The Alignment Problem: Machine Learning and Human Values

Why it's worth it

Brian Christian traces the whole problem of getting machine learning systems to do what we actually meant, from biased hiring tools to the researchers chasing fixes. It's the most substantial book on this list and the one that best explains the systems in the news.

Christian writes like a reporter. He goes into the labs, sits with the researchers, and builds each idea out of a specific mess — a hiring tool that learned to prefer men, a classifier that learned something ugly about faces. You get the technical concept and the human consequence inside the same story, which is why the ideas stay put.

If you want to understand today's chatbots, start here. The book covers learning from human feedback, reward hacking and the plain difficulty of specifying what you want — the exact machinery behind the systems everyone argues about. Nothing else on this list gets that close to how modern models are trained. Readers who work in machine learning say they keep it around as a reference, which is unusual for a trade book.

From our review of 10 books for understanding ai, by David Sinclair, Managing Editor

What we like

• +Christian writes like a reporter
• +If you want to understand today's chatbots, start here
• +It's long. Nearly five hundred pages, and the middle stretch on the history of machine learning is where reviews say people bog…
• +It names the problem better than it solves it

What to know

• –The density is real. Owners who loved the book still describe the early chapters as a drag, and the technical detail runs…
• –It also leans hard on failure