ai-for-software-engineers is twelve book links and a reading order nobody has to follow

The section marked optional is the one about engineering judgment

The list opens with a section of general resources and immediately downgrades it: consider this optional. That section contains a single title, Richard Hamming's The Art of Doing Science and Engineering, and the accompanying note explains why an engineer moving into AI should care about it.

The argument is two sentences. AI is fundamentally a research discipline, and continued learning matters as much there as it does in ordinary software engineering. Everything after that section is technical, so the one entry arguing that you need a research mindset is the one you can skip. That inversion is the author's own, not a summary, and it is the most opinionated thing in the repository.

The resource is also the only one offered twice. The full book sits behind a publisher's page, and the note points at a shorter essay version hosted by a university, free to read, for anyone who would rather not buy it. That free alternative is the single most useful practical detail in the whole file, because it is the only entry where the author has thought about cost.

The selection rule is that a book ships an accompanying repository

Scanning the twelve visible entries, seven point at a companion code repository: Géron's hands-on machine learning book, Danka's mathematics book, both Raschka titles, Lambert's RLHF book, and both Chip Huyen titles. Six of those seven repositories are explicitly described as containing notebooks, exercises, code or chapter summaries.

That is the filter. A title that asks you to read about building something, without giving you the thing to build, is mostly absent. The three exceptions are the free web resources, which are usable precisely because they are readable at no cost.

The effect on a reader is that this list rewards doing rather than reading. Each entry pairs a concept with an artefact, and the artefact is where the understanding actually lands. It also means the list ages at the speed of its slowest repository, since a title without code would have been excluded anyway. For someone deciding what to study next, the companion repository is the thing to check first, because it tells you whether the resource is still maintained.

Two authors appear twice and two publishers supply most of the list

Concentration is the other pattern. Sebastian Raschka is listed twice, once for building a language model from scratch and once for building a reasoning model from scratch. Chip Huyen is listed twice, for production machine learning systems and for engineering applications on top of foundation models. The reasoning-model title is the later addition, and it is the clearest sign that the list has been kept current rather than assembled once.

By publisher, four entries come from one house and three from another, with the remainder split across a third, a fourth and three free websites. Readers who take this list seriously are committing to two publishers' catalogues, which is a reasonable thing to do deliberately and an awkward thing to discover by accident.

The three web resources are the exception that proves the rule. A video series on neural networks and transformers, a free inference engineering text, and a scaling book from a research lab are all readable without a purchase. They are also the entries most likely to have changed since they were added, which is the trade a free resource makes with you.

The path runs from research judgment to serving, and stops before evaluation

The four sections form an argument about order. General resources cover judgment. ML foundations covers two books, one practical and one mathematical, the second being linear algebra, probability, calculus and optimization. The LLM section holds four entries and is described as the most important topic in modern AI: a visual series on attention and transformers, then three from-scratch titles covering tokenization and training, reasoning models, and reinforcement learning from human feedback.

The engineering section is the longest. It covers production machine learning systems where data distributions shift after launch, application-layer work on foundation models covering context, evals, agents and retrieval, data-intensive application design, inference serving, and large-model training.

What the sequence does not include is the part where most production work happens. Evaluation methodology appears once, as a topic inside one book's description. There is nothing on monitoring, on drift detection, on cost control, or on incident response. If your goal is to operate a model in production rather than to understand one, the list will get you most of the way there and then stop.

The repository is a reading list with no code and one folder of assets

The tree is four entries: a gitignore, the licence, the README and an assets directory. There is no source directory, no package manifest, no dependency file and no build configuration, which is consistent with a project whose primary language cannot be determined because it contains no code.

There are no published releases either. The last push is dated 26 August 2026, and the repository has no open issues. The licence is MIT, which for a list of links is more permissive than most people expect, though it is worth remembering that the licence covers the list and not the twelve books it points at, each of which carries its own terms and its own price.

So there is nothing to install and nothing to clone for its own sake. A reader who wants this content takes the README, and a reader who wants to be kept current subscribes to the newsletter the README ends by promoting, or follows the author's account on the social network the README links. The repository is the top of a funnel rather than a tool.

The author calls it a curriculum and then tells you to skip around

Two instructions sit awkwardly together. The README presents the material as four sections in a numbered order, which is the grammar of a course, and it says the guide assumes prior programming experience, which sets an entry point. Three paragraphs later it says to feel free to skip around between resources as you see fit, which dissolves the order.

The author is consistent about this elsewhere too. The mathematics book carries the note that most engineers skip this part, while arguing that the maths behind the models matters more as AI advances. The data-intensive applications book is flagged as not AI-focused, with the justification that it teaches concepts required for building AI systems at scale.

That candour is the list's real strength. A curated reading list that hides its own weak spots is marketing; one that marks its general-resources section optional and names the book most readers will skip is giving you the information to choose. The sequence is a good default order for someone with no map. It is a bad one to follow mechanically, and the README says so in the sentence most curated lists leave out.

The reading list is the free part of a commercial newsletter

The README closes by pointing outward three times: to an account on a social network for resources shared as the author finds them, to a newsletter subscription, and to a separate newsletter site. The project's own homepage sits on a plain http address, and the subscription link uses a second domain for what appears to be the same publication.

This does not make the list worse. It makes the arrangement legible. The repository is free, MIT-licensed and complete on its own terms, and it is also the top of a conversion path to a subscription. Both statements are true at once, and for an evaluator the useful question is whether the list would look different if it were not a funnel.

The answer, having read all twelve entries, is that it looks like a list written by someone who has worked through these books rather than one assembled from search results. The from-scratch titles come with code, the free alternatives are offered where they exist, and the weak spots are labelled. That quality is a property of the author's judgment, not of the business model, and it would survive the removal of the newsletter link.

Editorial conclusion

This list is a good starting point for an engineer who already writes software and needs a defensible order to work through, because the sequence from research discipline to model serving is sound and the entries are current rather than a pile of blog posts. Read it as a menu rather than a curriculum, since the author says so. Two things to know before you budget time: roughly two thirds of the list is behind a paywall, and the list stops short of evaluation, observability and cost, which is where most production work actually lands.

Frequently asked questions

Official sources

1. Issues
2. License: MIT
3. loganthorneloe/ai-for-software-engineers on GitHub
4. Project website
5. README