Inside the AI (How AI Works) (Books column)

I read the book “Inside AI” on the way to Vladivostok. How It Works by Ronald Kneuzel, who explains complex topics literally on her fingers. How AI Works: From Sorcery to Science 2023 year, and the translated book was published in 2026 year. The author himself, Ronald T. Kneusel, is a date Satanist who has experience in the medical domain.

It seemed to me that the author set himself the task of explaining AI so that the reader would stop perceiving AI as magic. To do this, Ronald starts not with ChatGPT or transformers, but with older and more useful classics: regression, decision trees, random forests, the nearest-neighbor method, and support vectors. This is not very fashionable, but the approach to gen AI is consistent, not a jerk.

I liked the explanation of interpolation and extrapolation when training the model with existing examples. Conventionally, the model sees the training sample, adjusts the parameters, tries to catch a stable pattern. And then the most interesting thing begins - inference on new data, which in the training sample was not. And here comes the practical question: does the model work inside the familiar area or not? In summary, the author shows that it is useful to think through interpolation and extrapolation: - Interpolation This is when a new example is similar to what the model has already seen, and it needs to restore the value between familiar dots. - Extrapolation. When we ask the model to move the pattern to where there was almost no data. In real engineering work, it is here that unpleasant surprises are often born: everything looked decent on tests, and in new conditions the model began to confidently respond without good support.

After that, the author gradually moves to neural networks and MLP. (multi layer perceptron) It shows the mechanics well: simple computing blocks, weights, layers, error, learning, repetition. This, of course, does not replace the course on deep learning, but gives the right intuition: the neural network does not “understand” the world in the human sense, but adjusts a large number of parameters for examples. Separately, the chapter on CNN is well presented. The author shows how the image is parsed through local features, filters, layers and the gradual assembly of a more complex representation. After this chapter, you start looking at computer vision more systematically, not just through math.

At the end, he goes to GenAI, where he looks at generative adversarial networks, diffusion models, large language models, and why text systems like ChatGPT have had such a strong effect. The book stops at information. 2023 It is a bridge to the topic, rather than an actual story. Therefore, if you need a fresh analysis of agents, RAG, MCP, evals, surveillance, production AI stack and how to implement all this in the company, then you should look for another book.

In total, the book simply explains the base and reminds that before talking about frontier models, it is useful to understand what training sampling, generalization, error, retraining, signs, parameters and inference are.

#Books #AI #ML #DeepLearning #GenAI #Engineering #Software #Math