People to Follow

Below is the curated set. Each entry names the specific contribution the person is known for, why that earns a follow, and what to read or watch first so the follow pays off on day one. Read it as a starting shelf, not a complete map: the point is that it is short and every slot is justified.

Simon Willison — runs the most consistent hands-on commentary on what actually works with LLMs, ships real experiments rather than takes, built the open-source llm command-line tool, and drew early attention to prompt injection as a security class before most of the field treated it as one. He earns the top slot because the signal is reproducibility: he posts the experiment, the prompt, and the result, so you can run it yourself instead of trusting a screenshot. Start with his ongoing writing on context engineering and his prompt-injection coverage, which named the attack category most people still underrate. Where the follow stops carrying you: the volume is high and the register is exploratory, so it is a running notebook of what one practitioner is testing, not a settled curriculum. Read it to see what is worth trying, not as a syllabus to complete.

Hamel Husain — the evals-first voice. His essay "Your AI Product Needs Evals" is the canonical practitioner argument for measuring AI output instead of eyeballing it, and his repeated insistence to "read your data" is the single most underrated discipline in the space. Evaluation is where most beginners stall, and his material is the cleanest on-ramp into it. Start with "Your AI Product Needs Evals" and his framing that the most valuable eval cases are your own production failures, not a generic benchmark. Where it stops: the material is pitched at people building AI products with a real failure stream to mine, so if you are still operating tools rather than shipping a system, take the mindset (measure, read your outputs) and leave the heavier harness machinery until you have failures worth labeling.

swyx (Latent Space) — maintains the AI Engineer Reading List, a curation of roughly fifty resources organized by topic, and runs the Latent Space podcast, which together act as the de facto hub for the field. The reading list is itself proof of the principle this page runs on: its value is the selection, not the enumeration, which makes it the fastest single jump from zero to a grounded map of what to read next. Start with the AI Engineer Reading List and work down the topics that match what you are building. Where it stops: a reading list is a map, not the territory, and any fast-moving link in it ages, so treat it as the index that points you onward rather than a finished course.

Cognition — the team behind the engineering essay "Don't Build Multi-Agents," which argues that for most applications today a single-threaded linear agent with carefully engineered context is more reliable than a multi-agent system, because parallel sub-agents make conflicting implicit assumptions and need the full agentic trace rather than summaries to stay coherent. They earn the slot because most public agent content sells complexity and this is the rare hard-won counter-position. Start with "Don't Build Multi-Agents." Where it stops: it is a strong opinion grounded in their architecture, not a universal law, so read it as the case for restraint you weigh against your own task shape, not a ban on ever running agents in parallel.

Drew Breunig — produced the most teachable taxonomy of how long context fails, naming four distinct modes (poisoning, distraction, confusion, and clash) and pairing each with a repair pattern. A named failure taxonomy is exactly the kind of citable, reusable framework that turns a vague worry into a checklist, which is why his write-up keeps getting cited. Start with his write-up of the four context-failure modes and their repairs. Where it stops: it is a diagnostic vocabulary, not a tool you install, so its value is letting you name what went wrong, after which the fix is still yours to apply.

Jason Liu — known for structured outputs and the instructor library, the reliable way to get typed, validated objects out of a model instead of begging for JSON in a prompt. He earns a slot the moment your AI work has to hand structured data to another program. Start with his work on structured outputs and instructor. Where it stops: it is a sharp, narrow contribution, so it is the person you follow for one specific problem rather than general orientation.

A few names belong here as depth pointers rather than fully vetted slots, because the specific thing to read first is yours to pick once you are building: Eugene Yan and Chip Huyen on production ML-systems engineering, and Lilian Weng on the mechanics of how agents actually work. Follow them when you cross from using AI tools into building a system that has to keep running, not for day-one orientation.

A follow list is only as good as the rule that keeps it short. Paste this as the vetting line before you add anyone, including the entries above when you revisit them:

The list works the same way in reverse. When an account stops shipping and starts restating, drop it, the same way an operator disconnects a server that no longer earns its place.