Research
Six projects on trustworthy AI.
Each project names one hard problem and the instrument we think could help. All are early. None claims a result.
Early research
Replay
Make every agent run something you can pause, inspect and replay.
Agent runtime →Early research
Recall
Memory an assistant can prove, correct and forget.
Memory and context →Early research
Proof
Tests a field expert would actually sign off on.
Evaluation →Early research
Bounds
The unglamorous failures that break deployed AI.
Safety and control →Early research
Lean
Capable systems that do not need the biggest cluster.
Efficient intelligence →Forming
Fieldwork
Start every program from a problem a practitioner can state.
Cross-field →