There is an enormous amount of writing about AI and almost all of it is downstream of a press release. Simon Willison's blog is the main exception I know of, and the reason is simple: he runs the thing before he writes about it. When a model is released he uses it, on tasks with known answers, and reports what happened including when it was unimpressive. That habit sounds unremarkable and is in fact extraordinarily rare, which is why this blog has become the de facto reference for a lot of working practitioners.
The most valuable body of work here is on prompt injection. Willison named the problem and has written about it consistently since, and the central argument, that you cannot reliably separate instructions from data in a language model context and therefore must design systems that do not depend on that separation, has held up completely. He has been especially clear about what he calls the lethal trifecta: an agent with access to private data, exposure to untrusted content, and the ability to communicate externally. Any two of those is manageable, all three is a data exfiltration channel.
I would make this required reading before anyone ships an agent with tool access, and I say that having seen several products that would not exist in their current form if their authors had read it. The model coverage is the other reason to read regularly. New release notes, hands on impressions, comparative testing, pricing analysis, and clear attention to what a model can be shown to do rather than what its announcement claims. He is careful with the epistemics, distinguishing benchmark results from vendor claims from his own observations, and he corrects himself in public when he gets something wrong.
He also covers local and open weight models seriously, which most commentary treats as a hobbyist sideline. Running models on a laptop is a genuinely useful skill for cost, privacy and understanding, and his writing on it is practical. The tooling he builds and documents, particularly the llm command line utility, doubles as teaching material for anyone who wants to work with models from a terminal. The honest limitations.
This is a blog. It runs chronologically, there is no syllabus, and a newcomer landing on the front page gets whatever happened this week with no context. There are tag pages and annotated link collections that help, and it is still fundamentally a resource for someone who already has enough background to place what they are reading. It assumes real development experience.
Code appears without apology, and the writing presumes familiarity with APIs, Python and the general shape of software systems. Someone learning what a language model is should start somewhere else and come here after. Volume is genuinely high. Several posts a week plus extensive link commentary is more than most people can absorb, and there is a real risk of substituting reading about the field for building in it.
My own approach is to skim weekly and read anything on security or evaluation properly. It is one person's perspective, and a well informed and openly stated one, but the coverage naturally reflects his interests: developer tooling, data, open models, security. Enterprise deployment, research methodology and the economics of the industry get less attention. My four point eight is a rating for what this is, which is the best free ongoing education available on practical language model work.
It loses a fraction for having no structure whatsoever, which is inherent to the format. Add it to your feed reader today.