I have a complicated relationship with roadmaps. The problem they solve is real. Someone deciding to become an AI engineer faces an unstructured pile of topics, tools and opinions, and the single most useful thing you can give them is an ordering. Embeddings before vector search.
Prompting before agents. Basic evaluation before you deploy anything a customer can see. This roadmap does that competently, and its focus is the right one, since it aims at applying pre trained models and existing tooling rather than training models from scratch, which is what the large majority of jobs with this title actually require. The interface helps.
Nodes are grouped, you can mark them complete, and each one links to resources. The site adds projects and other supporting material, and the whole thing is free to read. As a way to orient yourself in an afternoon, it earns its place. Then there is the failure mode, and it is a serious one.
A roadmap gives you the vocabulary of competence without any of the competence. I have met people who could recite every node on a diagram like this and had never built anything that survived contact with a real user. The map is not the territory, and the ticking of boxes is dangerously satisfying. If you use this, use it to choose one next thing, then leave and go build a small project that forces you to learn that thing badly and then properly.
Come back when you need the next choice. The volume of nodes is its own problem. Presented as a wall of boxes, everything looks mandatory and equally weighted, when in practice a few areas carry most of the value and several nodes are things you can safely learn on the job in an afternoon. Community maintenance produces this flattening, because consensus editing tends to add rather than remove and nobody wants to be the person deleting a topic they care about.
A single opinionated author would produce a shorter and more useful path, at the cost of being wrong in different ways. Resource quality behind the nodes is inconsistent too. Some links are excellent, some are the first tutorial someone found, and there is no signal telling you which is which. Treat every linked resource as a suggestion to verify rather than a recommendation.
My 3.6 reflects something free, well intentioned and genuinely useful in a narrow way, held back by the fact that its format encourages exactly the wrong behaviour in the people most likely to need it. Use it for twenty minutes of planning per month and ignore the urge to complete it. If you want the same sequencing with actual teaching attached, a structured course or a project driven curriculum will get you far further, even though the first thing you will do with either is check them against a diagram like this one.