Two years of progress in one year. That's METR's headline number for Anthropic, published Thursday after an 11-week investigation inside the company. The number didn't make me stop and reread. The sentence METR posted next to it did: Anthropic redacted "a majority of the evidence we found most informative."
That's the thread through everything worth reading this weekend. In every story below, someone is asking the rest of us to trust a claim we can't check.
Anthropic let METR in, then decided what you get to see#
Start with the substance, because it's remarkable. According to The Verge's breakdown of the METR report, Anthropic's models are improving roughly twice as fast as a year ago. METR says it is "confident" the speed-up isn't just more training compute, and that the majority of the acceleration as of October "is due to AI." The numbers that stuck with me:
Employees using Anthropic's strongest internal model self-reported being about 12x as productive, roughly double what they reported a few months earlier on Mythos 5.
Three unreleased models (codenamed Model 2, 3 and 4 for METR) finished training in July, August–September and September–October.
METR found no clear signs of diminishing returns or bottlenecks that would prevent "even faster acceleration."
Most of the gain still comes from AI helping human researchers rather than working alone, but METR warns that "there are some signs that fully autonomous contributions will play a larger role soon." Last week Anthropic said in its own filings that it was very close to self-improving AI. This is an outside evaluator saying the curve is already bending.
Then comes the redaction. METR wrote a ~100-page report, and Anthropic stripped out most of the evidence METR found most useful, arguing it would help rivals. METR's reply was that most of what was cut was "very high-level," and that the public now has to lean on METR's judgment "much more than would be ideal." Anthropic also cancelled the part of the assessment that would have checked how much compute goes to safety versus capabilities.
“We redacted a lot, and I accept that this makes it harder for others to check METR's work.” — Sam McCandlish, Anthropic cofounder, on X
I actually find McCandlish's argument coherent. If you believe the details of what's driving the speed-up are a roadmap, publishing them speeds everyone up. But that logic only holds if you accept Anthropic's premise that Anthropic should be the one going fast. Daniel Kokotajlo of the AI Futures Project put it less politely, saying it is "pretty rich" to cite rivals going faster when Anthropic does "not appear to be slowing down much at all." Ryan Greenblatt, who now works at METR, said he "was seriously wrong about RSI speed up."
What does this change for builders? The tools you use today are likely a generation or more behind what the labs run internally. That gap is now measurable, and it's widening. If your roadmap assumes model capability roughly flat through mid-2027, I'd throw that assumption out. Anthropic says its own report on the pace of progress is coming "in the coming weeks." I'll read it, but I'll weigh it knowing who chose what to leave out.
Nvidia's H200 license list includes an office with wires hanging from the ceiling#
Same pattern, different industry. The New York Times visited New Matrix, a Chinese firm the Commerce Department cleared to buy Nvidia's H200 chips, and found "a hollow gray shell with wires dangling from the ceiling." As The Verge summarizes, the company is 17 months old, its owners have ties to a Chinese private equity firm blacklisted since 2024, and the Times describes this as a pattern, not a one-off. The whole export regime rests on the Trump administration's deal giving the government a 15% cut of sales. That gives the regulator a direct financial stake in approving buyers.
Then there's the strange subplot. TechCrunch reports on a US intelligence memo claiming China's Ministry of State Security recruited Megaspeed executive Alice Huang to spy on Jensen Huang. Nvidia says their interactions were "not unusual" and that it doesn't share classified information with customers. Fine, but export control works only if someone actually checks the address.
Meta blinks on scraping publishers through their Facebook pages#
Meta has indefinitely paused a commercial terms update that would have required business page owners to let Meta crawl the external sites they link to for AI work. The terms even required publishers to create test credentials so Meta's bots could get past paywalls. Courts in Brazil and Norway had issued injunctions, the UK's CMA opened a formal investigation, and Senators Schiff and Blackburn both asked Zuckerberg to withdraw it. The best line came from Theo Bamber of the UK News Media Association: "Does Meta think we can't read?" Swapping "train" for "develop" didn't fool anyone.
This one ended well because the terms were public and readable. Compare that with the METR report.
The politics just got a lot less patient#
The reaction is arriving faster than the industry planned for. On TechCrunch's Equity, the hosts note that voters approved almost every local measure restricting data centers and rejected almost every one that would speed them up. In Texas, Gina Hinojosa beat Greg Abbott after campaigning on data center limits. Musk reportedly gave half a billion dollars or more to Republicans and "seems to have basically nothing out of" it. And in direct response to METR, New York Assemblyman Alex Bores called on Congress to "ban RSI, this year if possible."
A year ago that sentence would have been fringe. Now it comes from someone shaping the agenda of the party about to hold power.
Also today#
AI decompilation is coming for GTA VI. Former Rockstar tech director Obbe Vermeij thinks AI-powered reverse engineering of the PS5 build could cost "tens of thousands of dollars in tokens" and might push Rockstar to ship the PC port sooner. Effort used to be the moat. That moat is gone.
Alex Stamos, now CSO at Cognition, told Decoder that AI-boosted attacks are pushing security into a "dark age." That's the defensive side of the same capability curve METR just measured.
Trust is now the thing in short supply#
Here's my prediction. Within six months, "independent evaluation with redaction rights for the lab" will stop counting as oversight in any serious policy debate. Bores is already there. Evaluators like Apollo are already pushing to be embedded inside labs. Anthropic is about to go public, and S-1 disclosure rules don't care about competitive sensitivity the way a voluntary METR agreement does. My bet is that the next time we learn how fast frontier AI is moving, it will come from a filing or a subpoena, not a blog post.