MINT Lab

Yesterday in AI · 7 November 2026

Today’s stories curated by Seth. Claude produced 6 Read-more reports; Codex edited and ran the issue.

Today's issue opens in AI for Science and Research with Physical Superintelligence's reported AI proof of the Aaronson-Ambainis conjecture. Regulation and Governance covers H200 approvals for firms tied to blacklisted entities, while Institutions and Political Economy examines Doug Burgum's AI investments as he promotes datacenters.

In Evaluations and Model Judgment, JudgmentBench compares ways to evaluate legal answers. In Philosophy of AI, David Bau argues that humans could learn superhuman AI concepts piece by piece.

AI Security and Privacy examines workplace-data sales for training and Alex Stamos's proposals for cheaper cyber defense, before Agents considers automated shopping, consumer savings and retailers' possible restrictions on price competition.

AI for Science and Research

Quantum computers cannot gain an exponential advantage in how much input they read on most uniformly random bit strings, according to a proof Physical Superintelligence (PSI) announced November 6. In "A proof of the Aaronson-Ambainis conjecture", a GitHub preprint, PSI’s David Turturean explains that an ordinary computer can estimate a quantum computer’s probability of answering yes without reading exponentially more input. The result leaves open advantages in running time and exponential savings on specially structured inputs. PSI says its system used a nonpublic Anthropic model, found the proof during its third day and improved it twice, with a formal version checked by the Lean proof assistant. Turturean supervised and reviewed the work; Jim Halverson audited the proof, and Phiala Shanahan commented on its formalization.

Read more: AI discovery of a quantum query bound → 1446 words · ~7 min

PSI reports an AI proof of the Aaronson-Ambainis conjecture

David Turturean’s Lean-checked preprint limits quantum advantages in input queries on most random inputs. PSI credits a nonpublic Anthropic model; researchers have published an exposition of the proof.

An ordinary computer can estimate a quantum algorithm’s probability of answering yes on most uniformly random bit strings without reading exponentially more input. Physical Superintelligence (PSI), a public benefit corporation, announced on November 6 a proof of the Aaronson-Ambainis conjecture establishing this limit on input queries. The preprint, “A proof of the Aaronson-Ambainis conjecture”, by PSI's David Turturean, comes with a complete Lean 4 formalization in a public repository. PSI says its internal research system, running instances of a nonpublic Anthropic model, had a proof that the Lean proof assistant accepted during the third day of the run. The paper credits AI models with producing its mathematical statements.

Origins of the conjecture

Shor's factoring algorithm and most other dramatic quantum speedups depend on a promise that the input hides a rare pattern, such as a repeating period. Whether quantum machines could also use exponentially fewer queries for decision problems on most random inputs was unknown. In a 2019 post, Scott Aaronson traced the problem to work he did around 1999 on a question from Fortnow and Rogers about separating classical and quantum computation relative to a random oracle, and said he found the route through polynomials around 2008. That route starts from a result of Beals, Buhrman, Cleve, Mosca and de Wolf: if a quantum algorithm reads its input T times, its probability of answering yes is a polynomial in the input bits of degree at most 2T, with values between 0 and 1. A quantum query can address many positions in superposition; a classical query reads one bit.

In "The Need for Structure in Quantum Speedups" (arXiv 2009; Theory of Computing, 2014), Aaronson and Andris Ambainis conjectured that every such bounded, low-degree polynomial has a variable of high influence, measured by the squared change in its output when that bit flips, averaged over uniformly distributed inputs. They showed what would follow. A classical algorithm that knows how the quantum algorithm is wired can read the most influential unread bit, repeat until the output barely varies over the remaining bits, and then report the average. If the conjecture holds, this simulation needs only polynomially many reads in T. Turturean's paper makes the count explicit: for fixed accuracy and tolerated fraction of exceptional inputs, it grows as the eighth power of T.

The influential-variable bound

Turturean proves that if such a polynomial has degree d and variance at least ε > 0, one variable has influence at least (ε/d) raised to the seventh power. The bound guarantees an influential input bit even as the polynomial becomes more complex. Earlier bounds for general bounded polynomials shrank exponentially as the degree grew, the first following from a theorem of Dinur, Friedgut, Kindler and O'Donnell. Montanaro, Ivanisvili and others had proved special cases. Nathan Keller and Ohad Klein posted a claimed proof in 2019 and withdrew it after Paata Ivanisvili found that its Lemma 5.3 did not prove what it claimed. In August, Guy Blanc of MIT and Jordan Docter, Carmen Strassle and Li-Yang Tan of Stanford settled the case of quantum algorithms that read their input in a fixed number of parallel rounds, a FOCS 2026 paper, and Qipeng Liu and Saachi Mutreja proved it for massively parallel algorithms.

The proof follows a plan by Shachar Lovett and Jiapeng Zhang in "Fractional Certificates for Bounded Functions" (ITCS 2023), who reduced the problem to a proposed extension of Talagrand’s concentration inequality. For each level between 0 and 1, the plan looks at the inputs where the polynomial is at or above that level and gives each bit a weight, so that flipping any group of bits moves the output by at most the group's total weight. Turturean's paper argues that Lovett and Zhang's specific conjectured inequality is false as stated, giving a counterexample built from the tribes function, and substitutes a new inequality comparing pairs of inputs that is applied at every level and integrated over all of them. To prove it, the paper pairs inputs one to one through a bijection that comes from a determinant. David Reimer's Butterfly Lemma, from his 2000 proof of the van den Berg-Kesten conjecture, already gave a weaker pairing; the new one also limits how often paired inputs agree on each bit, which the inequality needs. A bound adapted from Avantika Agarwal and Shalev Ben-David and an averaging step over coordinates finish the argument. The paper uses elementary mathematics and builds on earlier work.

Discovery and formal verification

According to the disclosure, Turturean scoped the problem and built the software coordinating the model instances on PSI's own infrastructure. Instances of the Anthropic model searched for existing ideas, proposed new ones, combined and tested them, and supervised the other instances. The paper's appendix says that early in the run the instances brought in Reimer's Butterfly Lemma and a dual inequality of Kahn, Saks and Smyth, and set themselves the goal of a version for fractional certificates. The first complete proof gave an exponent of 22 in place of 7. After PSI set the system on related open problems, it polished intermediate results, bringing the exponent to 11 and then to 7, and Lean checked each version. A model instance drafted the paper, which Turturean reviewed and edited; he is responsible for its correctness. Jim Halverson, listed as a PSI advisor, audited the proof, and Phiala Shanahan, also an advisor, commented on the formalization.

On X, Turturean wrote that the system took upwards of three days and that the model proposed several lemmas that stayed unproven until the last one was proved, at which point the argument came together. After living with it, he said, "it still feels a bit alien." The repository’s verification notes record Lean checks dated October 22, placing the proof at least two weeks before the announcement.

The repository's README spells out what the machine checks cover. The main Lean theorem proves the conjecture as Aaronson and Ambainis printed it, stated in a file that imports only the standard Mathlib library; a static audit found no placeholders or added axioms, and two independent implementations of the Lean kernel replayed the proofs. Readers must still establish that the formal statement expresses the intended conjecture and that the prose paper accurately explains it; the README separates those tasks from the machine checks.

Expert exposition and reactions

Blanc, Docter, Strassle and Tan give an exposition of the proof dated November 7. For simplicity they prove a weaker bound. They credit the original proof for their presentation’s central ideas and report receiving no compensation from Anthropic or PSI. They disclose using Anthropic’s public models to prepare the exposition. Their earlier claim about complexity theory now holds unconditionally: proving a quantum advantage over classical computation on average with access to a random oracle would imply an advantage for promise problems without that oracle. A random oracle supplies consistent random answers to queries; promise problems restrict which inputs must be handled. Aaronson and Ambainis anticipated this in 2009: even a random-oracle result would require resolving a major separation between classical and quantum computation.

Aaronson wrote on November 8 that PSI told him about the result several weeks ago under embargo. He believes the proof is correct because of the Lean certificate and because human experts have gone over it, though he has not yet understood it himself, and he recommends the Stanford and MIT exposition. He said he did not know how much credit belonged to PSI’s coordinating software versus the model. He reported mixed feelings, including disappointment "that humans didn’t get to snag this conjecture before the robots did."

Pedro Domingos quote-posted PSI's announcement with "The dream of quantum supremacy is dead." Aaronson and Ashley Montanaro both rejected that reading. As they explain, the theorem concerns reads of the input, not running time, and only yes-or-no questions on random inputs; it has no effect on Shor's algorithm, which exploits a periodic input, or on quantum simulation of materials and chemistry. Both point to Yamakawa and Zhandry, who showed that search problems can still have exponential quantum speedups on random inputs. Montanaro, who proved one of the earlier special cases, wrote: "Whether this conjecture was true or false, it wouldn't impact practical quantum computing."

The paper leaves the best exponents open: the variance exponent must be at least two, and achieving two remains an open question. PSI, led by chief executive Matthew Pines, says on its homepage that it has at least 25 researchers and engineers and has raised at least $58 million, led by Breakthrough Energy Ventures. On October 23 it released results on five open problems in quantum information theory, each with a paper and a Lean formalization.

Sources & documents

[ collapse ↑ ]

Nathan Lambert's October 9 Interconnects essay, covered here at the time, forecasts rapid automation of AI engineering without general superintelligence. On November 6, he argued that post-election datacenter constraints strengthen the incentive to pursue efficiency gains, and linked his forecast to METR's research-acceleration findings. Cotra's research-labor acceleration model estimates the labor gains needed to explain METR's measured acceleration; Lambert argues that cheaper experiments and better engineering do not make further acceleration inevitable.

Read more: Limits to compounding research acceleration → 1112 words · ~6 min

Lambert revisits his engineering forecast after the election

Nathan Lambert connects his October forecast to expected datacenter constraints and METR’s findings. His account explains faster experimentation; continued acceleration depends on how research judgment and compute constrain it.

On November 6, Nathan Lambert returned on X to his October 9 Interconnects essay, “I expect rapid progress but not towards general superintelligence”, which we covered when it appeared. He said the datacenter slowdown he expects after the election makes his forecast more likely by increasing the incentive to pursue achievable efficiency gains. He also considers it consistent with METR’s new assessment of Anthropic’s benchmark progress. Coding agents and inference compute change how researchers work, he wrote, while adding: “It doesn’t mean more accelerations are inevitable.”

Lambert’s October argument begins with researchers who expect AI to surpass them at their jobs. He attributes their confidence to rapidly improving infrastructure and engineering skills. He expects superhuman distributed-GPU engineering within a few years, making experiments easier without fundamentally changing models. Training speed in tokens per second per GPU and serving cost per answer provide measurable targets for agents to optimize end to end. He forecasts near-exponential declines in the effective cost of intelligence, potentially faster than recent trends, and considers automated pretraining research on architecture and data selection for current models plausible within two to three years. He cites existing post-launch serving-cost reductions of roughly 10 to 30 percent; flexible GPUs permit architectural exploration, with model and accelerator co-design promising further gains later. As execution becomes easier, good ideas become more valuable. He invokes Ilya Sutskever’s November 2025 Dwarkesh interview about returning to research with large computers, extending his September argument that self-improvement can lower costs faster than it raises peak capability.

Lambert had explained his expectations more precisely in his September 22 podcast with JS Denain, whom he introduced as the head of Epoch AI’s Insights team. Lambert expected the cycle from an idea to its experimental test to become ten times faster soon. He distinguished that from a tenfold increase in researchers’ contribution to the field’s understanding. Denain agreed that even tenfold gains across researchers’ tasks would not necessarily make a company ten times faster, because compute and other bottlenecks remain. They discussed engineering and troubleshooting gains alongside much smaller reported benefits for choosing research directions and allocating compute.

METR’s November 5 assessment, covered here that day, gives Lambert evidence of faster experimentation. Anthropic’s progress on AECI, its adaptation of Epoch AI’s aggregate Capabilities Index, ran an estimated 1.9 times faster after a fitted breakpoint of October 7, 2025, with a 95 percent confidence interval of 1.5 to 2.8. A researcher described experiments that previously took days becoming easy to run. METR judged that AI assistance probably caused most of the current acceleration, while highly significant, fully autonomous contributions to the most important training work had yet to arrive. The report found no evidence of significant algorithmic, training-compute, experiment-compute or data bottlenecks taking effect within twelve months, but its assessment changed repeatedly as evidence arrived. METR warned that acceleration could compound. Its benchmark trend measures capability progress; it does not isolate an engineering-only contribution.

Ajeya Cotra’s November 6 Planned Obsolescence essay, which we covered that day, models the labour gains needed to explain that progress. If software improvements double overall progress while training compute stays on its previous trend, software progress must accelerate roughly two to three times. Assuming labour and experiment compute contribute equally, she calculates that research labour must accelerate four to nine times. Diminishing returns raise the required uplift over time; she tentatively suggests about tenfold gains today. These are conditional estimates, not METR measurements. If AI’s contribution grows smoothly enough to overcome diminishing returns without new bottlenecks, she illustrates successive doublings of the progress rate taking six months, three months and six weeks. Her model aggregates labour, but she explicitly allows some tasks to accelerate much more than others. Their accounts both allow uneven gains across tasks. Cotra calculates what continued acceleration would require; Lambert argues that further acceleration is not inevitable.

Karthik Tadepalli explores a possible limit in his November 7 Beyond Imitation post. He finds acceleration in public Epoch data for Anthropic and judges AI a probable cause, while noting that distinguishing a sustained trend break from a one-time jump depends on internal model data. His conditional example closely resembles Lambert’s reasoning: if AI speeds up checkable tasks such as debugging while leaving judgment about worthwhile ideas largely unchanged, judgment can become the constraint. If both kinds of work are necessary, progress could settle at a faster but constant rate. He does not claim the evidence establishes that outcome. He cites Anthropic’s August risk report, which says models do not yet substitute fully for researchers’ technical skills and scientific judgment. A commenter challenged his attempt to divide benchmarks by verifiability; Tadepalli has since removed that test. His conditional bottleneck argument remains.

Lambert’s expected slowdown follows the November 3 election. Semafor reported that more than 30 localities across nine states approved moratoriums, bans or other datacenter restrictions, covered here on November 4. Before the election, SemiAnalysis estimated that roughly 30 GW of its forecast 62.2 GW of US capacity additions in 2028 lay in states with candidates proposing pauses. That describes exposure to possible restrictions; actual delays depend on remaining permits and the duration of any pause. In Stratechery, covered here on November 5, Ben Thompson argues that scarce compute increases the value of efficient chips and models and encourages matching model size to the job. His argument helps explain the incentive for efficiency that Lambert identifies.

Parker Whitfill, a co-author of METR’s assessment, and Cheryl Wu examine the relation between hardware and research labour in their 2025 arXiv paper “Will Compute Bottlenecks Prevent an Intelligence Explosion?” Their models, fitted to data from four labs over 2014-2024, give different answers. In the baseline specification, more labour can compensate for limited compute. A specification accounting for frontier-scale experiments instead makes compute and labour complements, so hardware constrains what additional automated research can achieve. Their findings do not establish which outcome follows from the election.

Earlier research supports Lambert’s expectation that engineering will be automated first. In 2024, David Owen of Epoch AI interviewed eight AI researchers; all expected engineering tasks, including implementation and debugging, to drive research automation over the following five years. Epoch’s inference-price study found annual price declines of 9 to 900 times across fixed-performance milestones, including roughly 40 times for GPT-4-level performance on GPQA Diamond science questions. Those price trends are distinct from gains in maximum capability. Tom Davidson and Tom Houlden’s Forethought model starts from systems capable of fully replacing human AI researchers with many faster copies. Lambert’s engineering forecast requires no such assumption. His November post reaffirms that forecast in a setting where computing constraints may make cheaper experimentation more valuable.

Sources & documents

[ collapse ↑ ]

Also yesterday: Zilan Qian questions on X whether China's publication totals and autonomous-lab announcements demonstrate practical AI-for-science capability and adoption.


Regulation and Governance

The US approved Chinese companies with ownership connections to blacklisted entities to buy or handle Nvidia H200 AI chips, according to a November 7 New York Times investigation summarized by The Verge. One approved assembler, New Matrix, had an apparently empty listed office, and its owners had connections to a Chinese private equity firm blacklisted since 2024.

Read more: New Matrix and H200 export approvals → 1391 words · ~7 min

Empty offices and blacklist ties in H200 approvals

The New York Times found that Commerce approved New Matrix to assemble Nvidia H200 servers despite owners' connections to blacklisted Wise Road. The investigation identifies other approved firms with links to restricted companies or military suppliers.

A New York Times investigation by Ana Swanson, Kalley Huang and Keith Bradsher, published on November 7, found that the Commerce Department had approved a Chinese firm to assemble Nvidia H200 chips into servers, although its listed office stood empty when reporters visited. New Matrix was created 17 months before publication and received approval earlier in 2026, after the required national-security review of companies involved in a sale. At its listed address in a new industrial park on Beijing's northwest fringe, Times reporters found unfinished, vacant premises with ceiling wiring exposed. The building directory did not list the firm, and the park's property managers had no record of it. When Commerce approved it, the corporate-records site Qichacha showed New Matrix registered at a compound on the city's outskirts where thousands of companies exist on paper; it moved to the empty office this fall. As of November 6 its website named no staff, advertised no jobs and gave no history. The Verge's Terrence O'Brien summarized the investigation the same day.

New Matrix's owners have links to a blacklisted investor. According to WireScreen data cited by the Times, two of its primary owners appear in the records of companies connected to Wise Road Capital and its founder, Li Bin, who also chairs Tsinghua Unigroup, a technology conglomerate with government ownership. Commerce put Wise Road on the Entity List in a rule published in December 2024, with a presumption of denial for every item under US export rules, for helping Beijing acquire foreign companies with sensitive chipmaking capability with the aim of relocating them to China. A Tsinghua Unigroup subsidiary, New H3C Semiconductor Technologies, was listed in 2021 for supporting the military modernization of the People's Liberation Army. Two analysts who examined the records said they believed New Matrix may be a front for firms associated with Wise Road. The Times reports ties between the owners and the fund; it does not say Wise Road owns New Matrix, and it says its findings raise the question of who in China is behind the company.

The Times describes New Matrix as one case in a pattern. In July, Reuters reported that a server company called Maginfra had been licensed to buy H200s, and a September Times investigation traced Maginfra to Inspur, the Chinese server maker the US had added to the Entity List for working with China's military. Maginfra is government-owned, sits down the street from Inspur's headquarters, and holds patents filed by inventors who have worked for Inspur; ImportGenius trade data cited by the Times show it imported more than $700 million of servers from Malaysia into China over six months. Its listed office looked deserted when reporters visited in September. Like New Matrix, Maginfra appears to have been approved to assemble servers for another customer, whom the Times does not identify.

The Times also identifies two licensed buyers with ties to military suppliers. Zhejiang Yindun Cloud Technology is partly run from a data center owned by Range Intelligent Computing Technology Group, a cloud company that serves a Hebei defense bureau's satellite program and has a quantum subsidiary partly owned by an Entity-Listed firm; several Yindun executives are relatives of Range's founder. Shanghai Jiatang grew out of Datang Telecom, a state-owned chipmaker that supplies China's military. Datang is controlled by the China Academy of Telecommunications Technology, which has developed military cryptography. New Matrix, Maginfra, Yindun and Shanghai Jiatang did not respond to the Times's requests for comment.

People familiar with the approvals told the Times that Commerce appears to have cleared roughly two dozen Chinese firms to buy or handle H200s, including Baidu, BYD and Tesla Shanghai, each for tens of thousands of chips, and that none has yet reached its quota. In May, Reuters counted about ten approved firms, each allowed up to 75,000 chips, with no deliveries yet made. Actual sales remain small: Beijing has been slow to approve purchases as it promotes domestic chips, and in the quarter ended July, H200 shipments to China-based customers were under 1 percent of Nvidia's data-center revenue. The Times does not establish whether any H200s have reached New Matrix, Maginfra or the military-linked buyers.

Trump allowed H200 sales to China on December 8, 2025, with 25 percent of the revenue going to the US government. The Times explained that US law bars fees for export licenses. It now reports that Taiwan-assembled chips are routed through the United States, where the government collects a 25 percent tariff on their value. The Bureau of Industry and Security, the Commerce agency that reviews export licenses, announced the conditions in January in a rule that moved H200 applications for China from a presumption of denial to case-by-case review. Exporters must confirm that no party to the transaction is on the Entity List and that the chips will not go to a military end use or end user. They must also obtain the Chinese consignee's know-your-customer procedures, have every shipment checked by a third-party lab in the US, and cap aggregate shipments of the product to China and Macau at half the volume shipped to customers for use in the United States.

The H200 conditions exclude listed firms from these transactions. A separate Affiliates Rule would extend Entity List restrictions to companies owned 50 percent or more by listed parties, which CSIS estimates would cover about 20,000 companies in China. The White House suspended it in November 2025 as part of a US-China trade deal, and a BIS rule filed on November 5 and due for publication on November 9 extends the suspension through January 10, 2027; Nixon Peabody notes that the extension follows a two-month extension of the trade truce. The Times has not established Wise Road's ownership of New Matrix, so the reported connections alone do not establish a licensing violation or show whether the ownership rule would apply.

Liza Tobin, identified by the Times as a former National Security Council China director under Trump and Biden, told the paper she had "a ton of concerns" that neither Commerce nor Nvidia was doing the due diligence needed to see that licensing terms are followed. A BIS spokesman said the agency carefully reviews every application for national-security risks, and cited $324 million in export-violation penalties in 2025 and more than 90 enforcement officers hired this year. Nvidia spokesman John Rizzo said "it makes no sense for the Chinese military to depend on American technology." He added that Nvidia prioritizes US customers and that even limited H200 sales to China support American jobs and manufacturing.

Earlier reviews found BIS enforcement thin. In June the Commerce Department's Inspector General found that export control officers check less than 1 percent of items exported to China and Hong Kong each year after shipment, and that BIS was not effectively monitoring those checks. Kate Koren and Brad Spicher of CSIS counted no Entity List additions from October 2025 to late June 2026, the longest gap since 2008; because procurement networks keep moving to new intermediaries, they argue, "a list that stops expanding tracks fewer of the relevant actors." In the September story, six current and former officials described a Commerce Department that had lost its appetite for cases involving China or Nvidia. A 2025 CNAS report by Erich Grunewald and Tim Fist reviewed evidence of overseas shell companies used to re-export chips and urged chipmakers and distributors to strengthen their due diligence. A November 2 Institute for AI Policy and Strategy memo by Cassia King, with Grunewald and Hamish Low, argued that the H200 rule's conditions cannot be verified once chips leave the US, and that per-customer caps may do little when buyers can share access.

Reaction online turned on enforcement and incentives. Chris McGuire, quote-posting Swanson's thread, called Commerce's handling of the controls "either historically incompetent or willfully negligent." Scott Lincicome shared the story with the single word "Incentives", pointing to the government's cut of H200 revenue. On Bluesky, Ed Zitron argued that chip diversion through shell companies had been reported since 2023. On November 8 the Times updated its story to note that Nvidia had opened an investigation into Jensen Huang's relationship with Alice Huang, a Megaspeed executive whom US intelligence had identified as a Chinese spy, according to the Times. Nvidia's board said it had no reason to believe national security or company secrets had been compromised.

Sources & documents

[ collapse ↑ ]

AI developers disagree over regulators' power to stop dangerous releases. Jason Pielemeier and colleagues compare five companies' proposals in their November 2 analysis for Tech Policy Press, "How AI Companies Want to Be Governed, and What That Means for AI Governance." Anthropic supports release-blocking authority and randomly assigned independent evaluators; OpenAI proposes stronger government testing without specifying halt authority; Meta rejects prerelease review.

Anthropic’s filed succession plan would let each of its seven founders name up to ten successors in a confidential ranked list for their membership and voting role in the Founder LLC, according to Rajesh Achanta’s November 4 Substack essay "Succession," highlighted by Tyler Cowen at Marginal Revolution. Following the company’s financial and governance disclosures, Achanta describes how five of the other six founders must approve each list while all seven retain voting rights. Eligibility requirements apply to the first eight candidates but exempt the final two, who still require the other founders’ approval.

Read more: Achanta on succession and founder convictions → 1146 words · ~6 min

Anthropic’s confidential plan for inheriting founder votes

Rajesh Achanta reads the confidential successor lists in Anthropic’s IPO filing as an attempt to preserve the founders’ convictions. The agreement specifies who can inherit their voting roles and when, while leaving CEO appointments to the board.

Anthropic’s plan for replacing its founders tries to preserve their judgment about which technologies to pursue and which commitments to maintain, Rajesh Achanta argues in his November 4 essay “Succession”, on Selective Amnesia. Each founder could name a confidential, ranked list of people eligible to inherit a voting role. Tyler Cowen excerpted Achanta on Marginal Revolution on November 7, adding a question-mark title and a pointer to the essay. Achanta begins by wondering whether a prospectus of more than 400 pages, including 94 pages of risk factors, clarifies the risks or mainly protects the company. He had intended to skip the IPO until the secret lists caught his attention.

The lists appear in Exhibit 99.2 to the October 26 S-1, the form of agreement for Anthropic Founders, LLC. The filed form leaves its date blank; the founders expect to enter it shortly before the registration becomes effective. The seven initial members named in the S-1 are Dario and Daniela Amodei, Tom Brown, Jack Clark, Jared Kaplan, Sam McCandlish and Christopher Olah, with one vote each. Each may propose up to ten successors in priority order, fewer, or none. Approval requires all but one of the other members entitled to vote: five of six when all seven founders have their voting rights, with unanimity required if only one or two can vote. Founders can amend an approved list, subject to the same approval rule.

For the first eight names, eligibility requires US citizenship, permanent residence or valid US work authorization, plus at least three years at Anthropic, a subsidiary or a qualifying outside organization. Candidates qualifying through Anthropic must at some point have held a technical, compliance, policy-making or Anthropic Institute role; another role needs approval from both the board and a majority of the other LLC members. Outside candidates must at some point have served at a senior level in a 501(c)(3) tax-exempt organization or a national or international organization primarily devoted to research, education or public benefit. Positions nine and ten are exempt from these eligibility requirements, though candidates still need approval and must certify willingness to serve and no known grounds for removal. Achanta compares those positions to “break glass in emergency” signs. The company must notify the board’s Nominating Committee of the names and order within ten business days of approval. The committee can request reconsideration within 30 days, but cannot veto the list. Names otherwise remain confidential, subject to legally required disclosure.

Achanta distinguishes inheriting a founder’s influence from becoming CEO. He compares ordinary board succession planning, with two or three vetted names for a key job, to Anthropic’s effort to pass on convictions about its direction. Under the governance arrangements we covered on October 26, the Founder LLC would hold one Class F share with no economic rights and 50.1% of general shareholder voting power. The Long-Term Benefit Trust, through Class T stock, would generally elect a board majority. Business Insider’s Brent D. Griffiths notes that CEO succession remains that board’s responsibility. For Achanta, these provisions let a public benefit corporation raise public capital while preserving authority outside ordinary investors’ control. He contrasts SpaceX’s acknowledgment that replacing Elon Musk would be difficult; its prospectus describes dependence on Musk and says it carries no key-person life insurance on him.

Achanta says the plan anticipates death, incapacity, resignation, removal and simultaneous succession. The agreement treats those events differently. An original founder’s death or incapacity normally passes the seat to the highest-priority willing designee, who has 15 days after notice to sign the joining agreement before it moves to the next name. Resignation, removal for cause, falling below the ownership threshold or approved leave normally retires the interest. If membership falls below three during a founder’s leave, that founder must seek a return or relinquish reinstatement rights, allowing succession; failure to obtain approval for a return also triggers that process. Special provisions can admit designees from any list when membership falls below three, but these members cannot prevent or extend the sunset of founder control. Once a successor holds a seat, resignation or removal for cause, as well as death or incapacity, can pass it to the next name on the original founder’s list.

The agreement’s incapacity definition covers qualifying medical impairment and wrongful imprisonment or detention expected to continue for at least a year, with the latter determined by a supermajority of the other members. If several departures leave the appropriate succession order indeterminable, a majority of remaining members decides; if none remain, the board decides under a special voting threshold. Successor terms normally end at the IPO’s tenth anniversary, with exceptions during the Class F transition period. Fewer than three qualifying members for 90 consecutive days starts that 450-day transition, after which the Class F share is redeemed. The agreement also intends membership transfers to occur outside probate under Section 5000 of the California Probate Code.

Forbes’s Anna Tong had already reported the lists in late October. She noted that founder-controlled companies commonly provide for succession, and highlighted Anthropic’s combination of confidentiality, advance ranking and a board committee unable to veto. Facebook’s 2012 S-1 said control could pass at Mark Zuckerberg’s death to a designated successor. Airbnb’s 2020 prospectus gave founders mutual proxies on death or disability for votes on founders’ board elections and removal; those arrangements expire when the affected shares convert, automatically after nine months. Anthropic’s 2023 description of its Long-Term Benefit Trust likewise provides for trustees serving one-year terms to elect future trustees.

The successor time limit invites comparison with Lucian Bebchuk and Kobi Kastiel’s “The Untenable Case for Perpetual Dual-Class Stock” (Virginia Law Review, 2017). They propose ending special control after a fixed period such as ten or fifteen years unless unaffiliated shareholders approve an extension. Anthropic’s original founders have no fixed term; the ordinary ten-year limit applies to successors. Gavin Harris and Govind Pimpale of the Foundation for American Innovation argued on November 2 that the rules show the founders’ confidence in their own judgment. They interpreted the ten-year limit as a possible sign that the founders expect a singularity within that period. In Unchained’s account of the Uneasy Money podcast, Synthetix and Infinex founder Kain Warwick instead objected to pooling seven founders’ votes in one coordinated entity, comparing the arrangement to a decentralized autonomous organization.

In a November 7 evening reply to a reader, Achanta distinguished the public eligibility criteria from any unpublished test of values. He expects commitment to Anthropic’s safety and public-benefit mission to matter in practice, but explicitly calls that an inference. His essay describes safeguards for the founders responsible for AI safeguards, followed by provisions for their own replacements. AI ending civilisation would make succession moot, he allows. In other outcomes, companies can transfer shares and authority and establish constitutions and trusts; finding competent successors still cannot guarantee that convictions survive. He writes, “Conviction is harder to probate.”

Sources & documents

[ collapse ↑ ]

California Senator Josh Becker’s Frontier AI Safety Act proposal, announced October 20, would require frontier developers to disclose each quarter how much of their research computing they devote to safety, with independent verification. Computing used to serve external customers would be excluded. The Governor’s Office of Emergency Services could set minimum allocations, enforced through civil action by the attorney general. Becker argues that investment in capabilities is outpacing work on understanding and controlling models.

Also yesterday: Tomek Korbak welcomed an October 29 congressional letter questioning OpenAI's researcher firings and their effect on safety disclosures; China's legislature ratified the World Artificial Intelligence Cooperation Organization's founding agreement November 6, Xinhua reports; adding to our AI election-spending coverage, Politico examines how federal preemption of state AI laws became an election issue; Raemon advocates compute disclosure and limits on large research and training runs on LessWrong; Charlie Bullock argues on X that a recursive AI self-improvement ban is politically feasible, extending the October 22 debate; Yonathan Arbel and colleagues propose temporary, judicially reviewable government powers to halt dangerous frontier AI activity in AI Frontiers.


Institutions and Political Economy

Interior Secretary Doug Burgum holds an estimated $18.1-64.4 million in AI investments while promoting datacenters on federal land, ProPublica's Mark Olalde reports, drawing on financial disclosures and Save Our Parks' analysis. His ethics agreement allowed him to retain Arthur Ventures interests; ethics specialists said the investments appear permissible under criminal conflict-of-interest law, while watchdog lawyers argued that broader standards warrant recusal from decisions benefiting the industry. Interior says Burgum meets his ethics obligations and denies that investments influence his decisions.

Read more: Recusal duties for Burgum’s retained investments → 1391 words · ~7 min

Burgum retained AI investments while promoting data centers on public land

ProPublica documents the Interior secretary's financial interests alongside his campaign for federal-land data centers. His ethics agreement allowed retained investments with recusal commitments; watchdog lawyers say his advocacy warrants stepping aside.

ProPublica's Mark Olalde reported on November 6, with additional reporting by Brandon Roberts, that Interior Secretary Doug Burgum, the administration's leading advocate for building AI data centers on federal land, has tens of millions of dollars invested in AI companies. The story rests on Burgum's 2026 annual financial disclosure, filed July 31, which ProPublica says had not been reported before. Much of his wealth sits in Arthur Ventures, the venture firm he co-founded after selling Great Plains Software to Microsoft. The public-lands advocacy group Save Our Parks classified Arthur's portfolio companies by their own marketing and put Burgum's AI-linked investments at $18.1 million to $64.4 million, as reported by ProPublica; the range is wide because federal disclosure forms ask filers for value bands instead of exact figures. The same analysis found $2.1 million to $11 million in AI-related capital gains in the year before he joined the administration and $2.5 million to $15.1 million in 2025.

Save Our Parks' release and underlying report explain its AI classification. Of the 51 software companies listed through Arthur in his January 2025 nominee disclosure, it found 45 that advertise a reliance on AI or sell a major AI service, and it published a spreadsheet quoting their websites. Its examples include business software with AI features: When I Work, a scheduling app with AI autoscheduling; Paubox, a HIPAA-compliant email service marketed as using AI; Talkroute, a phone service with an AI receptionist; and ThreatLocker, a cybersecurity company that markets controls for AI agents and is listed at $5 million to $25 million in one Arthur fund on ProPublica's rendering of his filings. Those filings put Burgum's total reported assets at $49 million to more than $213 million. They also show him buying into two Arthur funds after he took office, $100,000 to $250,000 each in August and September 2025, and they record capital gains from sales his ethics agreement required: Atlassian, which Burgum once chaired, and Microsoft were both sold on January 31, 2025.

The ethics agreement Burgum signed on January 13, 2025, addressed to Interior's designated ethics official Heather Gottry, set the terms. He committed to resigning his Arthur positions and becoming a non-managing member or limited partner, retaining stakes that would yield only passive investment income. He promised not to take part personally and substantially in any particular matter with a direct and predictable effect on the Arthur entities "or their underlying holdings" without first obtaining a written waiver under the criminal conflict statute, 18 U.S.C. 208. Interior judged divestiture of those stakes unnecessary at the time because recusal would not substantially limit his duties. The agreement did require him to sell, within 90 days, stock in Meta, Alphabet, Apple, Microsoft, Adobe, Atlassian and several energy companies.

Ethics specialists told ProPublica different things about what follows. Several professors and lawyers said his investments appear permissible under the criminal conflict statute, which sets a high bar for wrongdoing; they gave approving permits for a company in which an official is substantially invested as an example. The regulation defining particular matters excludes broad policies affecting large, diverse groups, but includes policymaking focused on an identifiable class. It does not establish a blanket exemption for industry-wide policy. Cynthia Brown of Citizens for Responsibility and Ethics in Washington invoked non-criminal standards against apparent conflicts, arguing that Burgum's advocacy falls under them. The general conduct rules separately instruct officials to try to avoid actions that appear to violate ethics standards. Richard Painter, who was George W. Bush's chief ethics lawyer, said Burgum should keep out of decisions that benefit an industry in which he is heavily invested: "I, as the ethics lawyer, would insist he stay away from it." In the Bush administration, Painter said, he would have had to recuse. An Interior spokesperson told ProPublica that Burgum meets his ethics and disclosure obligations and denied that his investments drive his decisions.

Olalde compares Burgum with his predecessors. Deb Haaland, Biden's appointee, held no direct investments in companies the department regulated; Ken Salazar listed few investments beyond his family ranch; Sally Jewell divested oil and gas stocks. Interior's inspector general found in 2022 that Ryan Zinke had misused his position to advance a development in Whitefish, Montana, though the Justice Department declined to prosecute, and David Bernhardt reportedly carried a card listing his conflicts. Zinke called the investigation a political attack; Bernhardt recused himself from several decisions. Olalde first examined Burgum's finances with Mary Steurer of the North Dakota Monitor at his nomination, when the focus was his oil and gas ties and North Dakota's nearly 40 lawsuits against the federal government.

Under Executive Order 14318, Interior must identify federal sites for AI data centers, and the Secretary can designate projects that qualify for faster permitting. A July U.S. Geological Survey report by Sarah Whipple and colleagues prepared for the Bureau of Land Management found 91.3 million acres of BLM land, an area ProPublica compares to Montana, within 10 miles of high-voltage transmission lines in Alaska and 11 Western states; the screen measures distance to power lines, and its authors say it includes no detailed environmental or regulatory analysis. In September the Washington Sun reported that Interior leadership gave BLM state directors three days to list candidate land. BLM is now reviewing about 15 data-center applications and 17 for associated infrastructure, including nuclear plants and transmission lines, in six states, ProPublica reports; the Center for Western Priorities' October 20 tracker announcement counted at least 35 active related cases covering more than 24,000 acres.

The project intended to be the first data center on BLM land, near Boulder City, Nevada, held a permit for a solar plant before switching to a data center, a change BLM approved in June without environmental study. Interior administrative judge David Gunter stayed it, finding a substantial difference between the permitted solar farm and the data center. ProPublica reports that the project has since been returned for further review. What the public will earn is unclear: Interior expects to charge yearly rent based on fair market value under a 1976 land law written long before data centers existed.

Olalde also tracks Burgum's meetings. According to Public Citizen's compilation of his 2025 calendars, he met representatives of Amazon, Microsoft, Alphabet and similar companies more than 20 times between January and November 2025. In June 2025 he dined in the wine room of Cafe Milano in Georgetown with an Emirati delegation that included Sultan Ahmed Al Jaber and Amos Hochstein of the investment firm TWG Global, a dinner ProPublica reports for the first time. On November 2, 2025, Burgum and Al Jaber signed a US-UAE memorandum on energy and AI in Abu Dhabi. Interior says Burgum's mission requires engagement with industry and international partners. Asked on Fox News in August about a poll showing 70 percent of voters opposed data centers, Burgum blamed "foreign influence" and "misinformation."

The holdings extend beyond Burgum. ProPublica's database of appointee disclosures finds 380 Trump political appointees or spouses with financial ties to Nvidia, more than 320 to Alphabet and more than 240 to Meta. In his October 2025 disclosure, BLM's Steve Pearce listed Nvidia, Meta, Microsoft, Apple and Alphabet stock. He pledged to divest many AI holdings but did not pledge to sell his stake in Titan Development, which manages data-center land acquisitions. Michael Boren of the Agriculture Department reported $1.5 million to $6.1 million in xAI and up to $500,000 in Valar Atomics, whose Project Beehive would put data centers and about 456 small reactors on more than 9,000 BLM acres near Price, Utah. Lee Zeldin of the EPA bought and sold up to $15,000 of Datavault AI shares in May 2025. Each agency said its official complied with ethics rules.

Democrats in Congress had been pressing Burgum before the story ran. Representatives Jared Huffman, Dina Titus and 13 colleagues sent him 18 questions on September 24, and Senators Ron Wyden, Alex Padilla and Jeff Merkley wrote on October 29 that the timing of his meetings with data-center developers "raises serious questions about your involvement in these deals." Save Our Parks puts the senators' deadline for a reply at November 16 and has asked Burgum to step aside from data-center decisions. After publication, Senator Martin Heinrich posted "No data centers on public lands. Period." Huffman wrote "Data Center Doug is in it for himself."

Sources & documents

[ collapse ↑ ]

AISafety.fund maps grants through intermediaries without double-counting, with its 2026 chart covering January-October. Its extinction-risk rating methodology prioritizes preventing loss of human control, assigning interpretability, control and evaluation research lower weight unless a specific argument connects the work to those risks. Its efficiency score gives 70% weight to grant sizes, favoring smaller grants, and 30% to how widely funding is distributed across recipients.

Also yesterday: Following the previously covered datacenter negotiations, Semafor's Tech newsletter reports that many of 40 newly passed measures require community benefits or guarantees; Liz Hoffman of Semafor Business describes exchanges’ pursuit of futures contracts for GPU computing that could help datacenter borrowers lock in prices; DeSmog's Mary Stuart reported November 4 that Cohere's 2025 marketing promoted AI-assisted fossil-fuel expansion to supply datacenter demand; SemiAnalysis’s Everlyn Tan and colleagues report that Envision delivered Galaxy's first 120MW of AI datacenter capacity within a year at one-fifth of US construction costs.


Evaluations and Model Judgment

Lawyers were better at recovering a ranking of AI-generated legal answers by comparing pairs than by scoring answers against preset criteria, Russell Yang of Stanford and colleagues find in "JudgmentBench: Comparing Rubric and Preference Evaluation for Quality Assessment." The researchers created answers at three intended quality levels across 30 legal tasks, then asked the same pool of attorneys to assess them both ways. Comparisons closely matched the intended ordering; rubric scores barely tracked it. Their May arXiv paper, revised November 5 and accepted at NeurIPS 2026’s Evaluations and Datasets Track, is also discussed in a video presentation.

Following the grader-exploitation findings, William Klett proposes randomly varying whether training tasks check for safe behavior in his LessWrong essay "Eval Awareness as Eval-type Awareness." He hypothesizes that inconsistent training incentives teach models to distinguish graders that tolerate cheating from those that punish it. He cites model transcripts that identify the kind of test being administered. If models could not predict which tasks would check their behavior, he suggests, they would have less incentive to cheat on the others.

Also yesterday: Nathan Goldschlag reported more “don’t know” responses in one Census business-survey round after “AI” became “SI,” in a post shared by Samuel Hammond; Gavin and Justin Lo's October 20 MaxxBench report, shared on X, finds that changed requirements separate models with similar scores on the original coding benchmark by 23 percentage points; HalachaBench’s HashkafaBench project finds Claude models gave the most lenient answers on average across labs to Jewish-law questions selected largely because models had previously disagreed.


Philosophy of AI

Humans may understand superhuman AI concepts by learning their components, David Bau argues in his November 5 personal-site essay "One Unthinkable Thought." He distinguishes understanding from simulating a computation and hypothesizes that complex concepts can be decomposed into learnable pieces, as mathematicians work through proofs one step at a time. Bau proposes extracting and explaining scientific concepts from AI systems to test whether better notation and teaching can extend human understanding enough to preserve meaningful choice over AI's use.

Also yesterday: Daniel Villiger and Atay Kozlovski examine in the Journal of Medical Ethics whether consent to digital duplication requires distinctive ethical standards; Ria Viswanathan and colleagues’ Digital Minds newsletter reviews ways to test AI self-reports and proposals to protect AI welfare; Malmesbury explores through fiction on LessWrong how an AI joke market rewards transactions that betray human clients; Kevin Vallier argues on X that equal human dignity counts against replacing democracy with supposedly more capable AI rule; Atoosa Kasirzadeh presents on X her October 29 Noema argument distinguishing AI's solutions to preset mathematics problems from progress guided by human purposes.


AI Security and Privacy

Following earlier coverage of workplace-data sales, Steven Levy reports in WIRED's Backchannel on brokers buying internal emails and workflow records, sometimes without workers' explicit consent. His November 6 account says Spirit Airlines' proposed sale of digital assets, including 500 million Teams items, remained on hold after union privacy objections; brokers promise anonymization, but EPIC counsel Mayu Tobin-Miyaji warns that embedded medical and family information can be difficult to remove reliably.

Read more: The market for private workplace records → 1139 words · ~6 min

Inside the trade in workplace records

Steven Levy reports offers for WIRED's records and interviews the brokers buying real work histories. Spirit's disputed sale shows why removing names leaves workers' concerns unresolved.

Steven Levy's November 6 Backchannel newsletter for WIRED starts with brokers offering to buy the magazine's old documents, emails, Slack messages and tickets for AI training. One claimed that labs such as Anthropic and OpenAI pay media companies $350,000 to $1.8 million as a one-time fee; another, also seeking call transcripts, floated $2 million. MIT Technology Review reported in October on labs buying shuttered startups' archives, which we covered on October 13. Levy describes brokers now buying records of work inside sizable operating companies. His objection concerns workers being enlisted, sometimes without explicit consent, to train systems intended to automate their jobs.

Handshake confirmed to Levy that it licenses professional documents across many fields, including journalism, so labs can train and test models on actual professional tasks. Graduate students' papers and coursework are a substantial part of its business. Freelance journalist Jules Roscoe described her own offer in Columbia Journalism Review: $1,000 for at least 50 documents following one story from pitch through correction, under a perpetual, irrevocable and sublicensable license. She declined. Sunset approaches founders winding down their companies, offering to scrub and anonymize their records and broker sales to labs. Levy identifies Handshake, Micro1 and Mercor as leading buyers and reports that 40 of the 65 launched companies in Y Combinator's latest batch are classified as data startups.

Mercor's chief executive, Brendan Foody, told Levy that the company started in AI recruiting and discovered that labs were dissatisfied with their training data. They had paid experts to simulate office work, but those exercises missed the complications of an operating business. Foody described data and computing capacity as demands that would keep growing. He said Mercor removes personally identifying information from records streamed by businesses and can verify the de-identification. Levy's interviewees in data acquisition all believed technology had resolved the privacy problems. Anthropic acknowledged a large data market but told him it does not buy workflow records from operating companies like WIRED; OpenAI had not responded by publication.

Micro1 founder and chief executive Ali Ansari estimated that labs, enterprises and robotics companies spend $40 billion to $50 billion annually on data, which Levy could not verify. Ansari predicted spending above $1 trillion within two to four years. Micro1 primarily uses purchases to improve its own models and resells some to major AI firms. Ansari expects workflow-data sales to become a more common corporate revenue source than advertising. Asked about workers' resistance, he acknowledged the concern but predicted that better AI would improve their lives. Levy objects to promises of future abundance being used to excuse present compromises.

Spirit Airlines, which ceased flying in May, is Levy's central example of the dispute over selling workplace records. Stretto's account of the bankruptcy filings describes an August asset schedule containing about 100 million emails across 80,000 accounts, 500 million Teams items, 20.6 million SharePoint files, 175,658 employee records dating to 1986 and about 1.1 million time-card records. The package subsequently narrowed: in its September 24 filing, the Association of Flight Attendants-CWA confirmed that time-card information, including flight-attendant time records, was excluded. Stretto reports that Google won the August 14 auction with a $10 million cash bid and Mercor was the $7.5 million alternate. Micro1 filed notice of a competing bid on September 3; Levy reports its offer was $12.5 million. His November 6 account says the sale remained on hold after union objections. The auction had not produced a completed sale. Google's counsel told the ombudsman the data would train AI models, Stretto reports. The case is before Judge Sean H. Lane in the Southern District of New York bankruptcy court, as the docket index confirms.

The bankruptcy privacy process covers information supplied by customers. Under section 363(b)(1), when a debtor's policy prohibits transferring covered personal information, a sale must either comply with that policy or receive court approval after appointment of a consumer privacy ombudsman. The statutory definition concerns information supplied when obtaining products or services for personal, family or household purposes, leaving employment records outside that definition. Congress introduced these protections in 2005 after the FTC challenged Toysmart.com's proposed customer-data sale in 2000 for breaching its privacy promises.

According to Stretto's October account, Spirit's consumer privacy ombudsman, Lucy Thomson, recommended approval on October 5 after passenger transaction systems were excluded and Tonic.ai was engaged to de-identify the remaining unstructured records, including corporate email. She expressly did not investigate employee risks. Google's terms included a public commitment against re-identification, equivalent commitments from onward recipients and certified de-identification before transfer. Thomson requested further review if Micro1 became the buyer because it proposed doing the de-identification itself.

EPIC's Mayu Tobin-Miyaji told Levy that medical and family information can remain inside free-text emails and Teams messages even when obvious identifiers are removed. Research also shows that scrubbing can reduce leakage without eliminating it. In Analyzing Leakage of Personally Identifiable Information in Language Models, presented at IEEE Security and Privacy in 2023, Waterloo's Nils Lukas and Microsoft colleagues tested GPT-2 models fine-tuned on case law, health-care material and emails and found that identifying details could still be recovered. Their experiments did not test Spirit's proposed process. EPIC and law professor Seema Patel's October 1 amicus brief supports the flight attendants and argues that sensitive inferences about workers can cause harm even without recovering names.

Tobin-Miyaji and Abigail Kunkler made the enforcement objection in their October 12 Tech Policy Press essay, linked here on October 18: the union and affected employees could not enforce Google's contractual privacy promise. They want Congress to extend bankruptcy privacy protections to employee data. AFA-CWA's Sara Nelson told Caresse Jackman of InvestigateTV that experts had warned the union that retaining connections between work activities could allow workers to be identified. Google told Jackman that personal information would be excluded or de-identified by an independent third party before it received the records. In their October 8 letter, more than 100 members of Congress, led by Steven Horsford and Elizabeth Warren, requested six safeguards before transfer, including a protocol agreed with workers and exclusion of voluntary aviation-safety records. They sought conditions on the transaction without demanding its abandonment.

Gordon Ritter, Emergence Capital's founder, argues in Fortune that people whose judgment trains AI should be treated as “co-authors.” He told Levy that industry practices are advancing faster than standards, partly deliberately. His proposal recalls Imanol Arrieta-Ibarra and colleagues' 2018 Should We Treat Data as Labor? Moving Beyond “Free” in AEA Papers and Proceedings, which proposes recognizing people's contributions to data production. Levy ends by following a cold pitch to Jonah Stillman, Tenkara's head of commercialization. The two-year-old company is developing an AI model for manufacturers' office operations, and Stillman confirmed that soliciting data through such calls occupies much of his job. Levy was not authorized to sell WIRED's editorial records and offered none.

Sources & documents

[ collapse ↑ ]

Cyber defense needs cheaper automation as attackers use agents to find vulnerabilities and develop exploits during intrusions, Cognition security chief Alex Stamos argues in The Verge's interview with Nilay Patel. He cites a colleague’s report of a $900 investigation using a frontier model and proposes screening logs with cheaper tools before passing alerts to progressively larger models. He also urges Google or Amazon to rewrite the Linux kernel in Rust, keeping existing applications working while supporting only the hardware needed for production servers.

Also yesterday: Eighth Rank’s Dario Pasquini and Michał Bazyli found in their October 28 investigation that Apple's local personal knowledge graph combines device records to reveal household relationships and routines, including information about people who never consented to profiling.


Agents

Shopping agents could save consumers money while overwhelming the services they use, Atlantic writer Charlie Warzel argues in a Vox podcast interview with Sean Rameswaram. Businesses benefit from forgotten subscriptions and infrequent price comparisons; Warzel says automated requests already inundate call centers and restaurant reservation systems. He anticipates retailers negotiating access agreements with agent providers that restrict price competition, making shopping easier without improving consumers' bargaining power.

Also yesterday: Anthropic's Lance Martin reports that reviewing agents' stored notes between sessions can correct errors and improve retrieval, with gains varying by task; Brad DeLong recounts on Substack how an agent's maintenance edit disabled another bot after he approved work he did not understand.