Building: https://xitter.cf/t.co/r5YnSiqUYi - your own whiteboard on steroids. https://xitter.cf/t.co/gl1XpSYK3j — my own coding agent
Helsinki, Finland
Joined October 2012
- Tweets343
- Following154
- Followers208
- Likes243
5.6 models are so freaking slow, and eats $200 limits in speed of lightning. Went back to 5.5 😐
5.6 Max finally passed this benchmark, the first model ever to do so.
This isn’t some “draw me a pelican” BS. It’s a complex 3D engine built from scratch, with destructible voxel planets. 20+ hours of agent's working
The capabilities have taken a massive leap forward.
Today they broke Codex app completely
After the last update the Codex App is almost unusable. It asks permission on everything even if I've turned off every possible option in UI. Super annoying stuff
@OpenAI
Codex 5.3 was insane speed& quality in the first 24 hours of the launch. Now it is still decent, but like 10x slower.
Every time you open&close Mission Control your windows are just sorted randomly.
Stupid AF, best UX from #apple
I have tried Grok-4.1-fast recently in complex image recognition & reasoning tasks.
It beats every other comparable VL model in text recognition (qwen3 vl 235, gemini flash) and is very, very smart for its price.
Very impressive
I have started blocking anyone ranting about Gemini 3 pro right away. The feed immediately become much better.
Feel free to mention more of those clout chasers to block.
Making an app onboarding is the most annoying thing ever 😬 Everything is working, but I need to explain the concept to other people in a pair of screens. Pretty hard I must say
AGI is not in the near future, it is already in the past.
AI is _thinking_ better than the average human, it just lacks proper universal memory.
The best name for neural network calls in your software is “smart functions” (not agents).
Embrace this concept deeply, and you’ll see how to build scalable and reliable intelligent programs.
Developers will deny it -> Git pipeline and tooling is hell. It’s super efficient & You suffer every day.
Still figuring out how to launch this thing.
Coding agents are everywhere, hard to stand out. But I hacked together a vulnerability scanner for fellow vibe-coders 😁. Drop in a repo or folder and it’ll tell you what’s wrong.
I just realized that while building the new Shotgun, I’ve basically ended up making something… lovable 😂
Showed it on stream today — I’ve got a cloud server with the engine running. It pulls a Git repo, uses LLMs to work on tasks, and commits changes back ->
