The takeaway: DeepSeek has suspended its second funding round after remarks attributed to Liang Wenfeng went viral — and the leaked transcript behind it warns that its own numbers may be wrong. Meanwhile, three stories from Saturday deserve a second look: Congress came at the Pentagon's AI data centres three ways and only the quiet one landed, OpenAI moved medical records out of the health tab because most health conversations were happening outside it, and a wrapper took the same models from 43% to 99% on ARC-AGI-3 without touching a weight.

A leak with a disclaimer attached

Bloomberg reported on Saturday that DeepSeek had told some prospective investors in its second funding round it was suspending the deal for now, after comments widely attributed to founder Liang Wenfeng spread online. The pause "stemmed in part" from Liang's frustration at reports of what he had said during the first round — Bloomberg's hedge, not ours. It is unclear whether the company has told all prospective backers, and negotiations "remain fluid".

The striking part is the standing of the document at the centre of it. Bloomberg has not verified the viral posts. DeepSeek has not confirmed the transcript those posts concerned, and did not immediately respond to a request for comment. The transcript is a leak from a closed-door investor meeting held on 20 May, published by Tencent Tech roughly two months later; multiple Chinese outlets ran differing versions, and the WeChat links have since been taken down.

The document opens like this:

Audio "deepseek 0520.m4a", total length approximately 3 hours 44 minutes. This transcript was auto-transcribed from speech recognition and organized by AI, without distinguishing speakers… individual proper nouns and numbers may contain recognition errors, please refer to the original recording as authoritative.

That warning covers precisely the figures now circulating from it — a 4:1 Huawei-to-Nvidia chip ratio, a one-twentieth-of-the-compute claim, a two-year gap. All spoken numerals, all machine-transcribed, none confirmed by anyone. The filename, incidentally, is where the 20 May date comes from.

What the transcript argues is coherent enough: that one variable explains the field, and that "the talent gap is fundamentally a compute gap." What it does not argue is that money is beside the point. The same document names capital as a constraint in the same breath as chips, ties the previous round to option grants and retention, and Bloomberg says plainly that the current round is meant to fund "an ambitious expansion plan, including an increase in computing capacity."

The size of the round has never been agreed either. Bloomberg has DeepSeek seeking at least ¥10bn at a pre-money valuation of at least ¥480bn; Reuters had as much as ¥50bn at about ¥500bn. The two valuations may describe much the same thing — neither outlet states its basis — but a ¥10bn floor against a ¥50bn ceiling does not reconcile.

Our full piece is here.

Congress came at the Pentagon's AI data centres three ways

Two loud bills that would halt the military's AI buildout are going nowhere. The one that landed was narrow and quiet: a supply-chain provision on adversary-made components, which rode the must-pass NDAA through the House on 22 July — the bill passing 216–212 — over the Army's objection.

Restrictions framed as environmental lost; a restriction framed as China policy passed. The legible fight and the consequential one are rarely the same fight. And a provision that passes the House is not yet law. Read it here.

OpenAI moved medical records across your conversations

Health in ChatGPT is rolling out to US users, and the telling detail is why it changed shape. In OpenAI's own words: "Among people with access, more than 70% of health-related conversations took place outside the dedicated health experience." So connected records can now be drawn on across ordinary conversations — if you let them. US only, 18+, web and iOS.

A product decision driven by observed behaviour rather than design intent, with the privacy surface that implies. Read it here.

A harness, not a model

An Impossible Research wrapper reports about 99% on ARC-AGI-3's public set — self-reported, and never run on the held-back set at all, where the best ARC Prize-verified frontier score is 7.78%.

The authors are notably careful, and it is worth being specific about what they disclose: the results are unverified, they are public-set only, and the headline number comes from a fallback rule in which games scoring below 80 are rerun with a second model and the higher per-game score kept. They also decline the eighty-point comparison most coverage would have led with, offering instead a controlled one that holds the models fixed and changes only the wrapper — 42.83% to 98.98%. Same weights, same tasks, one variable. That is the finding, and it is the one least likely to reach a headline. Read it here.

The thread

Two of today's stories are about the distance between a number and what it licenses you to conclude: a machine-transcribed figure quoted as a founder's estimate, and a public-set score read as a held-back one. The Congress story is a different failure — not a misread number but a misread fight, where the restriction that passed was the one nobody was arguing about. In each case the record was available; the reading was the problem.