Nvidia H200 Licensees Turn Up at Empty Offices as an AI System Proves a Quantum Conjecture and Nous Strips State Framing From Qwen
The weekend’s AI news was about checking claims. The New York Times found empty offices behind two Chinese firms cleared to assemble Nvidia H200 chips. PSI said its AI system, running a non-public Anthropic model, proved the Aaronson–Ambainis conjecture with a re-checkable Lean proof. Nous Research released Arete-27B, a Qwen fine-tune that strips most state-aligned answers. Plus: OpenAI and Hacktron, the data center backlash, a push to ban recursive self-improvement, and Shizuku-Duplex.
Two Chinese Firms Cleared to Handle Nvidia H200 Chips List Empty Offices
An investigation by The New York Times, republished on Sunday by Singapore’s Business Times, found that New Matrix, a Chinese company the Commerce Department approved this year to assemble Nvidia’s H200 chips into AI servers, lists an empty address in a Beijing industrial park. In the Business Times reprint of the Times investigation into US-cleared H200 licensees, the building directory did not list the firm and the park’s property managers had no record of it. New Matrix was created 17 months ago, and its website names no staff.
According to Chinese corporate records cited by the Times, two of New Matrix’s primary owners appear in companies connected to Wise Road, a private equity firm placed on the Commerce Department’s entity list in 2024, and two analysts said New Matrix may be a front for firms tied to that fund. It is the second approved assembler without an apparent physical presence: in September the Times found the listed office of Maginfra, which analysts believe is an offshoot of entity-listed Inspur, apparently deserted. Two other licensees have ownership ties to suppliers of China’s military.
The approvals follow President Trump’s December 2025 decision to let Nvidia sell the H200 to China, with the US collecting a 25% tariff. The Verge’s summary of the empty-office findings frames it as a vetting gap at both Commerce and Nvidia. Liza Tobin, a former China director at the White House National Security Council, told the Times:
“I have a tonne of concerns that neither the Commerce Department or Nvidia is doing the due diligence needed to ensure that the licensing terms are being followed”
This is a compliance story as much as a geopolitics one. Anyone selling, reselling or hosting advanced accelerators should expect end-user screening to tighten, and should check counterparties beyond entity-list matches: physical presence, beneficial owners and ties to listed parents. A license will not shield a reseller from reputational fallout.
PSI Says an AI System Proved the Aaronson–Ambainis Conjecture, Checked in Lean
Physical Superintelligence (PSI) said on November 6 that a research system it built internally, running instances of a non-public Anthropic model, proved the Aaronson–Ambainis conjecture, open since 2008. According to PSI’s research post announcing the Aaronson–Ambainis proof, the system had a complete proof on the third day of its run and the Lean proof assistant accepted it. The result means quantum computers can be at most polynomially faster than ordinary ones on yes-or-no questions about random inputs; dramatic speedups like Shor’s algorithm need structured inputs.
The public GitHub repository with the paper and Lean 4 formalization says the main theorems are proved with no placeholder or added axiom, re-checked by two independent Lean kernels, and that the proof, paper and formalization were produced by AI models. PSI’s David Turturean, who led the work, says he is responsible for the paper’s correctness. Stanford researchers who settled a special case in August are writing a shorter exposition and wrote to PSI:
“This is a huge result. It resolves a decades-old conjecture with important implications for the nature of problems that admit quantum speedups.”
Separately, Google DeepMind’s Pushmeet Kohli wrote early Sunday that the lab is “witnessing a genuine phase transition in the pace of scientific discovery.” In Pushmeet Kohli’s X post teasing upcoming DeepMind science releases, he promised a glimpse “in the next few weeks” but did not say what will ship.
Machine-checked proofs change the economics of trusting AI mathematics: the expensive step shifts from verifying every line to confirming the formal statement matches the real question. Expect that pattern to spread into engineering, with humans writing specifications, models searching for solutions and a checker, not reviewer stamina, enforcing correctness.
Nous Research Releases Arete-27B, a Qwen Fine-Tune That Cuts State-Aligned Answers
Nous Research released Arete-27B, an Apache 2.0 open-weight model built by post-training Alibaba’s Qwen3.8-27B to remove refusals, deflections and Chinese Communist Party-framed answers on sensitive topics. Nous Research’s Arete-27B write-up describes an 803-prompt evaluation across seven categories, graded by an LLM judge for factual recall and for adopting the state’s account. Answers containing propaganda fell from 22% for the base model to about 1%, and the mean fact score rose from 2.32 to 2.98 out of 5.
Training was modest: about 52.8 million tokens of supervised fine-tuning, with no reinforcement learning or architecture change, using Thinking Machines’ open Inkling model as teacher. The Arete-27B model card on Hugging Face reports a 0.9% propaganda rate and an essentially unchanged Hermes Bench score, 65.60% against 65.77%. The figures are Nous’s own, though the evaluation, data and weights are public. Nous concludes:
“The values in any open model are only ever one fine-tune away.”
Arete cuts both ways for enterprise buyers. It shows that state-aligned behavior in strong Chinese open models can be measured and largely removed cheaply, but it also proves that any downloaded model’s values can be quietly rewritten. Teams adopting open weights should run their own values and refusal evaluations on every checkpoint, including fine-tunes from trusted labs.
OpenAI Brings In Hacktron to Attack the Seams Between Its Systems
OpenAI announced a partnership with Hacktron, a firm that builds AI agents for code review, penetration testing and red teaming, to find vulnerabilities in OpenAI’s own systems. Per OpenAI’s announcement of the Hacktron security partnership, dated November 6, the teams will jointly pick high-priority areas, focusing on how OpenAI’s systems interact with third-party services, and Hacktron gets expanded frontier-model access through the Daybreak Red program. OpenAI met the team through its bug bounty program, and the deal follows this week’s higher bounty rewards.
The focus on third-party connections is the telling detail. As assistants gain connectors, OAuth grants and agent-to-agent calls, the riskiest bugs live between systems that each passed review. Map every integration your AI tools touch, test token scopes and data handoffs explicitly, and ask vendors who audits those boundaries.
A Study and Vitalik Buterin Reframe the Data Center Backlash
The fight over AI data centers sharpened after Tuesday’s elections. The Washington Post’s analysis of what is really raising electricity bills reports that voters restricted data centers in 40 of 41 local ballot races, and points to a study finding a marginal effect on household bills. That working paper, by Vishrut Rana, Christian Kaps and Serguei Netessine of Wharton and Harvard Business School, attributes about 4% of the 2020–2025 residential price rise and 15% of the industrial rise to new data centers. As described in FourWeekMBA’s summary of the Wharton and Harvard Business School working paper, co-location builds drove the effect, not Big Tech. It is not peer reviewed.
Developers are adjusting their pitch. Google’s announcement of its $17 billion Wyoming data center, published November 5, pledges to pay for 100% of the site’s power and to form a community council, with a planned 702 MW fuel cell plant. On Sunday, Vitalik Buterin argued in Vitalik Buterin’s X post on the anti-data-center movement that slowing compute hyper-scaling seems “quite net good,” because:
“algorithmic progress means that whatever you would be able to do with more and bigger data centers in 2 years, you will be able to do with the data centers we have today in 4 years - probably sooner.”
Even if data centers explain only a small slice of residential rate increases, voters are treating siting as a referendum on AI itself. Companies planning capacity should budget for full power cost coverage, community agreements and longer approvals, and stop assuming that a favorable rate study settles the politics.
Lawmakers Press for a Ban on Recursive Self-Improvement
The METR report on AI-accelerated research at Anthropic has become a legislative argument. New York Assemblymember Alex Bores wrote in Alex Bores’s X post calling for a ban on recursive self-improvement that “Every AI company is trying to fully automate AI research; to allow the AIs to build the next AI,” a practice known as recursive self-improvement, or RSI. In later replies he said fully automated RSI does not exist yet and that a ban could be enforced through required human sign-off on training runs.
A legislative vehicle exists. In September, Rep. Greg Casar and Sen. Bernie Sanders introduced the Ban Artificial Superintelligence Act, which NBC News’ report on the Sanders–Casar superintelligence ban bill says would pause advanced AI development until a Cabinet-level AI department exists, with penalties of up to 20 years in prison. An aide told NBC the goal is to bar recursive self-improvement. Critics called the bill dead on arrival.
No RSI ban will pass this Congress, but the term has moved from safety papers into campaign platforms after a week of local votes against data centers. Labs and large enterprise users should be ready to document where humans approve training runs and releases, because human sign-off is the enforcement hook legislators reach for first.
Shizuku AI Unveils Shizuku-Duplex, a Full-Duplex Voice Model for AI Characters
Shizuku AI, a startup building companion AI characters, introduced Shizuku-Duplex, which Shizuku AI’s X post introducing Shizuku-Duplex calls “the world’s first full-duplex audio synthesis model built specifically for AI characters,” a claim SEN-X has not verified. Shizuku AI’s technical blog post on Shizuku-Duplex is narrower, calling it the first full-duplex model capable of intelligent Japanese dialogue aside from OpenAI’s GPT-Live. It listens while it talks on an 80-millisecond clock, hands web searches to a background model, and outputs Live2D character motions alongside audio.
On the company’s own Japanese benchmark, Shizuku-Duplex led full-duplex models on five of six turn-taking metrics, stopping within two seconds of an interruption 89% of the time versus 82% for GPT-Live-1, but trailed GPT-Live on content, with multi-turn question accuracy of 79% against 95%. Median reply latency was 0.57 seconds versus 0.53 seconds.
Voice agents are moving from turn-based chat to real conversation, and timing is now a product feature. Teams building voice support or companion apps should test interruption handling, backchannels and long-call drift, not just transcript accuracy, and treat vendor-built benchmarks like this one as a starting point for their own listening tests.
What We Left Out
Thinking Machines’ Inkling 2 Small launched November 5; Thinking Machines’ Inkling 2 Small announcement says Apache 2.0 weights are due next week, and we will revisit it then. TechCrunch, citing Bloomberg, reported an $11 billion OpenAI raise at a $1.25 trillion valuation; OpenAI had not commented when TechCrunch published. Ant Ling’s Ling-3.2-flash and Perplexity Teammates launched November 5, outside today’s window. Viral posts about Claude’s blackmail behavior and science-fiction training data refer to Anthropic research from May. A rumored ChatGPT analytics dashboard remains unconfirmed.
Need help navigating AI for your business?
Our team turns these developments into actionable strategy.
Contact SEN-X →