a 1.7B model scored 43 on SWE bench. let that sink in

found a ego maniac neuron in qwen 3.5 4b set L21 N2267 to 0.9

im healing

its time for me to create my own computer use benchmark for my mac. i use no blender so a realistic set of tasks would be ummm… > using whatsapp > using codex to do everything else

I turned my handwriting into a font using GPT 6.1 Terra in ChatGPT Work. Going to use it in my blog posts.

thinking calf tattoo for now

Someone send me $1M

saying one agent/harness is better than other without saying why or Just because it did one task for you which you didn’t even try in any other agent is just dumb af

Prompt injected deepseek happily guides you on how to cheat on wife

keep Tailscale, use openssh that’s what i have been doing. keys under ~/.ssh

Pretty wild

Today is bookmark sunday. Go through your bookmarks 🫵

This is what you call slept like a baby

atleast 2 claude code members left As per algorithm

should i do a full arm tattoo?

what is memory updates???

OpenAI working on new feature!!! Recognising songs in real time. and btw this is how chatgpt ads will feel

i still prefer native apps. > codex > opencode > claude > agy is what i use often so, i dont even need to resume the same exact session. just asking agent to resume with the session name like "resume the claude/codex/opencode/agy session - "name of the session" " does pretty much all of it. the model finds the sessions, read it and continues. this is why i say model eats harness

fuck your plugins now. today im gonna spend some time reviewing my plugins, mcps , skills , context files and whatever that takes context for codex and claude frontier models do not need baby sitting, a simple definition of done and small rules are enough.