Writing
Occasional long-form notes on computer systems, distributed systems, hardware acceleration and AI. Subscribe via RSS.
State Is Not Context: What Long-Horizon Agents Need to Preserve
Why remembering more is not the same as knowing where you are
A long agent run accumulates records, not a current position. The harder capability is knowing what the past has made true and what has ceased to be.
Read the essay → · 12 min readThe Geometry of Reachable Futures
Thinking about long-horizon intelligence through what remains possible
A long-horizon agent needs more than the current state and the goal. It also needs to represent how each choice changes the futures that remain reachable.
Read the essay → · 22 min readTeaching AI the Art of Choosing the Right Problem
A systems engineering view of how long-horizon agents might learn where to spend their effort
Frontier models are strong at knowledge and reasoning. The neglected third capability is direction: deciding which problem is worth working on next.
Read the essay →