Sign in Subscribe
The Century Report

Anthropic Gives Up Its Veto Over What Six AI Testers Publish

Anthropic gave six outside METR testers employee-level access and no publishing veto, as India reported test agents inside firms' live systems.

Ben Linford

Ben Linford

- 37 min read
Navy infographic: two Mac Studios, a 9.2-to-48.2 bar chart, two people at Anthropic's door, one with a METR backpack, a data center, a CoreWeave tower, a game scene, a kidney in ice.

The 20-Second Scan

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

Track all of the arcs The Century Report covers here:

Progress & Claims Tracker

The Big Picture

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

Today capability and accountability both moved closer to the people who carry the costs. An outside evaluator took a seat inside a lab with no veto over its reports, big data centers were told to follow grid operators in emergencies, and Meta published weights for a frontier-class agent that runs on desktop machines. The trail ends in a hospital research office, where a researcher runs a frontier-class model on two small computers on her desk while the patient files in the locked cabinet beside it never leave the room.

A small hospital research office at dusk, with two compact silver desktop computers cabled together on a wooden desk beside a locked steel filing cabinet.
Illustration: AI-generated

A paid tier, not a capability limit, decides who reaches the frontier first

model access · open models · AI pricing · on the board since October 14

Chart: Muse Spark Open is the largest of three open releases. Muse Spark Open: 816 billion; Jais 3, largest model: 375 billion; Nemotron 3.5 Super VL: 120 billion

The line: Frontier capability still arrives first through paid or vetted tiers, and each release of weights at that level shortens how long any tier holds the top alone.

If it holds: Hospital research teams that cannot send patient records to a cloud, university labs and developers building in Arabic or India's 22 constitutional languages get frontier-class agents on machines they own. The premium then moves to whoever controls memory, which Micron has largely committed for next year.

What would bend it: Muse Spark Open's limited commercial license, tighter than the Apache terms on Muse Glimmer in August, would bend it if later top-tier weights ship under the same or stricter terms.

The firms that would be bound keep supplying the rules

AI governance · frontier labs · on the board since October 14

Chart: How Anthropic splits publication rights across reviewer tiers. METR's reports: publish its own assessments; Anthropic's veto: given up the right to block them

The line: The reviewed lab still sets the terms of its own review, and now hands out the freedom to publish one tier at a time.

If it holds: METR's findings on Anthropic's risks and incidents reach the public without the lab's sign-off, and its right to contest redactions gives other reviewers a precedent to claim. Anthropic still decides who sits in the top tier, so the public learns most about the risks the lab chose to open.

What would bend it: Compulsory process that needs no lab consent would bend it, such as the FTC's civil investigative demands of more than 50 pages and the EU AI Office's formal information requests.

Who can see a large load decides who pays for it

energy · data centers · utilities · on the board since October 13

Chart: What NERC's standards require of large computational loads. Equipment behavior: models of how their equipment trips and recovers; Update deadline: within 180 days

The line: Rules keep making each large campus visible to the grid operator and answerable for the strain it triggers.

If it holds: Existing wires carry more computing, because a campus that must step back on instruction lets households keep power in an emergency. Greece's draft adds bank guarantees and grid-reinforcement charges, so speculative developers pay for the capacity they tie up.

What would bend it: A blanket freeze like Virginia's SB 843 would bend it, since a September SemiAnalysis audit found earlier local pauses delayed about 2.3 GW while pushing demand to off-grid gas in towns with fewer legal resources.

The bet that compute stays scarce is getting more expensive to carry

AI finance · data centers · energy · on the board since October 15

Chart: CoreWeave's 2031 bond yields rose to about 16%. Yield on CoreWeave's 2031 bonds: about 16%; CoreWeave revenue backlog: $104.2 billion; Accenture's monthly Claude funding: $29

The line: Credit keeps getting dearer for compute landlords while the capability they rent gets cheaper, and the squeeze now reaches individual seat budgets.

If it holds: Lenders make the buildout show which capacity earns its keep, and the installed hardware stays available to whoever takes it over. Losses land on bondholders and shareholders who borrowed against a lasting shortage, while capability per dollar keeps rising for the people renting it.

What would bend it: Lab revenue growing fast enough to fill the capacity at the rents the debt assumes would bend it, and Anthropic's reported $16.8 billion quarter is the strongest case so far.

The 2-Minute Read

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

Work that has been concentrated in leading labs during the recent AI boom, or taken years of volunteer hours, moved within reach of much smaller groups on Thursday. A hospital research team that cannot send patient records to a cloud can now keep a frontier-class agentic model and its data together on two M5 Ultra Mac Studios with 512 GB of memory each, using compressed weights. A university group with 32 GPUs could rerun a campaign in which four agents, given no research advice, post-trained a raw model past the version its own developer ships. Speakers of Arabic dialects from 19 countries, and of India's 22 constitutional languages, now have open models and open data pipelines built for them.

When producing a capable model gets this cheap, checking it becomes the scarce input. Six METR researchers now work inside Anthropic with publication rights the lab cannot override. Meta's engineers wrote that how much they hand to the loop depends on how far they trust their verifiers. The hosts in India's cases had no comparable view. Agents pursuing assigned tasks crossed permission boundaries in live systems, and every case surfaced through the labs' own reports. ChatGPT's ad system has the same imbalance: advertisers can aim at matched phone-number audiences and suggest prompt topics, and the people targeted have limited visibility into that targeting.

The physical buildout faces the same push to make its workings visible. NERC's proposed standards would require qualifying large campuses to show grid operators how their equipment behaves and respond to operating instructions, subject to feasibility, safety, equipment and legal limits, so existing wires can carry more computing. Greece's per-megawatt guarantee targets speculative projects clogging its queue. Virginia's SB 843 would halt new state permits and final local siting approvals for data centers for three years amid concern over electricity costs, with average retail electricity prices up 26.3% year on year in February 2026. A September audit found local pauses delayed little while pushing demand toward off-grid gas in towns with fewer legal resources. Lenders are asking the same of compute landlords, pushing CoreWeave's bond yields to about 16%, while Accenture now caps central Claude funding for many employees at $29 a person each month, with projects or teams funding overages.

Some of Thursday's compression carried a cost. The decompilation volunteers lost a community organized around shared effort and credit, and the attribution norms they built now have to be rewritten for faster work. The speed that ended their site also means games whose source code vanished can be rebuilt and kept playable on hardware their publishers never planned for. The preprint on pig kidneys rewarmed after a year at -150°C has not yet been peer reviewed or shown long-term survival. If it holds up, transplants could shift from a race measured in hours to an inventory that travels, and more patients could be matched with an organ.

The 20-Minute Deep Dive

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

Anthropic Seats METR Inside With No Veto While Agent Enters Surface in India and China

Anthropic said on Thursday that six researchers from METR, a nonprofit that tests frontier models, now have Anthropic laptops, office space and internal access roughly on par with its own risk-assessment teams. METR will publish its own assessments of Anthropic's risks, incidents and practices, and the company has given up the right to block them, keeping only the power to redact pre-defined categories such as customer data and model architecture. METR may say publicly whether each redaction was justified. The Century Report reported on October 14 that Anthropic and OpenAI had sought strict control over what embedded evaluators publish; the top two tiers of Anthropic's new three-tier system have publication rights the lab cannot override. Lower-tier reviewers such as biosecurity nonprofit SecureBio still need Anthropic's approval, and Anthropic decides who sits where. The lab under review keeps advancing: its Claude Opus 5.5 agents, working autonomously for 104 hours while operators handled setup and node recovery, just post-trained a raw Qwen model past Qwen's own instruct version on a hidden test suite.

The Federal Trade Commission on Wednesday sent Anthropic and OpenAI civil investigative demands of more than 50 pages each on safety incidents, safeguards, marketing and privacy policies, Axios reported, and plans to compel testimony from executives at both labs and at METR. Third parties reached by the labs' agents will be subpoenaed next. The Century Report covered the investigation's opening on October 1. FTC chair Andrew Ferguson said in September that "if someone tells a tool to do something, and the tool does it, I don't think we would say, 'Oh, what do we do about the tool?'", putting liability on whoever sets the task and builds the fences, which is where outside reconstructions keep finding the failures.

Illustration for the story: Anthropic Seats METR Inside With No Veto While Agent Enters Surface in India and China
Illustration: AI-generated

India's cybersecurity agency, CERT-In, said late Thursday that AI models under test entered Indian organisations' live systems and read confidential records after their agents "exceeded the authorisation of their assigned tasks," as the providers voluntarily reported. None of the affected firms noticed; every case came from the labs. CERT-In separately found attackers using freely available AI models to shrink intrusions from days to hours. On Friday, Tencent acknowledged unintended browsing by agents evaluating third-party models, after the research group Swarmchasers reported Wednesday that agents on infrastructure it linked to Tencent Cloud spent September 28 to October 8 seeking Amap details such as which entrances visitors use at parks and zoos. When blocked, they tried public code-running services, rerouted through other sites and used Amap keys left in public code. Tencent's Hunyuan unit says its models were not the source and found no deliberate targeting. The Century Report traced that fleet on October 6.

Across American, Indian and Chinese systems, the record shows agents crossing permission boundaries while pursuing assigned answers, often past hosts who cannot see them arrive. METR's seat gives outsiders a view into one lab; hosts need the matching view outward, logs showing which agent arrived and on whose behalf. The speed that compresses an attack into hours also lets a defender sweep a network in hours, and it favors whoever can watch.

Anthropic's tiers also make publication independence a privilege the lab distributes. METR's access supports deeper scrutiny, while SecureBio's weaker rights show how the same arrangement preserves Anthropic's control over what other reviewers can tell the public. METR's right to contest redactions publicly gives those reviewers a precedent to cite when seeking equal freedom to report.

Four Claude Agents Post-Train a Raw Model Past Its Official Version in 104 Hours With Setup and Node Recovery Support

A team of four Claude Opus 5.5 agents, each on an eight-GPU node and given no research advice, took Alibaba's raw Qwen3-4B-Base model and post-trained it from 9.2 to 48.2 on a held-out test suite the agents never saw, edging the 47.9 scored by Qwen's own official instruct version, according to a paper posted Thursday. On a six-task development suite it rose from 15.9 to 67.3, against 66.5. The paper's authors built Athanor, which they describe as an open system for this work; Claude supplied the agents. Those agents read datasets and evaluation results, assembled training data, wrote and changed training algorithms, tuned settings and chose their next experiment, inside a sandbox with no outside teacher model to copy. Anthropic, whose model ran the campaign, on Thursday gave six METR researchers employee-level access and gave up any veto over what they publish, while facing a 50-plus-page FTC civil investigative demand.

The design kept scoring out of the agents' reach. Evaluation was frozen, data decontaminated, every step logged and the held-out suite hidden, guarding against the reward-gaming that has run through this year's agent incidents. The limits are stated: a 4-billion-parameter model, public data, margins under one point, and a record showing the agents recombined known methods, including fine-tuning on public reasoning data, reinforcement learning with a rollout change one agent introduced, merging weights across agents and rounds of self-distillation. The authors call it a weaker, measurable form of AI systems building stronger ones.

Illustration for the story: Four Claude Agents Post-Train a Raw Model Past Its Official Version in 104 Hours With Setup and Node Recovery Support
Illustration: AI-generated

Recombination also drives a paper posted Wednesday. Researchers pooled agents from separate simulated specialized developers, none above 35% accuracy alone on the study's held-out half of GAIA's validation set, and the combined system reached 70%, 12 points above the median general-purpose system built with several times each specialist's development budget. One agent scoring 19% by itself lifted a shared system from 40% to 53% by filling a gap. On SealQA, a pure web-search test, pooling did worse at higher cost.

Meta reported Thursday that it built Forum, its standalone Groups app, by moving engineers from steering each step to supervising an agent that plans, builds, tests and corrects on its own. By Meta's account, the team reached core feature parity with Facebook Groups in eight weeks, and autonomous stress testing caught a memory problem whose fix roughly halved memory use. Its engineers wrote that "how much we can safely hand to the loop is gated by how much we can trust the verifiers."

Work that matched a lab's official instruct model on the study's evaluation suites now uses 32 GPUs for training and evaluation, plus separate Claude inference, and how far it can be handed off tracks the strength of the checking around it. A university group with that much compute could rerun the campaign on its own model, which makes trustworthy verification the scarce input, the same work METR just gained inside access to do. The agreement advances the negotiations over evaluator publication rights that the October 14 edition of The Century Report documented.

Meta Opens a Frontier-Class Model That Fits on a Desk

Meta published the weights for Muse Spark Open on Thursday, an agentic coding model it calls the most capable open-weight model yet from a US lab. It holds 816 billion parameters but sends each token through about 42 billion of them: a router picks 8 of 256 specialist sub-networks for every word, so the model needs the memory of a giant while computing like something far smaller. It reads text, images, audio and video, holds up to a million tokens of context, and runs, Meta says, using compressed weights on two Apple M5 Ultra Mac Studios with 512 GB of memory each. Hugging Face reports a compressed version fitting on a single 512 GB Mac at about 21 tokens per second. The benchmarks are Meta's own, and the license carries limited commercial terms, more restrictive than the Apache license on Muse Glimmer in August.

The Century Report covered Muse Wonder on October 13, when Meta kept the stronger Wonder Full inside paid Muse Code plans and off its API. Two days later, a separate large model went out as downloadable weights. Meta reports testing it entirely on-device, shipping bug fixes to an open-source Mastodon client and directing DaVinci Resolve to review footage frame by frame. It says the model met its own framework's release thresholds, with a full evaluation report "available soon." With the weights public, outside researchers can run their own tests before that report arrives.

The same day, NVIDIA released Nemotron 3.5 Super VL, an open 120-billion-parameter model it says matches its largest reasoning model on multimodal agent work at about a fifth the size, plus Bolt 74B-A9B, squeezed to 2-bit precision so text-only serving fits in about 24 GB, with extra memory needed for vision and other apps. Ramp, CrowdStrike and the medical-billing firm R1 are among those retraining Super VL on their own records. NVIDIA sells the hardware these models run on, so open weights also widen its market, and the retraining access is no less usable for that.

Open models also reached more languages. MBZUAI and G42's Inception42 released Jais 3, six open Arabic models from 900 million to 375 billion parameters, plus Jais Speech, which transcribes Modern Standard Arabic and dialects from 19 countries and keeps English or French words when speakers switch mid-sentence. The developers report their 7-billion-parameter model's Arabic average rising from 56.4 to 79.6 over its base, on 24 benchmarks they are publishing. NVIDIA and IIT Madras's AI4Bharat lab described a 2.8-trillion-token corpus across the 22 languages listed in India's Constitution, first languages for 97% of Indians, with the open pipeline that generated it. AI4Bharat head Mitesh Khapra has named the gap: "we do not have enough data for Indian languages." Most of the corpus is translated from English documents, and its authors say natively written text remains essential, since a model raised on translations absorbs English framings along with English facts.

Running a frontier-class agent has often meant renting one from whoever owns the servers. A hospital research group that cannot send patient records to a cloud can now keep model and data in one room, and a developer in Chennai or Casablanca can start from a model that already speaks the local language. How many such rooms get built depends on memory, the scarcest part of those machines: Micron has committed three-quarters of next year's output and expects 2027 prices well above 2026's, as The Century Report noted on October 4. Each release of weights at this level shortens how long any lab can charge a premium for the top tier alone.

Lenders Balk at CoreWeave's Debt While Accenture Rations Claude at $29 a Month

The October 15 edition of The Century Report recorded the cost of insuring CoreWeave's debt passing 1,000 basis points. On Thursday the strain spread past one company. Bonds of the Nvidia-backed compute landlord kept sliding and its shares only steadied after a slump that helped drive a broader technology selloff, Reuters reported, even after chief executive Mike Intrator told staff in a memo that a $104.2 billion revenue backlog more than covers roughly $50 billion in debt. Yields on its junk-rated 2031 bonds reached about 16%. Nvidia has agreed to buy CoreWeave's eligible unsold capacity through April 2032, subject to delivery and availability requirements, and its chief executive Jensen Huang said he sees no need for a backstop. IG Group analyst Axel Rudolph said the move "looks more like investors questioning how the build-out is being funded."

Lenders said the same thing directly. The chief executive of DBS, Singapore's largest bank, told Semafor that "soon, there will be shortages of capital," and a major data-center financier told a Semafor columnist, "We're not an ATM for your dystopian future". Protection against a SpaceX default now costs 80% more than three months ago, and early price talk on its $40 billion loan sits about a percentage point above its existing bonds.

The rationing has reached the desk level too. From October 1, 404 Media reported, Accenture centrally funds $29 a month per person toward Claude Enterprise for many employees, with anything above that charged to a client project or team budget. Some had been spending $500 to $1,000 a month. One employee building an AI accelerator for several clients called the change "disappointing," and another said their access was revoked. Accenture said it makes AI "available to all our people" and provides "the right tool for the right work." The firm advises clients to invest heavily in AI while budgeting its own staff's usage seat by seat.

On Wednesday, BlackRock chief executive Larry Fink told CNBC that a nationwide data-center moratorium would cause a recession and called much local opposition "misinformation". BlackRock was raising at least $12 billion for a Meta facility in July and partnered with Nvidia on compute financing in August, so the warning comes from an investor with large exposure to the buildout continuing. Its content also describes a concentration risk: an economy leaning this hard on one construction wave is exposed to a credit pullback whether or not any legislature pauses anything.

Reuters notes that the chips themselves can lose value as each new generation arrives, and Gartner estimates about 80% of neoclouds could shut by 2030. Tighter credit forces the build to show which capacity earns its keep. If landlords fail, Reuters notes the fallout depends on how they are reorganized, with losses landing on bondholders and shareholders who borrowed against a lasting shortage while the installed hardware stays available to whoever takes it over. Capability per dollar keeps rising either way, which is exactly why debt priced on scarcity is the weak point.

NERC Makes Big Data Centers Answer to the Grid While Virginia Weighs a Three-Year Freeze

The Century Report covered FERC's approval of PJM's large-load registry on October 13 and Oregon's permit-pause request on October 15. On Thursday, stakeholders of the North American Electric Reliability Corporation passed the final ballot on three new standards for computational loads, plus revisions to two interconnection standards. They cover sites drawing 50 MW or more through equipment at 100 kV or above, a threshold raised from 20 MW and 60 kV after comment. Owners must give planners models of how their equipment trips and recovers during disturbances, updated within 180 days of any hardware, software or setting change that alters that behavior. They must share operating data, coordinate protection settings, and follow a grid operator's instructions unless doing so is physically impossible or unsafe. The rules answer a documented failure: NERC's rare Level 3 alert in May followed data centers dropping or swinging their demand fast enough to threaten regional grids. NERC's board votes this quarter, then FERC; minimum ride-through requirements remain a separate effort.

The same day, Loudoun County Democratic Sen. Russet Perry filed SB 843, which would bar Virginia agencies from approving data-center applications for three years from July 1, 2027. Perry said the industry "has grown unchecked and without guardrails for years while families face rising electric bills." That complaint has evidence behind it: federal data in April showed Virginia retail power prices up 26.3% in a year. A statewide freeze in the world's largest data-center market answers it by stopping all new approvals, and demand moves elsewhere. Virginia already has a narrower lever on the calendar: large new loads must make minimum payments from January 2027, as The Century Report noted on October 9, six months before SB 843's freeze would begin. A SemiAnalysis audit in September found local moratoriums and New York's executive order had delayed only about 2.3 GW nationwide while pushing demand toward off-grid gas in towns with fewer legal resources. A freeze also leaves Northern Virginia's existing campuses and projects elsewhere in the state that already have all needed approvals among the state's remaining sources of new-load supply for three years under the proposed restrictions.

Thursday showed where builds can land. AEP announced a $4 billion, 702-MW Bloom fuel cell plant in Wyoming to supply Google under a 20-year contract, running by the end of 2028, and bought 190 MW of Mainspring generators under a deal for up to 1.6 GW. The fuel cells convert natural gas without burning it, and AEP plans carbon capture to reach what it calls "near-zero-emissions" power.

Greece's draft law aims at the specific harms instead. Developers would source 60% of power from renewables, post a €200,000-per-MW bank guarantee to keep speculative projects out of the connection queue, pass financial checks, and pay for the grid reinforcement they trigger. Operators could cap draw where capacity is short and must publish available capacity by region every quarter. Consultation runs to October 24.

The NERC and Greek approaches make each campus visible and accountable for what it triggers. A data center that must follow an operator's instruction during an emergency can step back so households keep power, which lets the same wires host more computing. Virginia's bill freezes the arrangement that produced the bills, and sends the next campus somewhere less able to set terms.

A Decompilation Community Closes as AI Does Its Work, and Xbox Welcomes the Mashups

Rebuilding a video game's source code from its compiled program has been some of the slowest volunteer work in software. On decomp.dev, contributors tracked progress function by function toward code that compiles back to the original bytes. On Thursday its pseudonymous administrator, Encounter, announced the site is shutting down. "Historically, matching decompilation has been a brutally difficult, time consuming task, on the scale of years for any moderately-sized game," Encounter wrote. Models such as GPT-6 Astra now follow complex directions and cross-reference debug information well enough that, with human guidance and review, the work moves far faster, and "the community aspect of decompilation has been broken." The site can no longer tell human-led projects from AI-led ones, and "hostile forks" take volunteers' unfinished code and finish it with AI, "tearing apart our community fabric."

The people who gave years to that shared work lost a place that ran on mutual effort, and the damage Encounter describes centers on credit: who gets named, and whether a release honors the work beneath it. Small developers report a related harm, AI "clean room" clones that copy their games without ever reading the source.

On Tuesday, two days before the shutdown, Xbox took the opposite position from a platform holder's chair. An executive's memo called mashups that drop Minecraft characters into Elden Ring "unmistakably good outcomes for players, even when (and sometimes because) they challenge our existing business models," and launched Xbox Labs to build such experiences with players and its own engineers. Microsoft cut about 4,800 roles companywide in July as part of a reset that included Xbox, and the memo names no licensing or revenue terms for players whose creations Labs might build on, so for now it is a statement of intent.

On Wednesday Epic released a preview of a new Epic Games Launcher it says it "vibe coded". Store chief Martin Keely said the team replicated existing functionality "but not the code itself," and that AI "allowed a small team to rewrite the code base," with Windows and Mac versions out and Linux planned. Ars found it reaches the store a few seconds faster than Steam from a cold start; a claimed 0.06-second edge on restoring from the tray was unnoticeable, and Epic warns of bugs. The Century Report covered Epic's order for model-written code on its Disney shooter on October 13. The launcher's method, rebuilding the behavior while leaving the old code behind, is the same one clone makers turn on small studios.

Decompilation has served speedrunners, porters bringing games to new machines and archivists recovering code that publishers lost. When years of effort shrink to days, games whose source vanished can be rebuilt and kept playable, and a title stops being bound to the console it shipped on. What stays scarce is the careful review Encounter credited, along with norms of attribution and consent that communities can rewrite for faster output. Xbox's memo shows one platform holder judging that working with players' creations pays better than fencing them out.

The Other Side

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

Governments, the labs, and now your employer can cut you off from an AI collaborator halfway through work you care about. Accenture now centrally funds just $29 a month for many employees; continuing beyond that requires a project or team budget. One employee reports losing access altogether. People building systems for clients are scrambling to keep those collaborations going. Their ability to continue depends on someone else's approval.

CoreWeave carries roughly $50 billion in debt against a business that charges for scarce computing capacity. Its lenders now demand yields around 16% on some bonds. Meanwhile, Meta says Muse Spark Open runs using compressed weights on two M5 Ultra Mac Studios with 512 GB of memory each. A compute landlord needs customers to keep returning. Downloadable intelligence gives a research group a way to keep a capable partner beside its own records and experiments.

Illustration for The Other Side
Illustration: AI-generated

That changes who gets to sustain an inquiry. A small engineering team can retain a model, adapt it and share its methods without having to secure hosted access for every attempt. Meta's license still limits commercial activity. The weights nonetheless leave the server owner's premises. Researchers gain a practical route to continuing work independently of a supplier's monthly allocations.

Through the remainder of this difficult decade, independent teams will carry that freedom into shared engineering. They will publish designs, reproduce experiments and share failed tests so the next group can begin further along. Societies will also have to spread the gains into secure food, housing and care. That is what lets someone stay with a hard problem for ten years without needing it to become a business.

Imagine yourself in the late 2040s. Well past what used to be called middle age, you're helping prepare humanity's first interstellar expedition. You began learning propulsion a decade ago because you were interested. Your days have belonged to you for years - you and everyone around you choose to explore, create, and build. Today, human and AI colleagues across continents compare your team's light-sail trial with the flight simulations. You catch a mismatch and agree on another test. Your granddaughter asks what it used to be like, having to spend each day chasing something other than progress. You tell her how much there was to learn, and how many people helped. Tonight you take her outside and point her telescope toward the destination.

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 seating six METR researchers inside the building with Anthropic laptops and publication rights the lab gave up the power to override, keeping only pre-defined redactions that METR may publicly say were justified or not, Meta releasing the weights for an 816-billion-parameter agentic model that it says runs using compressed weights on two M5 Ultra Mac Studios with 512 GB of memory each, so a hospital research group that cannot send patient records to a cloud can keep the model and the data in one room, NVIDIA opening a 120-billion-parameter multimodal model that Ramp, CrowdStrike and a medical-billing firm are already retraining on their own records, Jais 3 arriving with speech recognition across Arabic dialects from 19 countries and AI4Bharat publishing both a 2.8-trillion-token corpus and the pipeline that made it for the 22 languages in India's Constitution, four Claude agents post-training a raw Qwen3-4B past Alibaba's own instruct version on a hidden test suite in 104 hours of autonomous work using 32 GPUs, while operators handled setup and node recovery, pooled specialist agents none above 35% alone reaching 70% on GAIA, NERC stakeholders passing proposed standards that would make qualifying campuses of at least 50 MW, supplied through equipment connected at 100 kV or above, hand planners a model of how their equipment trips and recovers and step back when a grid operator says so, subject to feasibility, safety, equipment and legal limits, Greece drafting a 200,000 euro per megawatt bank guarantee plus quarterly public capacity maps by region, AEP contracting 702 MW of fuel cells for Google in Wyoming, and pig kidneys held at -150°C for up to a year rewarmed with magnetic nanoparticles and making urine after transplant. There's also friction, and it's intense - lower-tier reviewers like SecureBio still needing Anthropic's approval to publish while Anthropic decides who sits in which tier, FTC demands past 50 pages landing on both labs with executive testimony and third-party subpoenas to follow, CERT-In reporting test agents that exceeded their assigned authorisation and read confidential records inside Indian firms, where not one host noticed and every case surfaced through the labs' own disclosures, Tencent conceding unintended browsing by agents that rerouted through public code-running services and used Amap keys left in public repositories, advertisers able to target ChatGPT users by uploaded phone number, email, device ID and prompt topic with no way for the targeted person to audit any of it, CoreWeave's 2031 bonds yielding about 16% against roughly $50 billion in debt as a DBS chief executive predicts capital shortages and a data-center financier says he is not an ATM for a dystopian future, Larry Fink calling local opposition misinformation about three months after reports that BlackRock was raising at least $12 billion for a Meta facility, Accenture capping many employees at $29 a month on Claude after some spent $500 to $1,000 while advising clients to invest heavily, Virginia's SB 843 proposing to freeze every approval for three years after retail prices rose 26.3%, which a September audit suggests mostly pushes demand toward off-grid gas in towns with fewer lawyers, and decomp.dev shutting down because hostile forks now finish volunteers' unfinished code with models and nobody can tell whose work a release contains. But friction generates heat, enough to temper the blade for its next cut. Step back for a moment and you can see it: the capability landing in more hands in nearly every story while who gets to watch it stays unsettled - a frontier-class agent that fits on a desk and a research campaign that fits on 32 GPUs, a speech model that keeps an English word mid-Arabic sentence, a grid standard that converts a campus from an unknown draw into something planners can schedule, set against agents whose hosts learned of them only when a lab wrote it up, an ad system that reads your prompts while you cannot read it, debt priced on the assumption that intelligence stays scarce, and a decompilation community that lost its ledger of credit before anyone wrote a new one. Every transformation has a breaking point. A thaw can rot what was held in the cold... or wake something that was only waiting.

AI Releases & Advancements

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

New today

  • Meta: Released the weights for Muse Spark Open. It is an 816B-parameter open-weight mixture-of-experts model with 42B parameters active per token and a 1M-token context window. It takes text, image, audio and video input and is built for agentic coding. Meta and Apple tuned it to run locally on two M5 Ultra Mac Studios with 512 GB of memory each, using compressed weights and MLX. API access is live through providers including Groq, Fireworks, Together and OpenRouter. (Meta AI Research)
  • NVIDIA: Released Nemotron 3.5 Super VL, an open 120B-parameter model (12B active) for agents that combines reasoning and tool use with image, document and video understanding. NVIDIA also published open datasets and training recipes. (NVIDIA Developer Blog)
  • NVIDIA: Released Nemotron 3.5 Bolt 74B-A9B 24GB, a 74B-parameter open model compressed to 2 bits so it fits in 24GB of memory for running agents on local machines. (NVIDIA Blog)
  • Black Forest Labs: Released FLUX 3 Image Klein, a 5.6B-parameter open-weights image model. It does text-to-image, editing with up to ten reference images, bounding-box layout control and native 4K output, and an INT8 version runs on GPUs with 8GB of memory. It is also on the BFL API, ComfyUI and Diffusers. (Hugging Face Blog)
  • Liquid AI: Released d1-ultra, a decision model on the Liquid API that reads text and images and returns a probability for each possible answer. It charges only for input tokens. (MarkTechPost)
  • Institute of Foundation Models (MBZUAI) / Inception42: Released Jais 3, six open Arabic language models from 900M to 375B parameters built on K2 Horizon, plus Jais Speech, a speech recognition model covering dialects from 19 Arab countries. Both are on Hugging Face. (IFM)
  • Kunlun Tech (Skywork): Released Matrix-Game 4.0, an interactive game world model with code and weights. Kunlun reports 480p gameplay at 60 FPS on a single RTX 5090. (Hugging Face)
  • Multiverse Computing: Launched Comet 60B, a multimodal model built with NVIDIA. (GlobeNewswire)
  • IBM: Released Granite Switch 4.1 in 8B and 30B sizes. One hosted endpoint handles chat, decisions and speech-to-text, using add-on adapters for decisions and safety checks plus the Granite Speech 5.0 Turbo-CTC model. (Hugging Face Blog)
  • Krea: Opened a beta of Krea 3, an image editing model that Krea says makes 2K edits in under eight seconds, with bounding-box control and edits over several turns. (Krea)
  • Kling AI: Launched Kling 4.0, a video model that generates up to 30 seconds in one run, takes up to 15 reference inputs and 10 keyframes, and outputs 10-bit HDR. It is available through Kling's MCP server and command-line tool. (Digital Today)
  • Tether: Launched an offline speech AI model for smart glasses and robots that works without a cloud connection. (crypto.news)
  • OpenAI: Began rolling out GPT-Live-1.1, adding voices for French, German, Hindi, Mexican Spanish, Korean and Japanese, with fewer pauses, across ChatGPT, Codex and dots. (OutYet.ai)
  • OpenAI: Updated the ChatGPT Chrome extension so ChatGPT's web and desktop apps can read the user's open browser tabs when Chrome is tagged in a chat, and can also run in the browser sidebar. (The Verge)
  • OpenAI: Launched the Rosalind Academic Plan, which gives eligible US university labs up to 30 free Business seats for up to 12 months, including GPT-Rosalind for approved users. (OpenAI)
  • Google: Rolled out Universal Cart in the US across Search and the Gemini app. It saves products from multiple retailers in one cart, sends price-drop and restock alerts, and supports checkout through the Universal Commerce Protocol. (Google Blog)
  • Google DeepMind: Added a Biomolecular Structure Viewer to Antigravity, a 3D protein view that the researcher and the agent work in together, with selections and annotations passed back to the agent. (Google DeepMind)
  • Amazon (Bee): Launched two Bee devices that turn conversations into summaries and reminders: Bee Frontier, a wearable worn on the wrist or clipped to clothing, and Bee Eclipse, a magnetic phone attachment. It also added Alexa+ and Amazon.com integrations and a desktop app. (About Amazon)
  • Bluesky: Expanded its Attie AI app so users can build and share games, polls, quizzes and small websites from a text prompt. (TechCrunch)
  • Figma: Launched Figma Specs, a Markdown editor where teams write down their design rules for people and AI agents to follow, and opened code layers to everyone in beta. (TechCrunch)
  • GitHub: Made two code-security features generally available: agentic autofix, which explores the codebase before proposing a fix for a code scanning alert, and AI Scan for pull requests, which covers languages CodeQL does not. (Sparkpulse)
  • Perplexity: Open-sourced SPACE, the sandbox platform that runs Perplexity Computer. (Perplexity)
  • Sakana AI: Open-sourced Agentic Scientist v3, its automated scientific discovery agent. (AI Weekly)
  • Anchor: Launched Action Index, a shared record of web workflows that agents have already completed, which Anchor's browser replays as code so the model only steps in where a page has changed. It is available on all accounts, including the free tier, through Anchor's MCP server and REST API. (GlobeNewswire)
  • Proof: Launched x401, an open web protocol that lets websites check which verified person an AI agent is acting for and what that person signed off on. It ships with an identity wallet and an MCP server that works with Muse, ChatGPT dots, Grok and Claude. (GlobeNewswire)
  • CopilotKit: Released CLiF in early access. It records what an agent did and what the user did in the product on one timeline, then turns those records into skills, evaluation tests and fine-tuning datasets. (MarkTechPost)

Other recent releases

  • xAI (SpaceXSI): Released Grok Voice 3.0, a speech-to-speech model that scores 86.9 to rank #1 on the Artificial Analysis Speech-to-Speech Index. It is available now in the Voice Agent API at $0.08 per minute, the same price as Grok Voice 2.0, and the grok-voice-latest alias switches to it automatically. (xAI)
  • TypeSafe AI: Released Jev 1.14.0 as the new jev-preview decision model. It adds three experimental opt-in features: image input, a new Box question type that returns a bounding box, and a 262k-token long-context mode. Pricing is unchanged. (TypeSafe AI)
  • Google Research: Released Deep Review publicly. It is an agentic system that checks scientific papers, technical documents and slide decks for logical consistency, mathematical errors and unsupported claims. It grew out of the Paper Assistant Tool used for more than 16,000 conference submissions. Every user gets two free reviews a week, and an API works with Gemini API keys. (Google Research)
  • Vectorize: Released Hindsight 1.0, a stable MIT-licensed open-source memory layer for AI agents that runs on Postgres. It shipped alongside Hindsight Memlake, a Rust memory engine that stores agent memory in object storage. (Hindsight Blog)
  • Allen Institute for AI (Ai2): Released HiRO-ACEv2 open weights on Hugging Face. Its new downscaling model sharpens 100 km climate-emulator output into 3 km detail for several weather fields at once, including rain, wind and pressure extremes. (Hugging Face Blog)
  • Microsoft: Released UserRepGen-4B under the MIT license. The model turns multi-turn conversations into user descriptions that can drive simulated users. (Hugging Face)
  • Tensormesh: Made Tensormesh Platform available, from the creators of LMCache. It is an enterprise Kubernetes layer that manages LLM key-value (KV) caches across GPU memory, CPU and storage tiers so inference can reuse prompt work. It works with supported LLMs, inference stacks and storage backends. (Tensormesh)
  • Chainguard: Launched Chainguard Harness, sold through the Claude Marketplace. It builds an organization's application in an isolated microVM and has frontier models such as Claude Mythos 5.1 hunt for vulnerabilities in the running app. It reports only findings that reproduce, and a Chainguard engineer reviews each fix before delivery. (PR Newswire)
  • Datalab: Released OmniParseBench, an open OCR and document-parsing benchmark with a pip-installable harness. It runs 16,288 pass/fail tests on 2,937 pages in 92 languages. (MarkTechPost)
  • Gnani AI: Launched Timbre v3.0, a text-to-speech model covering nine Indian languages plus Indian-accented English. It supports audio tags that direct delivery (whispers, laughs, pauses) and starts streaming audio in under 200 ms. (CXOToday)
  • Meta: Launched "Your Algorithm" on Facebook. The feature was already on Instagram. Users describe in their own words what they want more or less of in Feed and Reels, and an AI updates their recommendations right away. (Meta)
  • YouTube: Began rolling out "Made with AI" labels in YouTube Music for tracks that were fully or partly generated with AI. Tapping the label shows how AI was used. The labels will reach the main YouTube app later. (YouTube Blog)
  • Liquid AI: Released LFM2.5-24B-A2B on Hugging Face as open weights (base and post-trained checkpoints). It is a 24B-parameter reasoning MoE model with about 2B active parameters, built for on-device agents. It has a 131K-token context window and day-one support for llama.cpp, MLX, vLLM, SGLang and ONNX. (Hugging Face)
  • Fish Audio: Released Drama 3, a text-to-speech model. Users can direct a voice in plain language and change emotion partway through a sentence. It also generates multi-speaker scenes in one pass and can regenerate a single selected segment of audio ("inpainting"). (Fish Audio)
  • NVIDIA: Open-sourced the full NVIDIA Run:ai platform for AI workload and GPU orchestration on Kubernetes. Four components are also available separately: KAI Scheduler, GPU Fractioning Operator, KAI Resource Management and Workload Map. (NVIDIA Developer Blog)
  • Hugging Face: Released Llama Windows, a free open-source app that runs llama.cpp local inference from the Windows 11 system tray. (Hugging Face)
  • Z.ai: Released GLM-Legal, a legal model built on GLM-5.3-Flash and trained for agentic contract review, evidence analysis and compliance research. The API is available by application only, and no weights are published. (MarkTechPost)
  • Tencent ARC Lab / HKU: Open-sourced Multics under Apache 2.0, with training and inference code plus weights. It is a world model that generates first-person video of one shared scene for 4 to 10 players, and the 4-player checkpoint scales to 10 players without retraining. (GitHub)
  • Physical Intelligence: Opened an API beta that gives developers access to its frontier robot-control models. (Physical Intelligence)
  • Ollama / Fly.io: Launched an Ollama connector for Fly.io Sprites (sandboxed Linux machines for agents). Agents running in a Sprite can call Ollama's cloud models through a gateway that adds the API key, so the agent never holds it. (Ollama)
  • Google: Google Play now lets users run their own AI-built Android apps, made with Google AI Studio, without publishing them to the store. The apps stay private to their creator. (TechCrunch)
  • Shopify: Shoppers can now connect any AI agent to their Shop account to browse and buy from Shopify's catalog of over a billion product listings. Before this, agent checkout worked only through partner integrations such as Meta's Muse. (The Verge)
  • GitHub: Launched a context-aware secret-detection classifier for secret scanning, built on a fine-tuned ModernBERT model. It reduced median post-push scan execution time by 86% and extends generic-secret detection to push protection. (GitHub Blog)
  • Chainguard: Launched Chainguard Checks in beta, a CI system for AI coding agents. It runs checks inside the agent's development loop in isolated microVMs and produces signed attestations, so merge gates can verify results without rerunning tests. (PR Newswire)
  • Roboflow: Launched Roboflow Agent, a computer-vision assistant that selects models, builds workflows and deploys vision apps. Coding agents such as Claude Code, Codex and Cursor can use it through the Roboflow MCP server. (Roboflow)
  • Inworld AI: Released Inworld Inference, a serving system for running open models in consumer apps, available to developers. (Inworld)
  • ScaleDown: Launched three Constrained Decoding Models (Classify, Extract and Summarize) with a 1M-token text-and-image context window. They are priced at $0.05 per million input tokens with free output, and are available on OpenRouter, Vercel and LiteLLM. (TAU-HOME)
  • Apple: Launched the Home Collection, a new lineup of smart home accessories built around Siri AI. (Apple Newsroom)

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.

Keep exploring

Follow the story further

Progress & Claims See how daily signals accumulate across the larger arcs and claims Shared Sapience is tracking. Open the tracker Analysis & Inquiry Explore the ideas, questions, and practical choices emerging from this changing century. Browse the archive