The AI writing investors actually read

More has been written about AI in the last three years than any subject in our index, and most of it restates a model release with adjectives. A smaller group of writers do the work: read the paper, run the numbers, or talk to the people buying.

The technical end

Interconnects by Nathan Lambert is the best bridge between research and everyone else. Post-training, open models, evaluation, written by someone who does it.

SemiAnalysis for the hardware and supply side. Dense, occasionally paywalled, and the most reliable read on what capacity actually exists.

Stratechery for strategy. Ben Thompson's framing of who captures value in a model-plus-application stack has held up better than most.

The investor side

a16z publishes heavily on AI and the quality varies with the author; the infrastructure pieces are the ones to read. Tomasz Tunguz tracks what AI is doing to software metrics, which is the question most of his readers are actually being asked by their boards.

Clouded Judgement covers how public markets are pricing all of it, which is the sanity check on private enthusiasm.

What to skip

Model launch recaps. By the time a newsletter summarises a release, you've seen it three times, and the summary rarely adds a judgment.

Prediction posts with no mechanism. "Agents will do everything by 2028" is not a claim you can evaluate, act on, or be wrong about in a useful way.

Anything whose argument depends on a benchmark number without saying how it was measured.

Following it here

We tag every post by industry, so the AI writing collects on two pages as it comes in: AI for the broad pieces and AI infrastructure for compute, models and tooling. Both rank by how much the piece actually says rather than when it was published, so the top of the page is usually worth more than the top of your timeline.