Overview
Brian Christian (Norton) — Comprehensive examination of the alignment problem: how do we build AI systems that do what we actually want? Structured as a three-part journey through the history and future of aligning machine learning with human values.
Structure
Part I: Prophecy
- Representation — How models encode reality determines what they can perceive and what they miss
- Fairness — ML systems inherit and amplify biases present in training data
- Transparency — The interpretability challenge: understanding why models make specific decisions
Part II: Agency
- Reinforcement — Learning through reward signals; the challenge of specifying the right reward
- Shaping — Reward engineering to guide learning toward intended behavior
- Curiosity — Intrinsic motivation: can we make AI systems curious in productive ways?
Part III: Normativity
- Imitation — Learning from human demonstrations (behavior cloning, inverse RL)
- Inference — Drawing conclusions from incomplete evidence
- Uncertainty — The critical importance of knowing what you don’t know
Historical Foundation
Opens with the McCulloch-Pitts story (1943): Warren McCulloch (neurologist) and Walter Pitts (teenage logician) proved that simplified neuron networks could compute any logical function. Their “Logical Calculus of Ideas Immanent in Nervous Activity” launched the field — but real neurons proved much messier than their elegant model.
The Pitts paradox: The gap between theoretical elegance and biological reality that haunted Pitts in the 1950s mirrors today’s alignment gap between what we build and what we intend.
Core Insight
The alignment problem is not primarily technical — it’s about the gap between our ability to optimize and our ability to specify. We can build extraordinarily powerful optimizers, but we struggle to define exactly what they should optimize for.
Related
- situational-awareness-agi-trajectory — Aschenbrenner’s argument that alignment must be solved urgently
- reinforcement-learning-introduction — Sutton & Barto’s RL foundation that Christian examines critically
- ai-guide-thinking-humans — Mitchell’s complementary skepticism about AI capabilities