Who to Actually Follow: A Signal Map for AI

Article 01 ended with a shelf – the people and benchmarks I read to tell honest AI signal from sponsored noise. A shelf is a start. It is not yet a system: it does not tell you what each source is for, how often to check it, or who to tune out so the good signal can get through. This is the shelf turned into a working diet.

// in one breath

Why "who should I follow" is the wrong question, and the sharper one: what job is this source the best in the world at.
The five jobs a low-noise AI diet has to fill, and the one or two names that fill each – no duplicates, no padding.
A weekly rhythm and a short mute list, because the discipline is as much what you stop reading as what you start.

Open any feed in 2026 and the loudest voice on AI is selling something. A course that gates the real content. A tool reviewed by the channel it sponsors. A post promising a six-figure income from five apps and no experience. The signal underneath is real and it matters more than it ever has – which is exactly why it is worth the work of digging it out from under the people whose actual product is your attention.

before you trust anyone

The Incentive Test

Before a source earns a place in your week, run it through three questions. They are the same questions I ask of a vendor benchmark, pointed at a person.

// three questions, in order

What are they selling – and does the claim survive it? Everyone has an incentive; that is fine. The test is whether the insight still holds once you assume they want you to buy the thing. If the value evaporates without the sales pitch, it was the pitch.
Do they grade themselves? The strongest credential is not being right – it is publishing predictions, then scoring them in public, or demonstrating every claim with running code. People who show their corrections are people who have a process. People who only ever announce are performing.
Would this be worth reading with zero engagement? If a piece only makes sense as something designed to be shared – the screenshot, the hot take, the thread with a hook – it was built for the algorithm, not for you.

the map

Five Jobs, Not Forty Follows

You do not need forty AI accounts. You need five jobs filled by people who are genuinely the best at them. Here is each job, what it is for, and the sources I trust to do it. The one-liners I gave in Article 01 said who they are; these say what to use them for.

// Job 01

Calibrate the future

For timelines with error bars instead of vibes – and forecasters honest enough to mark their own homework.

// Job 02

Check a capability claim yourself

For the moment a model launches and everyone repeats the same chart – so you can go look at the number that was not chosen by a marketing team.

// Job 03

Cross the bridge from paper to practice

For understanding how models actually get good – the engineering behind the capabilities, without needing a PhD to read it.

// Job 04

See it built, with the code shown

For the difference between someone who talks about what AI can do and someone who shows you, running, on a real machine.

Simon Willisonsimonwillison.net

The most honest practical LLM engineering blog there is – every claim demonstrated with running code and real experiments. If one name on this map is non-negotiable, it is this one.

// Job 05

Read the rules being written

For the policy and governance layer that will shape what you are allowed to build – and the labs' own numbers, read with the discount they deserve.

A note on that last job: the labs themselves are worth reading, with a rule attached. Weight their operational numbers – the ones describing their own engineering reality – above their predictions about the future, which always point the same direction as their fundraising. I followed that rule when I cited the labs in Where This Is Going: their internal metrics earned trust, their timelines earned a discount.

the rhythm

A Diet, Not a Drip

Sources are half the problem. The other half is cadence – matching how often you check something to how fast it actually changes. Capability does not move hourly, so reading it hourly only buys you anxiety. This is the rhythm I keep.

// Daily

Nothing, on purpose. The feed is where signal goes to get buried. Skipping the daily scroll is the single highest-leverage move on this page.

// Weekly

One digest (Zvi or Import AI) and one practitioner (Willison). An hour that replaces a week of scrolling and leaves you better informed.

// Monthly

One bridge piece – Raschka, Lambert, or a Friendly Paper Review walkthrough – to keep the mental model current, not just the news.

// Quarterly

Check METR's time horizon. One number, one trend line. It will tell you more about where things are than three months of launch-day coverage.

the other half of the diet

What to Mute

A signal map is also a map of what to ignore. These are the genres I have learned to tune out – not because the people are bad, but because the format is built to convert attention, not to inform it.

// tune these out

The sponsored reviewer. If the tool pays the channel, the review is an ad with production values. Useful as a demo, worthless as a verdict.
The urgency merchant. "You are already behind. Buy now." Manufactured urgency is the oldest sales tactic there is; AI just gave it a new coat of paint.
The never-scored oracle. The permanent bull and the permanent doomer both make loud, dateless predictions and never grade a past one. Confidence without a track record is theatre.
The screenshot. A clever exchange with a chatbot is an anecdote, not a capability. One impressive output proves a model can, never that it reliably will.

// I believe this

In a field this loud, your edge is not how much you consume – it is how well you choose. Curate your inputs the way you curate your code reviews: a few trusted sources, read closely, beat a thousand you skim and a hundred you never should have let in.

That closes the series. You started at the lineage, walked through the engine room, the economics, the forecasts, and the wider risks – and you end here, with the small, durable habit underneath all of it: knowing whose signal to trust, and checking the rest yourself. Go back to Article 01 when a new model drops, and read it again with this map in hand. The tools will keep changing. A good information diet is how you stay the one deciding what that means.