An editorial note: what DataSynth & Co. is for

A great deal of writing about machine learning is either marketing in lightly-academic costume, or marketing in lightly-journalistic costume. The first is what AI labs publish about themselves; the second is what most technology press publishes about the labs. DataSynth & Co. is, with effort, neither.

We are a small editorial operation — currently one editor, occasional named contributors — that publishes a few well-edited pieces a month at the intersection of three things: how machine-learning research actually progresses, how the systems built from that research behave when deployed, and where the capital funding the whole enterprise is coming from and going to. Roughly: science, engineering, and the markets that finance both.

6/mo

Approximate publication cadence. We'd rather publish six pieces that hold up than thirty that don't. Stories that need additional reporting wait; stories that can't be supported get killed. Editorial standards · revised May 2026

What an article here is allowed to claim

Every article we publish should be readable as a self-contained argument. That means three things, mechanically:

The numerical claims are sourced inline, with the source named (paper, dataset, interview) and dated. Where the source is our own analysis, the methodology is described well enough that a competent reader could reproduce it.

The argument's load-bearing structure — the two or three claims that the rest of the piece rests on — should be findable in the first 200 words. We do not bury the thesis. Decks exist for a reason.

When we make a forecast, it carries a date and a falsifiable form. "AI labs will adjust their cap-ex by 2027" is not a forecast. "At least one of the three named labs will reduce announced cap-ex commitments by ≥ 20% before end-Q2 2027" is.

What we don't cover

We do not write about model releases as news. There are competent trackers for that, the best of which is ai-tldr.dev. We do write about what happens after a model release — what the benchmarks actually say, what changes in production, what gets quietly walked back.

We do not write opinion essays on AI policy, AGI timelines, or the alignment debate. There is plenty of good and bad writing in those genres elsewhere; we have no particular comparative advantage.

We do not publish sponsored content. Our companion sites (pomegra.io for markets, ai-tldr.dev for tracking) operate independently; cross-links between them are editorial, never paid placements.

Errata, updates, and reader corrections

The web makes it embarrassingly easy to silently rewrite history. We try not to. When we get something materially wrong, we publish a dated Correction note at the bottom of the piece, leave the original error visible (struck through), and credit the reader who flagged it. When new information meaningfully changes our analysis, we publish an Update rather than rewriting the original.

The tip line for both is [email protected]. Anonymous tips are welcome and respected; please include enough context for us to verify independently.

A note on style

We use serif type because we want you to spend time with the writing. We try to lay charts out so they can be read without zooming. Headlines are not designed to be optimised for click-through, which is why some of them are quite long. None of this is virtuous; it's just what the publication is.

If you find any of this useful, the easiest way to support us is to subscribe to The Brief — one email a week, no tracking, no upsell — or to write back when we get something wrong. Both are equally welcome.

— Lena A. Petrova, Editor · May 2026