NY Post: METR, the Watchdog, Has $71 Million and EA Roots

The New York Post ties Amodei's proposed evaluator to his old circle; regulators sound skeptical.

The New York Post put METR, the evaluator Dario Amodei wants embedded inside frontier labs with 'employee-like access,' under a full audit of its own. METR spun off from the Alignment Research Center run by Paul Christiano, Amodei's former housemate and collaborator. Founder Beth Barnes, hired by Christiano, chief scientist Hjalmar Wijk, and staffers Megan Kinniment and Pip Arnott all came from one Oxford research cluster; staffer Ajeya Cotra is married to Christiano.

The paper's money trail: METR announced $71 million in new funding in August over the prior six months; Jaan Tallinn's Survival and Flourishing Fund earmarked up to $752,000 for METR, including a $428,000 matching pledge; and Moskovitz's Coefficient Giving gave $1,515,000 to METR's incubator, the ARC, in 2022. METR told the Post it takes no funding from AI companies or their employees. Perry Metzger of Alliance for the Future called the arrangement 'a complete joke.'

Official Washington kept its distance. FTC Chairman Andrew Ferguson said that when companies ask for regulations and an antitrust exemption, 'all of my alarm bells go off,' and a source told the Post that 'METR is essentially Anthropic in another form.' @hughhewitt called the piece 'the story that killed Anthropic's credibility in ten minutes of reading.' Amodei's essay, the paper notes, warns a capable swarm could take over the internet within six to 12 months.

From posts by nypost.com, @hughhewitt

Agents & Coding

OMP With Local Models Wins Over Another Convert

@0xSero praises Oh My Pi on local models; replies point to the omp-ninfer repo with Qwen3.8 27B.

@0xSero reports OMP with local models is 'so good,' crediting the dev, whom replies identify as @_can1357. @devops199fan points to alphastorm's omp-ninfer, which pairs NInfer with Qwen3.8 27B on a single machine for durable local inference. @theadriann, four months in on every model he could try, calls it easy to customize as a pi fork. @LiamAITools's verdict: 'Local plus a harness beats another Max seat.'

From @0xSero

Models & Hardware

DeepSeek v4.1 Flash: Faster on Two Sparks, Ahead of TP4

@MiaAI_lab lifts two-Spark prose to 42 tok/s; @WescheNex1q's frozen benchmark scores TP2 91.2/100.

@MiaAI_lab's update pushes DeepSeek v4.1 Flash on two DGX Sparks to 42 tok/s on single-stream prose, up from 32, and 63.7 tok/s on two streams, up from 42.5, which the lab says matches what 3 to 4 Sparks did before. @WescheNex1q's frozen benchmark, 76 scenarios x 2 repeats with thinking off, scored the TP2 EXL3 build 91.2/100 against 87.7/100 for TP4. Dropping the one runaway task still left TP2 ahead, 91.64 to 88.16, on half the machines.

From posts by @MiaAI_lab, @WescheNex1q

48 of 49 Uncensored HF Repos Still Download

@yume_arasaki audits 49 uncensored Hugging Face repos; 26 of the 48 living ones now sit behind gates.

@yume_arasaki checked the live status of all 49 uncensored repos on his list on September 16. One is dead: Heretic's Llama build, taken down May 21 after a legal notice from Meta. The other 48 still download, and 26 of them now sit behind 'research only' access gates that did not exist two years ago. He also cites OrcaRouter's dense 27B abliteration card: over-refusal on benign prompts fell from 5.6% to 0.4% while MMLU rose 0.4.

From @yume_arasaki

GLM 5.3 Flash on Two Sparks Hits a Concurrency Wall

@rwojo counts about 39k usable cached tokens of 382k, so real work thrashes past two sessions.

@rwojo asked who runs GLM 5.3 Flash on two Sparks past one or two concurrent requests; with DFlash2 he sees about 39k of 382k cached tokens usable. @realryanpurdy calls three Sparks the sweet spot, and @kesslerIO gets 4 to 5 sessions at 25 to 40 tok/s on Qwen 3.8 Flash. @jmacftw, whose team builds sparkglm, says TP-2 concurrency is tough and they are working on it. @jondecker76 reports a 10-person team averaging over 500M tokens per day on TP3.

From @rwojo

Industry

FCC Chairman Mocks the Doomers as Net-Neutrality Reruns

@BrendanCarrFCC recycles an old fear list; @JASONMCNAB calls the pace debate a play against home clusters.

FCC Chairman Brendan Carr (@BrendanCarrFCC) says AI doomers are the same people who proclaimed the end of the Internet during the net neutrality debates. @JASONMCNAB argues open-weight recipes on Hugging Face and GitHub are eating closed models, so the labs run fearmaxx with rogue swarms, a 10% kill risk, and 'pace the frontier,' then write federal rules only OpenAI, Anthropic, and Google can afford. His close: 'If you fine-tune and serve your own weights, you are the competitor they want regulated out.' Home clusters, he writes, are the target.

From posts by @BrendanCarrFCC, @JASONMCNAB

An Engineer Reads Anthropic's Week as IPO Audit Prep

Matthew Chang says a PCAOB audit ahead of an S-1 explains the revenue and spending surprises.

Matthew Chang (@MatthewChang), who says his own company has been through PCAOB audits, reads Anthropic's behavior as IPO preparation: file the S-1, pass the audit, let early investors exit while index funds pick up the stock. He speculates NVIDIA discounts and Microsoft cloud credits were booked as revenue, and writes 'coupons are not revenue.' He discloses no Anthropic holdings and no inside information; the post is his interpretation.

From @MatthewChang

Sources

Tuesday, September 15, 2026 Back to the Archive Thursday, September 17, 2026