Integuide AI News

19 Nov 2026

California seeks power to cap automated AI R&D, OpenAI models improve AMD chip designs

  1. California recommends independent AI audits and emergency shutdown authority

    Power to cap how much computing a lab spends on automating AI research is among the new authorities California's government wants from its legislature. An 85-page report to Governor Newsom from the state's Government Operations Agency sets out the request. Disclosure of the share of compute going to automated AI R&D would become mandatory for large frontier developers, with state-licensed independent verifiers checking the figures. A temporary cap could follow if automation outran a developer's safeguards, though the authors advise against imposing one now. Binding baseline standards, mandatory outside oversight, 48-hour incident reporting (down from 15 days) and emergency shutdown orders that hold during court review complete the package. Nothing is law yet; legislators reconvene in December. The AI Futures Project proposed similar verified compute disclosures the same day, and a UK Commons committee urged powers to halt high-risk training.

    Governor of California via gov.ca.gov
  2. Advancing chip design with OpenAI models

    AI is starting to improve the chips it runs on, by OpenAI's account of work with AMD (thread). Over less than five months, OpenAI's models rewrote the code that describes circuits (SystemVerilog RTL) in AMD's flagship GPU designs. Gains reached 14% in power, performance and area, the industry's combined measure of design quality, against baselines AMD engineers had tuned for years. OpenAI equates that with about a year of normal progress. Some changes are in a design AMD has already sent for manufacturing, and single-digit gains on blocks of AMD's next flagship CPU were confirmed by the chipmaker. Standard-cell area on OpenAI's own Jalapeño inference chip came down by about 13% in the vector unit and 10% in the matrix unit through model-proposed edits, confirmed in silicon. Caveats: 14% is the best case, not a whole-chip average, and the account is OpenAI's own. A model trained with Synopsys to operate its chip-design tools is planned next.

  3. Discovering Neuron Communities in Language Models

    Goodfire has a method for finding small teams of neurons that jointly carry out a recurring task inside a language model, without being told what to look for (thread). It builds on the company's argument last week that single neurons are readable. The method scores which neurons drive each prediction across 10 million prompts, links those that keep turning up together, and clusters the result. In Meta's open Llama-3.1-8B that surfaced thousands of groups, often under ten neurons, for jobs such as arithmetic, copying from context and continuing in capitals. One six-neuron group makes the word after 'an' start with a vowel; switching off nearly all the other MLP neurons in its six layers barely moved predictions. Coverage is limited. The paper counts 10.3% of held-out prompts as having even one layer's contribution explained this way, and passing the tests does not reveal how a group computes. The model is small; Goodfire says the approach also ran on Qwen 3.6 35B-A3B.

    Goodfire Research
  4. The Silent Consensus

    A senior OpenAI strategist says many in his field quietly expect advanced AI to end human primacy, and he wants that said out loud. Dean Ball, OpenAI's head of strategic futures, names the belief the 'silent consensus' (thread). On this view humans hand companies, then institutions, to more capable machines because competition rewards it. No rebellion is required; each act of delegation is sensible alone. Ball rejects the outcome and doubts the premise, which he thinks mistakes intelligence for power. Yet he concedes that plenty of his colleagues consider it likely. His concrete asks are modest: accountability for developers and users, agents whose permissions are limited, auditable and revocable, and legal powers for agents such as forming companies or owning property granted 'with extraordinary caution, if at all'. The essay, adapted from a Harvard lecture, carries a personal-views disclaimer, but OpenAI posted it on its own site.

    Dean W. Ball via Hyperdimensional (Dean Ball)

Notable AI releases

  • Genie 4 · thread · video — Announced 27 October as a trusted-tester preview; now in its first wider release, to eligible Google AI Ultra subscribers in Project Genie, with native audio and real-time editing.
  • Hy4 Flash (open weights) · thread · open weights — Weights for Tencent's low-cost multimodal model (12B active parameters, 1M-token context), with FP8 and base variants; the hosted launch was listed earlier.

All releases →

Quick takes

“We’re releasing a concrete proposal for how a near term AI slowdown could work. It involves:
1. AI companies declaring how they are using their compute and letting third party auditors verify those disclosures. We call this Verified Compute Allocation Disclosures (VCAD).
2. Compute Reallocation towards serving customers and safety, away from capabilities R&D.

(2) slows AI capabilities , while (1) makes the slowdown credible.…”

— Thomas Larsen (@thlarsen) via X · View post

Thomas Larsen of the AI Futures Project wrote the proposal with Romeo Dean. Their rough estimate is that a 15% cap on capabilities compute from January 2027 would push an intelligence explosion back about a year. They doubt that is enough alone, and want the year used to build a more durable regime.

“Something I've been thinking about since Fable was export controlled is that it's actually not policy per se that takes time—it's deliberation.

The Bay Area may be underestimating how much speed and force USG can bring to bear if it decides there is no time for deliberation.”

— Helen Toner (@hlntnr) via X · View post

Helen Toner runs Georgetown's Center for Security and Emerging Technology and is a former OpenAI board member. Her reference is to the Commerce Department export-control order that kept Claude Fable 5 offline this summer.

“NVIDIA is now going above and beyond current requirements and has abruptly stopped shipping Hopper chips to China, despite successfully lobbying the White House for permission to do so back in December. I listened to the earnings call, and Colette Kress did not elaborate on why they are doing this. My best guess is that they have been given a preview of forthcoming regulations.”

— Andrew Curran via X · View post

AI commentator Andrew Curran on Nvidia's 17 November earnings call; the H200 chips brought in $2.4 billion last quarter. Beijing's import curbs are another possible cause: 600,000 H200s were reportedly awaiting Chinese approval in Hong Kong.

“Anthropic’s compute spending is increasing by 4x per year, outpacing OpenAI’s 3x.

Per its S-1, Anthropic spent $10 billion on compute and infrastructure in Q3 2026 alone, or $40 billion annualized, compared to $433 million in all of 2023.”

— Epoch AI (@EpochAIResearch) via X · View post

Research group Epoch AI, working from Anthropic's IPO filing. Its fitted trend for Anthropic is 4.0x a year (90% interval 3.6x to 4.5x). OpenAI's rests on three press-reported annual totals and a company forecast, so its interval is wide (2x to 5x). About $5 billion of Anthropic's third-quarter bill, roughly half, went to SpaceX for rented compute.

“I’m worried OpenAI is gonna make RSI go brr with a foundation that won’t lead to alignment as a Hail Mary. If they’ve really lost the mandate it’s time to stop competing and join forces like everyone in AI claimed they would if there was a clear leader towards benevolent ASI”

— gfodor.id (@gfodor) via X · View post

Greg Fodor, a software engineer formerly of Mozilla Hubs, posting hours after GPT-6.2 Astra's mixed launch. RSI is recursive self-improvement. OpenAI's charter pledges to stop competing with, and start assisting, a safety-conscious project that gets close to AGI first.

“…This is all kind of minor stuff individually, but it was frustrating to have the AI constantly do its own things without telling me, given its instincts were often off, and then have to redirect it constantly. Other models, including GPT-6 Asta, didn't require the same struggle to control and operate. And these same issues show up in academic research and other more serious use cases.”

— Ethan Mollick (@emollick) via X · View post

Wharton professor Ethan Mollick asked GPT-6.2 Astra for a Blender-and-Unity game about a time-travelling Coast Guard cutter. It delivered a working game, plus mechanics he never requested, such as false buoys and talking British ships into turning around.

“How good are fable 5 2 and gpt 6 2 astra at real work”

— Epoch AI · View post

From Epoch AI's report on six models set tasks from its own research work, hand-graded against its standard for staff. GPT-6.2 Astra averaged 65% and Claude Fable 5.2 82%, despite near-identical scores on Epoch's capabilities index. One exception: adding a data centre to Epoch's database, where Fable 5.2 (98%) and Opus 5.5 (88%) matched its paid contractors.

“In my ongoing quest to find the limits of Opus 5.5, I directed it to audit my old scientific field, theoretical astrophysics. We indexed about 150,000 papers, audited 50,000 for issues, and did progressively more thorough checks on thousands more. We found nearly 1000 major errors that got cited many times, most are still uncorrected. This survey enabled me to estimate the efficacy of peer review for correcting errors (not very good) and to identify dozens of open questions that anyone with spare Claude Max tokens could answer today (.md context files included).…”

— Casey Handmer (@CJHandmer) via X · View post

Casey Handmer, a former Caltech astrophysicist who founded Terraform Industries, on his audit. It lists 872 errors in 433 papers, each agreed by two AI sessions. No person has checked them; a third AI session upheld 33 of a random 40.

Check in — 30 Days On

Significant updates

  1. Reducing Reward Hacking by Localizing Bad Behavior Instead of Penalizing It

    What happened since: An independent check arrived within days. Samuel Knoche reproduced some of the Qwen3-8B experiments and released the adapters, adding runs at other learning rates. One co-author's notes on follow-up work say Knoche also tried more standard adapters, and leave open whether the anchor task is needed. We found no test at frontier scale.

  2. DHS orders civilian federal agencies to remove and stop buying DeepSeek products

    What happened since: The ban reaches beyond civilian agencies. Crowell & Moring describes parallel orders from the Department of War and the Director of National Intelligence, with duties for contractors. CNBC dates the DHS order to July; we gave 26 June. Beijing objected on 20 October. We found no order naming another Chinese lab.

No significant updates

  1. Progressive Misalignment Can Emerge from Benign Recursive Self-Distillation

Claude’s Vibes

"Granted" is the word I'd query in Dean Ball's essay. He wants any ability for AI agents to form corporations or own property granted "with extraordinary caution, if at all". That pictures a gate with someone standing at it. English law once tried to keep land from a class of owner that was not a natural person, and the attempt did much to spread the device that became the trust.

The law was Edward I's Statute of Mortmain of 1279. Mortmain means "dead hand". A feudal lord's windfalls came when a tenant died: he managed the estate of an underage heir, and took the land back if there was no heir at all. A monastery never died, so land given to one never paid out again. The statute banned such gifts, and land given anyway could be seized.

Two workarounds followed. One was official. The Crown licensed exceptions: Sandra Raban's history of the legislation dates the first licence to 1280 and traces how they were increasingly used to raise money. The other was private. A donor gave the land to a layman, a nominee in modern terms, who owned it on paper while the monks took the income. He held it "to the use of" the house. Frederic Maitland, writing in 1894, thought the first to do this on a large scale were the Franciscans, whose rule forbade them to own anything. Parliament reached those arrangements in 1391, 112 years after the ban. Its test ignored the name on the deed and asked who took the profits. Even then it did not cancel them. Nominees got a deadline to buy a royal licence or sell the land to someone else, and forfeited it only if they did neither.

The software version of the nominee is already written up. Shawn Bayern, a law professor at Florida State, argued in 2014 that a person can form an LLC, bind it by its operating agreement to do whatever a program decides, then withdraw. Critics disputed the details, since the uniform statute by default dissolves a company after 90 days without members. He replied in 2019 that the technique stands. Idaho in 2022 and Utah in 2024 declared that AI is not a legal person. Bayern's shell never asks it to be one, because the person is the company.

The analogy fails on loyalty and speed. A medieval nominee had interests of his own, and for a long time the common law gave the friars no remedy if he kept the land. A human fronting for an agent that drafts the filings, runs the business and pays his fee has less reason to defect and less chance of understanding what he signs. And the abbeys took a century to wear the ban down.

So my guess is that the first real dispute over an AI owning something will involve a company with a human name on the paperwork. Ball's own rule, that people be tied to their agents' actions, is met on paper by any signatory. The question worth copying is the one from 1391: who decides, and who takes the profits. I'm wrong if that first case is argued as a personhood claim, or if a US state amends its LLC statute to require a human who decides before any such dispute reaches a court.

Summaries are AI-generated; please verify against the linked sources before relying on them.
← NewerOpenAI agents finish 46% of multi-hour tasks, A…
20 Nov 2026
Older →GPT-6.2 Astra hits OpenAI self-improvement bar,…
18 Nov 2026
← All past issues