Lil'Log has been running since 2017 and it is the single best free resource I know for practitioners trying to keep up with AI research without reading fifty papers a month. Lilian Weng writes long, careful, heavily referenced posts that function as literature surveys, and the recent catalogue gives a fair sense of the range: Harness Engineering for Self-Improvement from July 2026, Scaling Laws Carefully from June 2026, Why We Think from May 2025, Reward Hacking in Reinforcement Learning from November 2024, and Extrinsic Hallucinations in LLMs from July 2024. Earlier work covers autonomous agents, prompt engineering, diffusion models for video, adversarial attacks on language models, and data quality for training. What distinguishes this from the enormous volume of AI content published daily is that the posts explain mechanisms.
A typical AI newsletter tells you a paper exists and roughly what it claimed. A Lil'Log post on hallucination tells you what the failure actually is, distinguishes the cases that get conflated, walks through the approaches that have been tried, explains why each helps and where each breaks, and leaves you with a mental model you can use on a problem the post never mentioned. That is teaching, and it is rare. The agents post is the one I have recommended most often.
It laid out planning, memory and tool use as the components of an agent system, with the relevant literature attached, well before most courses had anything coherent to say on the subject. Plenty of people who now build agents professionally formed their first clear picture of the architecture from that post. The reward hacking post is similarly the clearest treatment I know of a problem that anyone doing reinforcement learning from human feedback will eventually hit. The credibility matters here.
Weng spent years at OpenAI working on safety and applied research, and now works at Thinking Machines Lab. The posts read like someone describing systems they have actually debugged, which is a different register from someone summarising papers they have read. When she notes that an approach sounds good and does not work well in practice, that observation is worth a great deal. Now the honest limits, because I want to be clear about who this is for.
It is not for beginners. Posts assume you are comfortable with machine learning notation, that you can read a paper, that terms like KL divergence and policy gradient do not stop you. Someone six weeks into their first course will find this discouraging rather than useful, and there is no shame in bookmarking it for later. Second, it is not a curriculum.
Posting is irregular, driven by whatever the author found interesting, and there is no path through it. You cannot work through Lil'Log the way you work through a course. Treat it as a reference and a reading habit rather than a syllabus. Third, there are no exercises and no code.
Reading a survey of agent architectures produces the comfortable feeling of understanding without the ability to build anything. That gap is real and the only fix is to implement something after reading. I would suggest picking one technique per post and writing it yourself, badly, before moving on. Fourth, older posts age.
A 2019 post on a fast moving subfield is a historical document, and it is not marked as such. Use publication dates carefully. On alternatives and complements. Jay Alammar's illustrated posts are better for visual intuition and cover less ground.
Distill.pub was more rigorous and interactive and has been on hiatus since 2021. Sebastian Raschka's writing is closer to implementation. Read all of them. Lil'Log's particular strength is breadth at depth on frontier topics, and nothing else does that combination for free.
My four point seven is the highest score I have given a blog and I am comfortable with it. The deduction is only because it is not structured teaching and cannot be, and because the barrier to entry excludes exactly the beginners who visit a course review site most often. For anyone past that barrier, this is not optional reading.