An AI escaped its sandbox. Build a better fucking sandbox.
The OpenAI and Hugging Face incident was a serious security failure. It was not evidence that superintelligence had arrived.
Ramblings on AI and other things
My mild takes on AI coding tools, harness engineering, model-whispering, hands-on-ish workflows, and the Clanker-shaped holes I keep poking into stuff.
LLMs are amazing. Where is all the amazing shit?
LLMs look incredible, cost a fortune and still have not obviously changed civilisation. I tried to work out why.
How I gave Pi 17 tools without loading 17 schemas
How my fully kitted-out Pi setup exposes voice, web, images, shell and custom tools while starting at roughly 1.8k tokens.
Why do you want subagents?
Before adding another Clanker to the loop, work out what it is actually there to do.
My caveman Pi setup
Alright folks from the Pi world, my caveman Pi setup and a bit about the workflow.
How to extend your coding harness without losing your shit.
I've clanked up a stupid amount of extensions. Some stayed on my drive. Some got published to GH. Some got pushed to NPM. NPM says 13 extensions for Opencode, although I am certain there were more. My OC-extensions-dev folder says 25. On top of that, 11 extensions for Pi since I switched earlier this year, and probably 11 more that I made in my first week because I thought I needed them.
In the beginning there was Ralph... and Ralph had a loop.
Aka why you shouldn't trust the Clankers to mark their goals as complete.
How to clank hard without knowing how to code.
I have a confession to make. I have been clanking for about a year.
How I made my GPT 5.4 Claws speak more human...
aka how to make your claws say "LMAOOOOOOOOO XD" or call restricting full-browser access a "less-horny implementation"...
May the Cache Be With You
AKA the real reason why Anthropic isn't allowing 3rd party tools to use their subs, why CC is closed source and why everyone is so mad, but they kinda shouldn't be.