Long context isn't a capability. It's a standing charge.
Long context isn't a capability, it's a standing charge. The same task can cost seventeen times more read in full than retrieved.
Where high-performance engineering meets LLM systems.
Exploration, building, and the things worth shipping. The work that earns the build - skipping the boilerplate, aimed at meaningful impact.
Lenses borrowed from running high-performance engineering teams, applied to context, inference, memory. When the open-source gap is worth closing, the work goes on this site, in the posts, and on Hugging Face when it's a quant.
Steven Batchelor-Manning.
Latest posts
All posts → 
You bought a 1M context window. You got 50x less than you paid for.
The advertised context window is 2 to 8 times larger than the effective context for multi-hop work, and 50 to 100 times larger than the effective context for reasoning. The number on the slide is the size of the door. The number that does work is the size of the room.
Your memory system does not need to decide what the agent sees. The agent does.
Nineteen agent-memory systems quietly reversed their biggest design choice: they stopped injecting context and gave the agent tools instead. Here is why.