As a note, i'm picky when it comes both to writing and thinking. And i hate most books written by academics. Even the ones with good information (eg, Fodor's Modularity) are hard to read and filled with confusing, field-specific words. Not this book. It's really well written. Written in plain English, very few assumptions, very thorough analysis, lots of self-criticism, lots and lots of data (OK, that part is boring and can be skipped, but it's comforting to know it's there)

What's it about? Common AI, psych and economic decision and learning algorithms (decision trees, neural nets, Bayes, multiple linear regression, etc.) are compared to several absurdly simple algorithms the authors believe real humans use. The various approaches are compared and evaluated on the basis of performance, accuracy on training data, accuracy on test data (generalization) and amount of input data required. Tests are on the standard UC Irvine data learning test sets. Comparisions, outcome explanations and relevance to the human mind and the real world are provided. Explanations and analysises are easy to understand and pretty convincing

i've decided to use a lot of what was in this book in my software, things that have made my agents more natural and easier to implement. i absolutely love this book

My own background is in philosophy, where this type of work has been very important in undermining the assumption that humans are rational. We aren't. You should probably read Kahnemann and Tversky's books before coming to this though, since this work adds an interesting spin to the old irrationality debate: maybe some of it is GOOD for us!