on this post!). My podcast is on a bad trajectory.
The "trajectory" metaphor gets really scary for codebases. Codebases are stateful - they accumulate state over sessions. So every codebase has a trajectory, and most trend down.
The only way to fix a bad trajectory is with better steering, and by changing the environment the agent works in.
Cool stuff. Thanks Dex
I could’ve just gone to the gym and paid for a PT. Instead, I combined my training and dev experience, used the @HevyApp API, and built an AI personal-trainer agent. Now I’m both the client and the engineer responsible when leg day ships with bugs.
Hello there... first tweet after a while.
My old account got blocked for violating some mysterious rules.
Apparently, understandig X is harder than debugging legacy code.