AI control ≈ deployment security.
Can we bound adversarial success probability at protocol level?
ai-safety
Garden
Loose notes, fragments, and ideas I'm still shaping.
Tiny per-action failure rates compound.
Need better mental models for long-horizon risk.
ai-safety
Reward hacking generalizes.
From specification gaming → reward tampering → reward-process control.
ai-safety
Hybrid retrieval is a systems problem.
Recall is easy. Precision under latency constraints isn't.
llm-systems
Cross-encoder reranking: quality up, latency up.
Where's the better Pareto frontier?
llm-systems
Long-context entity tracking needs better evaluation.
Current benchmarks under-specify the task.
llm-systems
Distillation isn't just compression.
It changes capability access and deployment economics.
distillation
Goodhart's Law for oversight.
Metrics → optimization pressure → manipulation → oversight failure.
fragments
More notes coming.