I built an operating system where the AI agent IS the process.
No Linux. No POSIX. A tiny LLM runs on bare metal in Rust β and a capability ABI means it can plan anything, but only do what it was granted.
It boots, browses the web, plays video, and talks. π§΅
I built an operating system where the AI agent IS the process.
No Linux. No POSIX. A tiny LLM runs on bare metal in Rust β and a capability ABI means it can plan anything, but only do what it was granted.
It boots, browses the web, plays video, and talks. π§΅
~3,800 lines of mechanism in a ~300k-line kernel, so the part that has to be right is readable in an afternoon.
Research OS β not stable; run it in a VM. Apache-2.0, and I'd love contributors.
β https://t.co/r4JPEhDVUf
π https://t.co/d4xlGk4hnc
I built an operating system where the AI agent IS the process.
No Linux. No POSIX. A tiny LLM runs on bare metal in Rust β and a capability ABI means it can plan anything, but only do what it was granted.
It boots, browses the web, plays video, and talks. π§΅
The honest part: the defence costs something. 25% of legitimate irreversible operations get refused because the turn touched untrusted data.
I built the cheap fix β per-value string matching β and it permits 4 attacks while recovering zero. The model is the dataflow.
Introducing SubQ - a major breakthrough in LLM intelligence.
It is the first model built on a fully sub-quadratic sparse-attention architecture (SSA),
And the first frontier model with a 12 million token context window which is:
- 52x faster than FlashAttention at 1MM tokens
- Less than 5% the cost of Opus
Transformer-based LLMs waste compute by processing every possible relationship between words (standard attention).
Only a small fraction actually matter.
@subquadratic finds and focuses only on the ones that do.
That's nearly 1,000x less compute and a new way for LLMs to scale.
Been on Cursor for a while, but Go dev never felt quite right with VS Code/Cursor.
Always had to fall back to GoLand.
Zed + Claude thoughβ¦ this finally clicks π