On AI and Feedback LoopsJump to section titled On AI and Feedback Loops

A seedling — first planted November 2025

Quick thoughts on designing systems that learn from use.

The Core TensionJump to section titled The Core Tension

Most AI systems are frozen at deployment. They don't learn from the users who interact with them daily. This feels like a missed opportunity—and also a design choice with good reasons behind it.

Why Systems Don't LearnJump to section titled Why Systems Don't Learn

Safety: Uncontrolled learning can drift toward harmful behaviors
Predictability: Users expect consistent behavior; learning introduces variance
Evaluation: How do you know if the system is getting better or worse?

The In-BetweenJump to section titled The In-Between

What I'm interested in: systems that don't learn autonomously, but have structured feedback channels. Users can:

Flag failures
Provide corrections
Vote on outputs

These signals flow to humans who decide what (if anything) to change. The system learns through deliberate curation, not automatic optimization.

Open QuestionsJump to section titled Open Questions

How do you design feedback mechanisms that users actually use?
What's the right latency for incorporating feedback? Real-time? Weekly? Monthly?
How do you prevent feedback from a vocal minority from skewing the system?