I integrating a cold war fighter pilot's decision framework into @karpathy's autoresearch. The result was 92% keep rate, better validation loss, same wall clock time. Fully autonomous. Here’s how it works and what I’ve learned so far (all open sourced, gist included below) Concept: Karpathy's autoresearch is a memoryless hill-climber: try a change, keep if better, repeat. No context for why things worked or failed. I wanted to know how the system would work with a realtime mental model so I built one based on John Boyd’s OODA decision-making model: observe → orient → decide → act Orientation is the Key: The implementation is a 40-line orient.md living document that the agent rewrites after every experiment. It tracks what works, what's closed (with specific experiment citations), the current best model, and ranked hypotheses for what to try next. More like a map than a log. More like intuition than a memory system. This is Boyd's key insight: Orient isn't just memory, it's a compiled mental model that shapes what you observe and decide next. Most Al agent loops skip this. They either have no memory, or they have a flat log the agent can't reason about efficiently. orient.md is the middle path. Caveats: These were single runs, ~9.5hrs each on an H100 - not claiming settled science. But the signal is consistent and promising. Modest val_bpb improvement margins, but a nearly 2x keep rate suggests decision quality can significantly improve with orientation. There’s more to find here. Misc. Architecture Insights : Several unique Boyd concepts shaped the architecture: Thin orient - maintaining experiment loop speed is essential to Boyd’s core argument that iteration rate is of utmost importance. The 40 line constraint explicitly eliminates orientation bloat. Deep orient - permission to unlearn is critical to learning so every 10 experiments automatically triggers tearing down and rebuilding the orient model from the full log. Boyd called this “destruction and creation” and it’s essential to preventing premature closure. Fingerspitzengefühl - German for “fingertip feeling” is a military concept that a good strategist needs a feel for the entirety of a situation rather than allowing their focus to narrow. Maintaining an awareness of architectural context also encourages broad exploration. Some Open Questions: - Does a 40-line compressed orient.md stay effective at 1000 experiments? - Does OODA improve exploitation at the cost of exploration? - Might a thicker orientation layer or even an in-the-loop tiered memory system be appropriate for certain types of research (at the cost of speed)?

2,012,724 crashes have occurred in NYC since 2012. Mapped and sorted by severity and mode. One major trend that I never knew about: crash volume dropped in half during COVID and never returned. Curiously the fatality volume was not impacted by the drop in total crashes.