Engineer picks マスターアルゴリズム and two other books to fix model-choice blind spots

The author's starting point is a specific failure: a demand-forecasting model built with scikit-learn performed well, but when asked in review why that model was used, he could only say he had followed a cheat sheet's branching logic. That gap — knowing the API but not the assumptions behind a method — is what the first book targets. マスターアルゴリズム, published in 2015 and written before Transformers, is presented not as a current deep-learning reference but as a map: five schools of machine learning, each with its own premise about what learning is, from combining knowledge to adjusting weights through backpropagation to updating probabilities.

The second gap the author describes is post-build: a model handed off without a plan for how frontline staff use its predictions, or who notices when they are wrong, was ignored within half a year — not because of accuracy. HUMAN+MACHINE is framed as a design document for that problem, covering role division between people and systems and human-in-the-loop routing, which the author says translated directly into operational flow notes. He notes it lacks implementation detail.

AI世界秩序 supplies the wider frame: the argument that AI's deciding ground has shifted from discovery to implementation, with data pipelines and speed mattering more than new methods, organized into four waves — internet, business, perception and autonomy. The author flags its 2018 publication date as making names and figures outdated, but says its point about data pipelines and implementation speed still holds for LLM app development. The closing advice maps each book to a specific current problem rather than treating them as a sequence.