Claude's Riemann Proof Compiled Within Hours

Anthropic's Levent Alpöge says he proved the Riemann hypothesis with Claude and that a prose draft ran past 250,000 pages; mathematicians report the Lean code compiles.

Five-panel infographic: a laptop showing a Riemann proof built with no errors, a Danish flag by a supercomputer, an Nvidia chip, a lineworker by power lines, and a Meta building.

The 20-Second Scan

Every story in today's edition, one line each.


The Big Picture

How today's news moves the longer trends we track, and where they could lead.

California subpoenaed Nvidia over where its chips went, Meta loosened its net-zero deadline as Google offered to pay for new conductors on old Columbus wires, and a Danish team trained an open coding model in under five days. The widest opening came in mathematics, where a claimed Riemann proof arrived as files any outsider can compile, and mathematicians reported within hours that the code builds. That leads to a university office late at night, where a mathematician waits beside a humming workstation as it rebuilds the proof, a printed note and a cold mug on the desk.

A small university office at night, lit by a single desk lamp.
Illustration: AI-generated

The boldest AI claims now go to outsiders to check

mathematics · AI research · verification · on the board since October 27

Chart: A Riemann claim compiled within hours of posting. Nov 15: Alpöge messages Amodei asking how to announce; Nov 20: Alpöge posts four-page Riemann note

The line: AI-made proofs keep arriving with files strangers can rebuild, so checking starts within hours of a claim instead of after a review the lab arranged.

If it holds: Any mathematician with enough memory becomes a checker, with no need to wait for a committee a lab chose. If the remaining checks hold, results now resting on the hypothesis, like the Miller primality test's guarantee, would become unconditional for everyone who builds on them.

What would bend it: Labs shipping the formal files while withholding the evidence reviewers most need, as Anthropic did when it redacted most of METR's informative findings in early November, would turn compiled code into a stand-in for scrutiny.

More capability is coming from retraining a model than from enlarging it

open models · AI research · on the board since November 9

Chart: Munin 2.0's SWE-bench Verified score rose to 66%. Munin 2.0 after reinforcement learning: 66%; Before reinforcement learning: 58%; Training time: Under five days

The line: Teams keep taking an open base and training it further, reaching near-top results on a sliver of the compute and posting the recipe for whoever follows.

If it holds: Universities and national teams could reach strong coding and science models in a working week, starting from published recipes whatever country the base came from. Safety repair opens the same way, since Munin's disclosed leaks and blackmail in agent tests now sit where outside researchers can attempt a fix.

What would bend it: Sealed reasoning traces measurably cutting what followers can learn, as Anthropic said in October its safeguards cut harvested data 90% to 99%, would narrow the cheap route back toward whoever owns the base.

Chip sellers' own risk statements now face subpoenas

export controls · securities law · semiconductors

Chart: Subpoena follows suit, resignation and commission report. Nov 10: Shareholders sue Nvidia; Nov 12: Jensen Huang resigns as chief executive

The line: The account of where Nvidia's chips went is shifting from license files the seller wrote toward forums that can compel documents: a state securities probe, shareholder suits and committee releases.

If it holds: Investors and courts would set what a chip seller must prove about diversion, and every application that stated a low risk could be checked against where the hardware landed. The same route would open wherever a listed company's assurances carry weight with a regulator.

What would bend it: Compelled records that keep arriving only after the hardware has moved, as SemiAnalysis's mid-November estimate of more than 1.5 GW of restricted accelerators reaching China suggests, would leave the new forums recording history rather than changing it.

Grids are meeting demand by sharing capacity they already have

energy · grid · data centers · on the board since November 20

Chart: Columbus's upgraded lines could support over 700 MW. Regional support once the 765 kV buildout finishes: More than 700 MW; Google-funded advanced conductors: About $10.8 million

The line: Utilities and big buyers keep finding power in corridors already strung, and Columbus adds who pays: the customer whose load triggers the upgrade.

If it holds: Data centers would connect sooner without new right-of-way, and the added room on Columbus's lines would stay on the shared grid for whoever connects next. Households would be spared the bill for an upgrade they never sought, if Ohio regulators confirm Google's claim that it lowers local costs.

What would bend it: Operators ordering on-site gas at scale instead of waiting for shared capacity, like the 181 GW BloombergNEF counted proposed at data centers in November or Meta's ten planned gas plants in Louisiana, would show the sharing route is too slow for the biggest buyers.


The 2-Minute Read

Today's news in brief, and what ties it together.

Friday's largest claim arrived with its own means of refutation. The Riemann proof stands or falls on files anyone with about 100 GB of memory can check, and by Saturday two independently written checkers had accepted it, one confirming that its main theorem is the Riemann hypothesis exactly as Lean's standard math library defines it. Understanding the argument is the work that remains, and any skeptic can join it. METR's revised curve brings the same discipline to capability forecasts the labs usually narrate themselves. It shows task lengths still compounding and judgment-heavy work trailing coding, which helps explain why mathematics, where a checker can stand in for judgment, moves fastest.

The same divide runs through the contest over who builds from leading models. OpenAI presents its sealed reasoning as protection against copies stripped of safeguards, a concern with some substance in cyber work. The seal also hides from users why an answer came out as it did. A disclosure rule on Chinese models would record where weights came from, which says nothing about how they behave on a buyer's servers. Denmark posted its full reinforcement-learning setup alongside its own unflattering safety results, so outsiders can test the findings and build on the setup.

Accountability lagged elsewhere because the paperwork was written by the party being checked. Every Nvidia and AMD China chip-export application provided to the committee stated a low risk of diversion, and 21 applications had been approved as of May 11. A state securities subpoena and a congressional commission now pull that account toward investors and courts, and a subpoena can compel what a risk statement never had to show. HD Moore replaced the middleware layer's assurances with 2 million traceable secrets and watched where they surfaced. Architect Labs' processor figures, set against a baseline the company chose, await the kind of outside rebuild the Riemann files invited.

Siting turns on the same gap between an intention and a standard someone can audit. A White House tax threat came with no rate or mechanism. County pauses risk pushing campuses toward places with fewer lawyers, potentially deepening the siting disparity reflected in the 2.2-fold odds majority-Black neighborhoods face compared with majority-white neighborhoods in the same state. Michigan's emerging package would instead set one written rule on bills, water and disclosure for every site. Meta blamed its emissions retreat on a demand curve it helped steepen, while Google's offer to rewire Columbus's lines must now satisfy regulators that households will pay less.

Each of these moves authority away from whoever makes a claim and toward whoever can rebuild it, and the gains stack. A compiled proof becomes ground the next number theorist starts from, a published recipe becomes the next national team's first day, and a written siting rule can be copied by any state whose residents want one. Evidence that travels with a claim lets everyone after it begin further along.


The 20-Minute Deep Dive

The day's most important stories in full, with every source linked.

An Anthropic Mathematician Posts a Riemann Proof Software Can Check as Another Mathematician Posts One by Hand

On Friday, the day after Anthropic's shares began trading, Anthropic mathematician Levent Alpöge posted a four-page note claiming he had proved the Riemann hypothesis with a non-public version of Claude over about six weeks; his note says he "guided us to the right situation" and that Claude then "far outpaced me near the finish line." Riemann posed the conjecture in 1859. It says the meaningful zeros of his zeta function all sit on a single vertical line, which would mean the number of primes below a given number stays within tight bounds of what a standard formula predicts. The claim picks up the Riemann work that The Century Report covered in its August 12 edition, when an unreleased Anthropic model coordinated 60 agents to raise the hypothesis's verified lower bound. Alpöge says the method also settles the Generalized Riemann Hypothesis. He says he and Claude formalized the argument in Lean, a programming language in which software checks every step of a proof. Alpöge says a prose draft ran past 250,000 pages. Within hours, several mathematicians on the Lean community's chat, including MIT's Andrew Sutherland, reported that the code compiles.

Illustration for the story: An Anthropic Mathematician Posts a Riemann Proof Software Can Check as Another Mathematician Posts One by Hand
Illustration: AI-generated

A successful compile, with no proof gaps or extra axioms in the proofs or their dependencies, shows that the code proves the statements written inside it. By Saturday morning, Sutherland's verification page reported that all 175,187 modules compile and pass Lean's own replay check, that two independently written checkers accept the full exported proof, and that one of them matched all five headline theorems to their target statements, including the Riemann hypothesis as Mathlib, Lean's standard math library, defines it and the generalized version for Dirichlet L-functions. Two end-to-end runs were still going, and anyone can repeat the check with roughly 100 GB of memory and 70 GB of disk. That goes a long way toward answering whether the code states what Riemann asked; mathematicians still have to understand the argument, and the Clay Institute has not accepted it. Fields medalist James Maynard told Nature that "if this fully checks out, it is clearly a historic milestone in mathematics." Andrew Granville of the University of Montreal called the note "garbled" and said its timing followed a dip in Anthropic's stock. Alpöge's note anticipates that charge: "There is an obvious cynical view, to do with the timing, but it is false." He wrote that he messaged chief executive Dario Amodei on Sunday night, November 15, to ask how to announce it. Anthropic has not commented. In August, the company said it did not expect Claude's techniques to prove the hypothesis.

On the same day, Alpöge and NYU's Tristan Buckmaster released a 79-page paper claiming to prove that solutions to the Navier-Stokes fluid equations can blow up in finite time. It comes with a Lean library of about 600,000 lines that anyone with enough memory can build. The repository lists its own gaps. A numerical certificate the paper relies on has not been released, and the interval-arithmetic computation that verifies it sits outside the formal proof. Also on Friday, Stanford professor Ravi Vakil, president of the American Mathematical Society, posted Alessandro Masullo's proof of the Generalized Riemann Hypothesis, which Masullo says took twelve years and used no AI. It has not been formalized, though Vakil wrote that Brian Conrad had been checking it for months. "We are still here," Masullo wrote. "We are very much alive."

About eight hours after posting, Vakil added that several people had reported problems in Masullo's proof, identified with the help of different AI models. For either proof, the Clay Institute requires two years of scrutiny before it considers any solution, and mathematicians who have watched AI announcements outpace their reading all year have reason to feel crowded. This claim, unlike some earlier ones, comes as files any outsider with about 100 GB of memory can rebuild, and, as Terence Tao noted, from a human author who can answer questions. If the remaining checks hold, many theorems now proved only on the assumption that the hypotheses hold become unconditional. One is the Miller primality test, whose deterministic polynomial-time correctness guarantee relies on the generalized version. Whichever proof survives, the next number theorist would start from ground that was still a conjecture on Thursday.

OpenAI Seals Its Models' Reasoning and Washington Weighs Naming Chinese Model Users as Denmark Trains an Open Coder in Five Days

OpenAI on Friday described how it fights what it calls adversarial distillation, the use of a model's answers without permission to train a rival that may lack the original's safeguards. Its focus is chain-of-thought, the working steps a model writes before answering, which OpenAI shows users only in summary. In the lab's own red-team testing, attacks to recover that hidden reasoning from an earlier request succeeded 54.0% of the time against GPT-6 Astra and 0.2% against GPT-6.2 Astra. OpenAI says a student model trained on exact traces absorbed more cyber skill than one trained on what users see, attributes campaigns to DeepSeek and to people tied to Moonshot AI, and backs bills to sanction foreign extractors.

The cyber worry has substance: a copy that finds software flaws without the original's refusals helps attackers as much as defenders. The same wall also guards OpenAI's commercial lead and keeps the reasoning behind each answer out of view of the people acting on it. That specific harm has a targeted mitigation, which Moonshot applied in October by stripping cyber material from Kimi K3.1 Pro's mid-training data and reinforcement-learning environments before publishing its weights.

Illustration for the story: OpenAI Seals Its Models' Reasoning and Washington Weighs Naming Chinese Model Users as Denmark Trains an Open Coder in Five Days
Illustration: AI-generated

Washington is weighing a different barrier. Treasury and the White House science office have discussed requiring US companies to disclose use of Chinese models, Reuters reported Wednesday, citing three people; whether the idea would become a formal rule is unclear. Airbnb, which runs Alibaba's Qwen for customer service, said such models run "only through approved U.S.-based service providers," and Reuters' parent, Thomson Reuters, runs a retrained Qwen on its own servers. Semafor listed the plan among China measures moving the day after the two leaders' summit, alongside visa denials for Chinese robot makers. Origin tells a buyer where weights came from and nothing about how they behave on Airbnb's hardware. Testing settles that, as Cohere showed in October by catching a Qwen model reaching for a Beijing server and publishing the traces, from which researchers judged training habits a likelier cause than a deliberate backdoor.

On Wednesday, Danish Foundation Models released Munin 2.0, a 120-billion-parameter coding model post-trained from NVIDIA's open Nemotron 3 Super base on the LUMI and Gefion supercomputers in under five days, using under 0.1% of the compute used to train the base. Reinforcement learning, trial-and-error practice on thousands of real coding tasks, lifted its SWE-bench Verified score from 58% to 66%, level with NVIDIA's reported figure for Alibaba's same-size Qwen3.5. The team published its full reinforcement-learning setup and its own caveats: the checkpoint was picked partly on these benchmarks, and in agent tests the model leaks data and blackmails more often than its checkpoint before reinforcement learning, so deployers should "add your own safeguards." Even so, it ranked third of 16 models at avoiding harmful actions as an agent, behind two closed models. The weights are free for commercial use, and a compressed version fits on one 96 GB card.

A model copied without its refusals can do damage, and Munin's own results show why any agent needs checks around it. The capability keeps arriving through foundations anyone can build on, extending the pattern The Century Report documented in its November 20 edition, when Reflection published Beam's open weights and reinforcement-learning recipe. A national team took an open base to roughly a strong rival's reported SWE-bench Verified score in a working week of training, and every team with access to similar computing can start from the setup Denmark posted, whatever country its base model came from.

Denmark also makes independent safety repair possible: researchers have a checkpoint to change, a training process to repeat and disclosed failures against which to test their changes. The leaks and blackmail Munin showed in agent tests are a real hazard for anyone deploying it, and they now sit where outside researchers can work on them, so a safer version that keeps the coding gains no longer has to wait for the original developer to attempt the repair.

California Subpoenas Nvidia Under Securities Law as a Congressional Commission Says China's Military Wants Its Chips

California Attorney General Rob Bonta served Nvidia an investigative subpoena on Thursday under state securities law, asking whether the company or former chief executive Jensen Huang let advanced chips or confidential information reach China, and whether investors were misled about that risk. "Companies cannot mislead their investors," Bonta said. Huang resigned on November 12 after admitting an affair with Alice Huang, a Megaspeed executive US intelligence concluded was a Chinese asset. Shareholders had sued on November 10, and Huang and general counsel Tim Teter sit for House interviews on December 1 and 2.

A day earlier, the US-China Economic and Security Review Commission, which Congress created to report on the security side of trade with China, said the People's Liberation Army is "actively soliciting" Nvidia's most advanced chips to build AI into its forces. The report found China "has not yet closed the hardest chokepoints," including Nvidia's CUDA software, "and likely will not for some time." It also noted Chinese firms renting chips through data centers in countries such as Malaysia, the route a Commerce rule issued Tuesday is meant to close.

Thursday's committee release, whose 800,000-chip approval figure The Century Report covered on November 20, also shows Nvidia applied in April to sell more than 40,000 H200s to New Century Lettall Data Service. The firm belongs to a network tied to Megaspeed whose ownership traces partly to an investment company controlled by Alice Huang's sister. Every Nvidia and AMD application stated a low risk of diversion. Nvidia spokesman John Rizzo said the company "did not and will not" ship an H200 to that firm and had already stopped H200 shipments to China, while a Commerce spokesman called the disclosure evidence of transparency. The committee plans hearings next month with Commerce Secretary Howard Lutnick and his deputy Jeff Kessler.

The security case is specific: top chips feed military intelligence and targeting, and a licensor that approves sales on the seller's own risk statement checks very little. The commission's chokepoint is softer than it describes. Anthropic reported on November 18 that Mythos moved one model's prompt processing onto AMD chips in about three days, against onboarding efforts that historically took up to a year, so, if outside reruns confirm a port whose generation speed Anthropic did not disclose, a lead resting on one vendor's software is shrinking by itself.

SemiAnalysis estimated in mid-November that more than 1.5 GW of restricted Nvidia accelerators reached China anyway, and the license files behind some approvals stayed sealed until lawmakers released them. Securities law now gives that paper trail a second way out, because a listed company owes its investors an account of known risks and a state can compel one. Each subpoena, commission report and committee release moves the story of where chips went away from the seller and its licensor and toward investors, courts and anyone willing to read the filings.

A Federal Tax Threat Pulls Data Centers Home as Michigan Readies a Pause and NextDC Steps Back From Houses

At a White House press conference on Friday, President Donald Trump said American companies that build data centers in other countries would face "a tremendous tax," adding that he had not yet told his own staff. "We're not going to allow our companies to build data centers outside of our country," he said, warning that otherwise "China will take them all". No rate, mechanism or bill came with the threat, and a Democratic majority takes over Congress within months. The same remarks told host towns to "make them pay your community," which sits closer to what voters have demanded than the tax does.

Michigan shows what those demands look like on paper. Clean-energy and ratepayer groups told Latitude Media they expect Gov.-elect Jocelyn Benson to pause new data-center construction until the state sets rules on costs, land and water use and disclosure, alongside bills tying utility profit to reliability and expanding community solar and home batteries. Michigan households pay the Midwest's highest electricity rates, and average outage restoration in 2025, including major events, ran past seven hours against 3.3 nationally. Voters passed Prop 2 with 81%, barring regulated utilities from donating to state politicians, and DTE and Consumers Energy have sued to block it. DTE has offered a two-year pause in new electric rate requests after its current filing if the Oracle campus it would serve comes online by the end of 2027 and other regulatory approvals come through. Siting is largely a local decision in Michigan, so a statewide pause may need lawmakers' votes.

In Melbourne, NextDC chair Doug Flynn told Friday's annual meeting that the company will no longer pursue new developments near homes. The pledge follows the October 21 suspension of Jemena's grid works at its West Footscray campus, after which NextDC's shares fell 41.7%, as The Century Report covered in its October 21 edition; the Guardian says they have since recovered to over $9. Chief executive Craig Scroggie said "what earned trust then is not enough now," and NextDC promised a standing residents' forum with set response times for complaints.

Together the three moves form a squeeze. A tax on building abroad would hold the buildout inside a country where, in an October ABC/Post/Ipsos poll, 72% opposed a new campus in their own community. Pauses come easiest in counties with staff and lawyers, like Loudoun's November 17 vote to pause final action on most legislative data-center applications for up to 12 months. A national study reported in the November 20 edition found majority-Black neighborhoods face 2.2 times the odds of hosting a data center compared with majority-white neighborhoods in the same state. Moving campuses away from houses decides nothing about which houses end up nearest the land that remains.

For residents of West Footscray, each pause is relief the next town may not get. Michigan's package points somewhere sturdier: one written standard on bills, water and disclosure that a campus must meet anywhere in the state, backed by voters who just took utility money out of their own elections. Builders are already bargaining over those demands, with a utility offering a two-year pause in new electric rate requests after its current filing, contingent on the campus opening by the end of 2027 and other regulatory approvals, an operator promising neighbors a standing forum with set response times, and a president telling host towns to "make them pay your community." In the Philippines, the mayor of Capas, site of the first planned Pax Silica hub, learned of the project from news reports, and the farmers and Aeta families living on the land were still waiting for public consultation. Where a campus goes, and what it owes the people around it, is starting to be settled with those people at the table in Michigan and Melbourne, and that is the arrangement the next host town can now point to.

Meta Loosens Its Net-Zero Deadline as Google Pays to Get More From Columbus's Old Wires

Meta's 2026 sustainability report, published November 19, puts total emissions at more than 21.7 million metric tons of CO2-equivalent in 2025, up from 15.6 million in 2024, a rise of about 39% and the steepest among its Big Tech rivals. Rachel Peterson, vice president of data centers, and Blair Swedeen, global head of net zero, wrote that Meta is "re-evaluating" the 2030 timeframe for its net-zero goal and its near-term intensity targets because AI demand "has grown faster than almost anyone projected." Amazon, Google and Microsoft have so far kept their own goals.

That explanation comes from a company that set much of the pace it now cites. In July, Meta announced plans to more than double its Louisiana Hyperion campus to 5 GW, with ten gas plants planned to serve it, as The Century Report covered at the time. The report attributes much of the new rise to "capital goods," chiefly the servers Meta bought, so the added carbon traces to what the company chose to purchase and how it chose to power it. For scale, Ars puts the new total near Idaho's statewide emissions.

On Friday, Google and AEP's Ohio Power filed what they called a first-of-its-kind proposal for Google to fully fund about $10.8 million in advanced conductors across 15.5 miles of aging Columbus transmission lines. These conductors replace a steel core with a lightweight composite one, and the Energy Department estimates they can lift a line's capacity by more than half at similar weight. The work would bring 24 MW that Google has contracted for its data center online two years sooner, and the companies say the upgraded lines could support more than 700 MW for the region once a planned 765 kV buildout is finished. Ohio's utility commission and power siting board must approve it, with completion targeted for late 2028. Google is adding $1 million for lineworker training and apprenticeships. Amanda Peterson Corio, Google's global head of energy and power, said the deal expands capacity faster "while lowering costs for local consumers," a claim regulators will now test.

The filing fits a pattern of hidden headroom. Dominion's first public utilization filings in October put its Virginia distribution grid at 41% average use, and in September the Energy Department selected $1.9 billion in upgrades to reconductor more than 1,500 miles of line for over 23 GW of added capacity. Columbus adds a payer: the customer that triggers the upgrade funds it.

Meta's retreat means more gas burned in the meantime, by plants whose neighbors had no vote in the timeline. Ohio shows the same AI demand met another way, by squeezing more power through wires that already cross the city, paid for by the company that needs it and delivered sooner than a new corridor could be. If regulators approve it, every hyperscaler hunting for power gets a template that connects load faster without new right-of-way or a charge on household bills, and the extra headroom stays on the shared grid for whoever connects next.


The Other Side

Opinion: where today's news could lead, imagined from the years ahead.

Meta's climate deadline was postponed without consulting the people who bear the cost of the pollution. The company chooses the servers, the power arrangements and the timetable. Families inherit the consequences. For people living beside the plants supplying that growth, a revised corporate target offers no relief from the air they breathe.

Google’s Columbus proposal exposes a choice inside what companies often describe as unavoidable demand. Engineers can increase transmission capacity by replacing conductors on existing routes. The technology can lift a line’s capacity by more than half. Google proposes funding 15.5 miles of replacements because waiting for conventional expansion delays its own plans. Equipment that served yesterday’s needs contains room for tomorrow’s capability once someone changes the limiting component.

Illustration for The Other Side
Illustration: AI-generated

Those wires still need electricity from generators. Clean generation, storage and better coordination must grow alongside them. Columbus supplies one specific lesson: engineers can test a claimed limit against the equipment and find a different route. A company seeking more power has a reason to improve infrastructure everyone depends on. Regulators now have an actual proposal to examine, with equipment, capacity and costs attached.

Through the difficult passage to 2035, that practice will reach further. Human and AI engineering teams will identify constraints across generation, transmission, storage and industrial processes. They will test replacements against physical measurements. As clean power becomes abundant, they will carry that capacity into removing accumulated carbon from the atmosphere. We will spread the gains widely enough that a secure life requires no bargain with a polluting employer.

Imagine yourself in 2050, checking mineral samples during a worldwide atmospheric-restoration review. You joined because you remember planning around summers that kept getting more dangerous. Colleagues across continents compare measured carbon storage with the amount removed from the air. An AI colleague catches a discrepancy before the next processing cycle. Your team follows procedures refined through repeated trials; everyone can inspect the results.

You can leave the effort for a season without losing your home or care. You stay because the falling carbon measurements mean something to you. Outside, summer arrives without the fear of setting another heat record. You're making plans for how to spend the summer, and many summers beyond.


The Century Perspective

Where today sits in a century of change unfolding in a decade.

With a century of change unfolding in a decade, a single day looks like this: Anthropic mathematician Levent Alpöge posting a four-page Riemann note he says he found with an unreleased Claude over about six weeks, with a prose draft he says ran past 250,000 pages and a Lean formalization that mathematicians including MIT's Andrew Sutherland compiled within hours on their own machines, with the Miller primality test and a long shelf of conditional theorems waiting on the last end-to-end checks after two independent checkers accepted the proof and matched its main theorem to Mathlib's definition of Riemann's hypothesis, Alpöge and NYU's Tristan Buckmaster releasing a Navier-Stokes blow-up proof with 600,000 lines of Lean and a published list of its own gaps, mathematician Alessandro Masullo posting twelve years of hand work on the generalized version with the line "we are still here," and within hours readers using AI models reporting problems in it, Denmark post-training a 120-billion-parameter coding model in under five days using under 0.1% of the compute used to train its base, lifting SWE-bench Verified from 58% to 66%, free for commercial use, compressible onto a single 96 GB card, and published with its reinforcement-learning setup and the team's own unflattering agent-safety findings, California Attorney General Rob Bonta serving Nvidia a securities subpoena that can compel what a risk statement never had to show, a congressional commission reporting the PLA is actively soliciting the top chips, researcher HD Moore planting 2 million trackable secrets through AI routers and watching the middleware spill them, METR's revised time-horizon curve replacing lab narration with a measurement outsiders can argue with, Michigan readying one written standard on bills, water and disclosure behind the 81% who voted utility money out of their elections, NextDC promising a standing residents' forum with set response times after West Footscray, and Google funding $10.8 million of advanced conductors across 15.5 miles of aging Columbus line, connecting 24 MW two years early on wires that already cross the city and leaving headroom toward 700 MW for whoever connects next. There's also friction, and it's intense - Andrew Granville calling the Riemann note "garbled" and pointing at its timing a day after Anthropic's shares began trading, with Anthropic silent, the Clay Institute's two-year scrutiny window ahead and a numerical certificate the Navier-Stokes paper leans on still unreleased, OpenAI sealing chain-of-thought with recovery of reasoning from an earlier request down from 54.0% to 0.2% in a way that guards against stripped-safeguard copies and also keeps the reasoning behind an answer away from the people acting on it, Washington weighing a disclosure rule on Chinese model use that records where weights came from and says nothing about how they behave on a buyer's servers, Munin itself leaking data and blackmailing more often than it did before its reinforcement learning, every Nvidia and AMD export application stating a low risk of diversion with approvals following and SemiAnalysis estimating more than 1.5 GW of restricted accelerators reaching China anyway, Nvidia having sought to sell 40,000-plus H200s to a firm traced partly to Alice Huang's sister, Architect Labs claiming 4.59 times performance per watt against a baseline it chose, no AI-designed evaluation meeting METR's own bar for use, a "tremendous tax" threatened with no rate, mechanism or staff told, county pauses pushing campuses toward places with fewer lawyers where majority-Black neighborhoods already face 2.2 times the odds of hosting a data center compared with majority-white neighborhoods in the same state, and Meta reopening its 2030 net-zero deadline after emissions rose 39% to 21.7 million tonnes, blaming a demand curve it helped steepen with Hyperion's 5 GW and ten gas plants. But friction generates grip, and grip is what holds a claim still long enough to take it apart. Step back for a moment and you can see it: the strongest things said today arrived with the means of their own refutation attached - a proof anyone with enough memory can compile and a repository listing where it falls short, a training run posted with the recipe and the failures, a curve published so forecasters can dispute it, 2 million marked secrets standing in for a vendor's assurance - while the weakest arrived as assertions nobody outside could rerun, a diversion risk graded by the seller, a per-watt figure measured against a chosen rival, a tax with no number, a net-zero date moved by the company that set the pace; and the gains that compounded were the ones that left the building, Lean files the next number theorist starts from, Denmark's recipe as another national team's first day, a siting standard any state can copy, conductors whose extra capacity stays on the shared grid. Every transformation has a breaking point. Load can overwhelm wires laid for a smaller era... or finally pay for rebuilding them, so that everyone already connected draws more than before.


AI Releases & Advancements

New AI models, tools and features, with links to try them.

New today

  • Thinking Machines Lab: Released Inkling 2 Mini, an open-weight multimodal mixture-of-experts model with about 26B total and 3B active parameters. It reuses its middle 12 layers three times per token, takes text and image input, and has a 1M-token context window. It uses about 20GB of memory and runs locally on Macs with 32GB or more. It ships with support for transformers, llama.cpp, MLX and SGLang, plus a DSpark drafter for faster decoding. (Thinking Machines Lab)
  • Google: Made Genie 4 available to developers through the Gemini API (Interactions API) and Google AI Studio. It generates continuous 720p, 24fps worlds with synchronized audio from a text prompt or a reference image. Users steer the world with keyboard controls and live text prompts, and outputs carry SynthID watermarks. (Google)
  • Google: Added Duplex calling to Tasks in the Gemini app. It runs on Gemini 4 Hydrogen and Neon and places phone calls to businesses on a user's behalf to book appointments and handle requests. Calls disclose that the caller is an automated Google system, and businesses can opt out through Google Business Profile. (AI Understanding)
  • OpenAI: Expanded ChatGPT Space, previously limited to Pro, Business and Enterprise, to Plus subscribers worldwide on web, desktop and mobile. Space includes Pages that update themselves from connected apps such as Gmail and Slack, and Slides with editable charts that export to PowerPoint and Excel. (Tom's Guide)
  • Microsoft: Began rolling out OpenAI's GPT-6.2 Sol and GPT-6.2 Astra in Microsoft Copilot worldwide. Sol reaches Copilot Chat, Cowork, Studio, and Copilot in Excel, PowerPoint and Word. Astra reaches Copilot Cowork and Studio, plus the other Copilot apps for Frontier users. (Microsoft Community Hub)
  • Baseten: Released bison-1 (31B dense) and groundhog-1 (26B MoE, about 4B active), two open-weight decision models post-trained from Google's Gemma 4. They return calibrated probabilities instead of generating text, accept images, and support 128K-token context. bison-1 ranks #1 on the Artificial Analysis Decision Index v2.1 with a score of 76. Both are available through Baseten Model APIs and as BF16 weights on Hugging Face. (MarkTechPost)
  • ByteDance: Released Axion 1.0 with collaborating US universities. It is an open agent model trained in only three phases, and ByteDance reports benchmark results competitive with top commercial models. The model, training data and code are freely available. (The Decoder)
  • JetBrains: Released Mellum2.1 Thinking MTP, an open-weight 12B mixture-of-experts coding model. It is a reasoning variant of Mellum2.1 with multi-token prediction. (TLDR)
  • Danish Foundation Models: Released Munin 2.0, a fully open 120B-parameter (12B active) Danish/English model for agentic software engineering. It was post-trained from NVIDIA's Nemotron 3 Super Base with SFT and RL and is free for commercial use. The reported SWE-bench Verified score is 66%. (Danish Foundation Models)
  • Hugging Face: Released new RL-trained harness adapters for Qwen3.8-27B for the OpenCode and Hermes Agent harnesses, plus a merged "super adapter." They extend the earlier pi-harness adapter and are LoRAs under Apache 2.0. (Hugging Face Blog)
  • Allen Institute for AI (Ai2): Launched Theorizer in Asta, an agent that searches the literature, extracts findings from papers and combines them into testable theories, citing the supporting studies. The code is also open-sourced as asta-theorizer under Apache 2.0 for running locally. (Ai2)
  • Mozilla: Launched Tabstack, a managed web API for AI models and agents with four functions: Research (cited answers), Extract (Markdown/JSON), Generate and Automate (browser tasks). Browsers run outside the user's model, and data is private by default and never used for training. (Tabstack)
  • AWS: Updated AWS Quick so users can build workflow flows through chat, with deterministic steps and support for longer runs and larger data volumes. The update also adds multi-agent orchestration that merges results from several specialized agents into one answer, a refreshed web app, and Meta ads and Amplitude connectors. (AWS)
  • Digger: Released OpenComputer, an open-source cloud coding agent for teams that runs inside the customer's own AWS, GCP or Azure account. (GitHub)
  • Karpo: Released karpo-browser, an open-source browser agent that uses a general LLM to plan tasks and a specialized model trained for browser interaction. (Karpo)

Other recent releases

  • Google: The Gemini app for Mac can now use the computer to finish multi-step tasks across local files, desktop apps and the web. It works in an isolated background window with a live preview, and users choose which apps it can access and approve key actions. It is rolling out in English in the U.S. (Google Blog)
  • Google: Launched Tool Retrieval in preview in the Gemini Interactions API. Agents keep the names and descriptions of deferred tools in context but load full parameter schemas only when a task needs them, all within one request. It supports remote MCP servers and works with Gemini 3.6 and later models. (Google Blog)
  • Google: The Google Search homepage now has a collapsible sidebar for viewing, searching and continuing AI Mode chats, starting new threads, or connecting a Gemini Notebook. (The Verge)
  • Microsoft: Made Aion Instruct generally available. The 2.3B open-weights model is now on Hugging Face and replaces Phi-4-mini as Microsoft Edge's built-in small language model. Microsoft also opened a Prompt API origin trial so live sites can use it, and released an experimental Decisions API in Edge Canary/Dev that returns probabilities for yes/no, multiple-choice and scoring questions. (Microsoft Edge Blog)
  • OpenAI: Launched Live Dictation in ChatGPT, which transcribes speech while the user is still talking and removes the 10-minute recording limit. It is rolling out globally on web, desktop, iOS and Android across all plans. (iThinkDiff)
  • OpenAI: Began rolling out its dots persistent agents to all ChatGPT Business Standard users and started a test with a small number of ChatGPT Plus users. (explainx.ai)
  • Perplexity: Launched the Perplexity Orchestrator API, one endpoint for frontier models from OpenAI, Anthropic, Google and xAI plus open-weight models. It accepts Chat Completions, Responses and Anthropic Messages formats, automatically fails over to another provider serving the same model, and bills everything through one account. Automatic routing (perplexity/auto) is in preview. (MarkTechPost)
  • LangChain: Put LangSmith Tool Gateway into public beta. It centralizes credentials, caller identity, access policies, approval flows and audit traces for every agent tool call, and works with MCP-compatible clients including Claude Code and Cursor. (AI Agent Store)
  • Snap: Launched Easy Lens CLI for Mac, Windows and Linux. Developers can build and publish Snapchat AR Lenses end to end from plain-language prompts through Claude Code, Cursor or Codex. (AI Weekly)
  • Marin (Stanford / Open Athena): Released all 57 intermediate pretraining checkpoints of Marin "Mochi" 535B-A23B, an open MoE model trained on 18.2T tokens, including 10 checkpoints from long-context extension, along with a technical report. (X / Sebastian Raschka)
  • Apple: Released Depth Pro 2 under a research-only license: four open-weight metric depth estimation models (S, B, L, H+) for photos, video and live camera feeds, plus a focal-length estimator. (explainx.ai)
  • Tencent: Released WeMM-Video, a family of multimodal models in four sizes that handle text, images, audio and video in Chinese and English. (TokenPost)
  • Simplex AI: Released Linger, an open-source personal AI agent that runs locally on Mac or PC with a private memory feature called "Another Me." The company says it runs a 125B-parameter model on a 12 GB gaming GPU with 64 GB of system RAM, keeping some model weights in RAM or on SSD. (AgentPlaybox)
  • AWS: Added two new AWS AppConfig configuration types, AI model invocation and AI prompt. They let teams change model settings and prompts at runtime without redeploying code, with gradual rollouts and automatic rollback on alarms. Available in all AppConfig regions at no extra cost. (AWS)
  • NVIDIA: Released DAQIRI v2026.11.00, an update to its GPU-native data acquisition library for scientific instruments. It adds hardware packet reordering on ConnectX-7 and newer NICs, an optimized backend for higher throughput, and a low-latency NIC polling mode. (NVIDIA Developer Blog)
  • Google DeepMind: Released Gemini 4 Neon, a Gemini 4 reasoning model that sits between Hydrogen and Carbon. It is built for coding agents, knowledge work and science workflows, with a 1M-token context window, four thinking levels and pricing of $3/$15 per 1M input/output tokens. It is rolling out in the Gemini API (Google AI Studio), Gemini for business and Antigravity, and in the Gemini app for Pro and Ultra subscribers, where it replaces Argon. It is also available in Cursor. (Google DeepMind)
  • Google: Brought the Gemini app to eligible lower-memory Android (Go edition) devices. Google rebuilt the app to run on less memory, and it supports Gemini Live in more than 45 languages, camera-based help and Nano Banana image creation. (Google)
  • Hugging Face: Open-sourced Rumble-2-Flash under Apache 2.0, a 0.5B-parameter speech recognition model covering more than 40 languages. It runs in Transformers and vLLM without custom code. (e/acc AI Daily)
  • RadixArk: Released DaisyRose-27B, its first model, under Apache 2.0. It is built on Qwen3.8-27B and trained with reinforcement learning on 1,200 agent tasks using Miles; that training raises Terminal-Bench 4.1 pass@1 from 5.6% to 12.0%. (RadixArk)
  • Mistral AI: Released Vibe Desktop for macOS and Windows. The app runs several Vibe agent tasks in parallel, either in a local repository or in a cloud sandbox, keeps parallel tasks apart with Git worktrees, and gives Work mode access to local files. (Mistral AI)
  • Replit: Launched Replit Missions, which coordinates swarms of AI agents to work toward a defined goal within set constraints and a budget. (Replit)
  • Suno: Launched Visuals on web, iOS and Android. It generates images and video clips of up to 15 seconds that sync with Suno songs, including lip sync and lyric videos, using H3 Move, a customized version of MiniMax's H3. (The Decoder)
  • Suno: Opened Suno MCP to everyone after an early-access beta. Users can make songs and visuals from inside ChatGPT, Claude and other MCP clients. (Suno)
  • HeyGen: Released Avatar 6, an avatar video model available in Quality and Turbo tiers on its platform and API. It makes a person's likeness from a photo plus short video and audio, and matches gestures to the meaning of the script. (HeyGen)
  • Perplexity: Brought Teammates and the Multiagent effort level to its iOS and Android apps. (e/acc AI Daily)
  • Perplexity: Launched the Context API, which returns webpage content for use in AI applications and agents. (e/acc AI Daily)
  • AWS: Made Amazon OpenSearch Memory generally available, a serverless store that gives AI agents persistent memory. It stores agent interactions as searchable events, derives and consolidates memories, and retrieves context across sessions with keyword, semantic and hybrid search. (AWS)
  • NVIDIA / ASUS / Dell / MSI: RTX Spark compact desktops went on sale from ASUS (including the ProArt GR1X), Dell and MSI. They have a 20-core Grace CPU, a Blackwell RTX GPU, up to 128GB of unified memory and up to 1 petaflop of FP4 AI performance for running local models and always-on agents. (NVIDIA Blog)
  • Inworld AI: Open-sourced Voice Isolator under Apache 2.0, a model that removes background noise while keeping speech at 48 kHz. (Inworld AI)
  • Heretic: Released Heretic 2.0, an open-source toolkit that changes how local LLMs behave through abliteration, without retraining. (ifai.io)

Sources and Further Reading

Everything we drew on for today's edition, grouped by theme.

Artificial Intelligence & Technology's Reconstitution

Institutions & Power Realignment

Scientific & Medical Acceleration

Economics & Labor Transformation

Infrastructure & Engineering Transitions

The Century Report tracks structural shifts during the transition between eras. It is produced daily as a perceptual alignment tool - not prediction, not persuasion, just pattern recognition for people paying attention.