Writing
Posts on AI agents, search and interaction design.
- Hiding the hard partsInstinct feels magical because of two design decisions, not new technology. Decide the experience first and let it choose which parts the user sees.
- Talking in percentagesAI bugs are probabilistic, so one observation tells you nothing about how often it happens. Measuring a new behaviour should be cheap enough to do inside the investigation.
- Categories of hallucinations“The model hallucinated” describes very different failures. Naming the root cause is how you find the right fix.
- Dynamic contextContext, not raw intelligence, is the bottleneck for agents. Keep each step in the goldilocks zone by injecting and removing context as the task moves.
- Visual content search should be visualDescribing a style in words is lossy. What if searching for visual content felt more like navigating a map?