There is a category of knowledge that does not appear in courses or papers. It is the accumulated practical understanding of how machine learning systems are actually built inside companies, which patterns recur, what breaks in production, and how the technical work interacts with the organisational reality around it. That knowledge usually transfers through working alongside people who have it. Eugene Yan writes it down, carefully and at length, and the result is one of the more valuable resources available to anyone doing applied machine learning.
The pattern surveys are the standout contribution. Posts that survey how many companies have solved a recurring problem, synthesising engineering blogs, papers and practice into a coherent picture, do something neither individual sources nor academic literature manage. You come away understanding the design space rather than one point in it, knowing which approaches suit which constraints and why. That is exactly the knowledge that is hardest to acquire and most valuable when you have it, and it is here for free.
The recommendation systems material deserves particular attention. Recommenders are commercially enormous and educationally neglected, sitting awkwardly outside the standard curriculum which prefers classification and generation. The writing here covers architectures, the evaluation problems specific to recommenders, cold start handling and the serving constraints that shape what is feasible. For anyone working in this area it is the best free material I know of, and it is better than several paid courses on the subject.
The tone is careful rather than combative. Claims are referenced, alternatives are considered, and uncertainty is acknowledged where it exists. In a space full of confident assertion this restraint is a virtue, and it means the writing holds up when you check it against your own experience rather than collapsing on contact. The organisational content is unusually good and unusually rare.
Working with stakeholders, scoping projects that can actually ship, deciding when machine learning is the wrong tool, communicating uncertainty to people who want a number, managing the gap between what a model does and what a business expects. These determine whether projects succeed at least as much as the modelling does, and almost nothing in the standard education touches them. Writing that takes them seriously fills a genuine gap. The structural limitation is the same as any blog.
Years of posts written independently, no ordering, no progression, no indication of what to read first. The site is well organised and it is still an archive rather than a curriculum. Anyone hoping for a guided path will need to build one themselves, and my honest advice is to start with the pattern surveys and follow your interests from there. The assumed background is real.
This is applied writing for people who already understand the fundamentals, and readers without solid grounding in machine learning basics will find the discussion moves past them. That is correct for the intended audience and it does mean this sits in the middle of a learning path rather than at its beginning. Older posts need dating. Material written before the current wave of large language models reflects a different landscape, and while much of the systems thinking remains entirely valid, some specific recommendations have been overtaken.
The reasoning about evaluation, serving and system design is durable. The tooling references are not always. Check when something was written before treating it as current. Breadth creates unevenness.
The site covers a lot of ground and some topics receive one excellent post that leaves you wanting more, with no follow up. That is the nature of writing driven by interest and experience rather than a syllabus, and it means this works better as a supplement to structured learning than as a replacement for it. Four point five for writing that captures the practical knowledge of applied machine learning better than almost anything else available, with genuinely outstanding material on recommendation systems and system design patterns, and a rare willingness to address the organisational reality of the work. The absence of structure and the assumed foundations are the only meaningful limitations, and for anyone working in applied machine learning this deserves a place in the regular reading rotation rather than an occasional visit.