AI Engineering Blogs: Systems, Evals and Practical Work
Seven AI engineering blogs that help you decide what to measure, what to build and what to fix, with a reason to read each and a CSV export.
These are seven blogs that help with real engineering decisions: what to measure, what to build, and what to change when an AI system fails. If you read only one this week, start with Simon Willison, who writes up small experiments you can try yourself.
It's a small, hand-picked list, not a traffic ranking.
Last verified Sep 17, 2026. Recheck target: monthly. The CSV holds all 7 entries.
Start with five
Simon Willison for quick experiments, Hamel Husain for testing what your AI gets wrong, Chip Huyen for system design, Eugene Yan for applied methods, and Lilian Weng for research background. Before you assume a result will hold for your own work, follow the article's own sources.
Read him to see new releases actually tried: the prompt, the output, small programs and a link to the release. Good for deciding what's worth trying on your own machine.
Start here when an AI feature works in a demo but fails for real users. He focuses on reading through the errors, building realistic tests and deciding what to change based on them.
Read her for how data, evaluation and deployment fit together as one system. Especially useful before you pick the infrastructure around a model.
Read him to see how a way of measuring a model connects to whether the product works, with detailed examples from retrieval, recommendations and LLM apps.
Read her before you adopt the trendy name for a method. Each post explains the idea and cites the papers behind it. Save it for a slow study session, not a quick skim.
The directory
Picks favor writing that shows its methods, names its tradeoffs and links its evidence. These writers don't post on a schedule, so the cadence on each entry is a suggested reading rhythm, not a promise from the author.
Each note says what was checked. A dated article is a good place to start, not necessarily the newest post.
Read him to see new releases actually tried: the prompt, the output, small programs and a link to the release. Good for deciding what's worth trying on your own machine.
The checked feed identifies Simon Willison's Weblog. Its XML parsed successfully; the activity date is the newest item timestamp supplied by the publisher, not a check of every article's claims.
Start here when an AI feature works in a demo but fails for real users. He focuses on reading through the errors, building realistic tests and deciding what to change based on them.
The feed parsed as Hamel’s Blog. Eight of 20 items, including the two newest, link to the homepage; use the website to select the article. The activity date is publisher supplied.
Read her for how data, evaluation and deployment fit together as one system. Especially useful before you pick the infrastructure around a model.
The public feed parsed as Chip Huyen, but its latest returned item is from January 2025. This is feed activity evidence, not evidence that the website has had no newer work.
Read him to see how a way of measuring a model connects to whether the product works, with detailed examples from retrieval, recommendations and LLM apps.
The checked feed identifies Eugene Yan. Its XML parsed successfully; the activity date is the newest item timestamp supplied by the publisher, not a check of every article's claims.
Read her before you adopt the trendy name for a method. Each post explains the idea and cites the papers behind it. Save it for a slow study session, not a quick skim.
The checked feed identifies Lil'Log. Its XML parsed successfully; the activity date is the newest item timestamp supplied by the publisher, not a check of every article's claims.
Read this when your problem is serving models, not prompting them: reusing cached work, moving memory, running work in parallel and response time. Before you plan capacity off one of their benchmarks, make sure the model, hardware and workload match yours.
The September 10 article explains tiered KV-cache offloading, implementation choices and a defined benchmark setup. This directory does not reproduce its performance results.
Real agent design choices and what went wrong, from the team that builds Claude. Treat their examples as one way to do it, built around their own model and setup.
The inspected November 26, 2025 article describes initialization, incremental work and testing across context windows. This is a dated example from the archive, not the blog’s latest publication.
Choose by the question you have
| Your question | Start with | Read it for |
|---|---|---|
| Why is this feature failing for users? | Hamel Husain or Eugene Yan | Finding the errors, designing test tasks and the evidence to back a fix |
| What should I try with a new tool? | Simon Willison | A small experiment you can inspect, and its limits |
| How does this research method work? | Lilian Weng | How it works and the original papers |
| Where does this model fit in the application? | Chip Huyen | System design and where each part can fail |
| Why is inference slow or short on memory? | vLLM | Hardware, workload, cache behavior and measured tradeoffs |
| How should a long-running agent keep its progress? | Anthropic Engineering | Saving state, testing and recovering across sessions |
Chip Huyen's checked feed ends in January 2025. Visit the site directly before deciding she's stopped publishing. The Anthropic entry points to a useful 2025 article, not the latest post.
Turn one article into an experiment
Pick a current problem before you open your reader. For an unreliable assistant, that might be whether the search step missed the right document or the model ignored it.
Read one relevant article and write down what it says is going on. Test that against a small set of real failures. Save the result next to the article link, including where the advice didn't fit.
That habit beats subscribing to every launch feed. Revisit the list monthly and drop sources that no longer change a decision.
The RSS and OPML pack gives you six working subscriptions, with alternatives where a publisher's feed is stale or its links don't work. For lab announcements and news coverage, use the broader AI blogs and news directory. To move from reading to code, pick one agent repository.
Change log
Added serving and agent-harness engineering sources, bringing the selection to seven. Added dated source evidence and distinguished a stale feed from a quiet or inactive website.
Added a focused selection with primary-source links, a start-with-five and an export.
Follow changes to every directory with the directory changes RSS feed.