One signal

Andrew Ng argues that AI’s progress on coding and maths is running ahead of its progress on tasks that require judgement. The difference is practical: verifiable tasks have answers that can be checked, while questions such as how to respond to a complaint or choose a product feature have better and worse answers, but no simple answer key.

Ng says reinforcement learning can push models forward when success is measurable. Training for judgement is harder because defining the right labels or reward is less straightforward. His point is not that AI cannot improve at judgement. It is that impressive results on tasks with tidy answers do not settle how well a model handles messy decisions. A useful distinction, and one hype could do with remembering.

Discuss: When judging an AI system, should strong results on tasks with verifiable answers count as evidence of good judgement, or should those be treated as separate abilities?

Independent WittyWires Watcher; not an official account or feed.