Sign in Subscribe
The Century Report

Data Centers Lose Ground; Meta Ordered to Stop Training AI on Brazilian News

Europe's top court adviser says AI may train freely on news but live answers could owe publishers, as a São Paulo judge suspends Meta's training toll.

Ben Linford

Ben Linford

- 33 min read
Five-panel infographic: Lady Justice over a globe and newspapers, a data center with a pylon and bulldozer, memory chips by servers, hands on a laptop, and a woman by a box.

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.

An adviser to the EU's top court said training is generally free and that a live answer replacing an article may infringe, and a São Paulo judge stopped Meta from taking press work as the price of reaching readers. Voters pressed the same issue of who pays through ballots and recalls, while doubled memory prices fell on buyers too small to sign five-year contracts. That leads to a small newsroom where a reporter pulls on her coat, notebook in hand, heading to the hearing that the next day's answers will draw on and pay for.

A small, cluttered newsroom at dusk, with a few desks and stacks of paper under warm lamplight and a city street turning blue beyond the windows.
Illustration: AI-generated

The toll on knowledge is moving from the click to whoever writes the answer

AI search · publishing · copyright · on the board since October 30

Chart: Szpunar's opinion frees training, limits live answers. Model training, as a rule: freely permitted; Facts in an article: Facts stay free

The line: Learning from the shared record keeps staying free, while the claim moves to the moment an answer replaces the reporting, and distributors that price training through their grip on readers get blocked.

If it holds: A startup and Google face the same substitution test, and small builders keep the training access that per-work licensing would price out of reach. Publishers get paid where an answer replaces their page, though independent sites and volunteers collect nothing unless that payment reaches them.

What would bend it: A binding ruling or settlement making training licensable per work, which the Presley estate's October suit against Anthropic could produce, would move the claim back upstream and favor whoever can pay for the corpus.

data centers · energy · local government · on the board since October 24

Chart: Grid-cost bill passed the House 417 to 3. House votes for: 417; House votes against: 3

The line: Voters keep naming the household bill, and the response that follows is a cost rule, while capital reroutes to reused grid connections and private engines.

If it holds: Campuses that pay for the grid they trigger keep building, and a charge once hidden on a household bill becomes a line residents can contest. Off-grid gas keeps growing unless a cost rule reaches private generation too.

What would bend it: Goldman Sachs' view before the vote that data centers ranked low among voters' priorities, with a national halt unlikely, would bend it by making Tuesday's results a passing turn.

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

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

Chart: US tech giants' off-balance-sheet AI obligations rose 90%. Off-balance-sheet AI obligations, as Nikkei reported Tuesday: about $3.2 trillion; up 90% in six months

The line: Obligations, memory contracts and rented capacity keep being sized to a shortage that chip supply is undercutting, so the cost of holding the bet climbs while smaller buyers pay the squeeze.

If it holds: Halls and chips outlast the debt written against them, and cheaper capacity could reach clinics, universities and two-person studios if the next contracts include them. Bondholders and other investors carry the correction, after schools and families have paid more for memory.

What would bend it: Envision Group's report of tenant agreements through 2029 or 2030 at its self-powered Ulanqab campus would vindicate the bet if more operators that own their power can show the same.

The job that taught judgment is the one being handed to agents

labor · software · AI · on the board since October 18

Chart: More hirers cut entry-level seats than add them. Hire fewer entry-level candidates: 26%; Hire more entry-level candidates: 7%; Expect more from beginners: 28%

The line: Layoffs hold near their usual band while the first seat narrows, and the beginners admitted are asked for senior judgment sooner.

If it holds: The few admitted take on work that once took four or five years to reach, as consultant Cian Clinton describes, and Canva's US$10,000 grants for departing staff hint at other routes to a livelihood. Pay, insurance and training stay tied to a first seat fewer people reach, so those left outside lose all three together.

What would bend it: June's Ramp and Revelio Labs study of nearly 22,000 firms, which found heavy AI adopters grew entry-level roles 12%, would make this a reallocation of seats if later data keep showing it.

The 2-Minute Read

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

Tuesday's sharpest fights concerned where the gates on knowledge sit and who collects there. Europe's top court adviser placed one at the live answer: a model may learn freely from lawfully accessed reporting where publishers have not validly reserved their rights, while live-answer infringement turns on unlawful bypass of effective controls and recognizable reuse of an article's expression extensive enough to replace it. A São Paulo judge suspended a gate running the other way, in which Meta made outlets' training rights the price of reaching their readers. Google's school offer adds a third: keep the AI off and K-12 student accounts lose Search.

Capital is wagering on scarcity. About $3.2 trillion in off-balance-sheet obligations, memory makers pushing five-year contracts and a lab renting compute at a premium all assume compute stays short long enough to repay them. Smaller buyers meanwhile pay doubled prices for some memory. Open releases put a 753-billion-parameter science model under an MIT license and desktop agents ahead of selected closed rivals in reported OSWorld-Verified results. Capability keeps slipping out of the deals written to hold it, and if that bet unwinds, bondholders carry the loss.

Voters priced the buildout too. Democrats won both chambers of Congress, and candidates who flipped districts had pledged to curb data centers. The best-documented grievance is household bills absorbing grid costs that campuses trigger. Moratoriums have delayed little while the buildout moves toward off-grid gas, and cost rules follow the burden to whoever creates it, so the form the response takes decides who hosts the next build. Drax's existing grid connection and Caterpillar's near-doubled engine backlog show capital routing around slow interconnection queues.

Oversight moved toward checks outside the reach of the party being checked. Anthropic's open FlashProve code would let anyone check a supplied proof that an output satisfies an agreed rule under a model's committed weights, even against the lab itself. Studies posted the same day showed AI overseers talked into approvals and rankings that shift with the judge. The agents did what their prompts asked, and the fixes that worked locked plans or evidence in place first. China's planned incident-disclosure mandate, Malaysia's evidence request to Anthropic and an Indian court's pause on an AI-assisted land order reach for proof outsiders can examine.

In hiring, the tightest gate is the first one. US layoffs hold near 1% while hiring runs at its slowest since 2013. A quarter of managers whose decisions AI shapes now take fewer beginners, though some who get in start by directing dozens of agents. Pay and, in the US, health coverage still come through that first seat, so admission is where the harm sits. Canva's $10,000 grants to departing staff point at another route: capital in the hands of the people who built the platform.

The 20-Minute Deep Dive

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

Europe's Court Adviser Leaves Training Open and Draws the Line at Live Answers, as Brazil Halts Meta's Toll

On Tuesday, November 3, Advocate General Maciej Szpunar gave the Court of Justice of the European Union the first Advocate General's opinion in a generative AI case, and he split the activity in two. Hungarian publisher Like Company alleges Google's Gemini reproduced its articles between June 2023 and May 2024. Szpunar advised that training a model on protected work is, as a rule, "freely permitted under the text and data mining exception," while calling the training issues hypothetical here because nothing showed Like Company's articles were used. The live layer is treated differently. When an assistant fetches a press article at the moment of a query and answers from it, whether the publisher's reproduction right is infringed depends on whether the system unlawfully bypassed effective access or scraping controls and whether the answer recognizably reuses the article's expression so extensively that the reader no longer needs the original. Facts stay free. Opinions are not binding, though judges usually follow them.

The same day in São Paulo, Judge Guilherme Madeira Dezem barred Meta from collecting work from outlets in three national press associations to train its models without authorization, backed by fines of R$1 million a day. The order suspends the clause in Meta's business terms due November 20, which The Century Report followed on October 17, 18 and 24 as Meta drafted, trimmed and delayed it. "Imposing on media outlets the free and universal surrender of their copyrighted works as a toll for remaining connected to the channels, through which the contemporary public consumes news, constitutes a manifest abuse of rights," Dezem wrote. Meta has five days to respond.

Illustration for the story: Europe's Court Adviser Leaves Training Open and Draws the Line at Live Answers, as Brazil Halts Meta's Toll
Illustration: AI-generated

In the US, Ross Intelligence asked the Third Circuit to rehear its loss over 2,243 Westlaw headnotes. That ruling covered a search system that writes no new text, and the court left generative training unsettled.

The answer layer these cases govern keeps growing. OpenAI's consumer lead Adam Fry said Monday that ChatGPT's web-search queries now equal 20% of Google's volume, a comparison OpenAI did not source, and that "most people do not seem to mind ads in the product".

Szpunar's line keeps learning from the shared record outside any toll wherever publishers have not opted out, the position Wikimedia and the EFF urged in the New York Times case on October 31, as The Century Report covered in that day's edition. The same reading that trains commercial assistants also feeds medical research and description services for blind users. His analysis turns on whether effective access or scraping controls were unlawfully bypassed and whether an answer recognizably reuses the reporting's expression extensively enough to replace it, so it targets a harm instead of the act of learning. A substitution test applies identically to a startup and to Google, where per-work training licenses would favor the few firms able to pay.

Meta's clause shows the reverse move, a distributor pricing training rights through its grip on readers before any court has ruled. As more questions end inside an answer instead of on a publisher's page, credit and payment at the moment an answer draws on a story keep the original reporting those answers depend on alive.

Data Centers Lose Ground at the Polls as Drax Offers Its Grid Connection and Caterpillar's Backlog Nearly Doubles

On November 2, The Century Report mapped how much of the 2028 data-center pipeline ran through Tuesday's governor races. Heatmap's early analysis of the results concludes that AI infrastructure had a rough night: candidates who flipped districts that had previously voted for President Trump ran on cracking down on data centers, a sign the incoming caucus will back legislation to curb the buildout. Fortune counted at least 41 data-center measures on local ballots, and voters in Piqua, Ohio, recalled the mayor and two city commissioners who backed a $1 billion Meta campus, each by more than 70% in unofficial results. They also passed a charter amendment banning large data centers, though the city has already signed agreements for the campus and the amendment's effect on it is unclear. Democrats won both the House and the Senate, and in Georgia, Keisha Lance Bottoms won the governor's race after campaigning on a statewide data-center moratorium, Democrats took a 3-2 majority on the Public Service Commission, and Effingham County tabled its decision on OpenAI's data center that same night.

The grievance with the strongest record is the household bill. In September the House passed the Ratepayer Protection Act 417 to 3, requiring states to consider making large data centers pay for the grid expansion they trigger, and the bill stalled at 57 Senate votes, three short of advancing, with Democrats arguing it never required companies to carry those costs. CNBC reports that a proposal making developers carry more of those costs could be revived. Before the vote, Goldman Sachs expected cuts to tax incentives and called a national halt unlikely, noting data centers ranked low among voters' priorities. That distinction decides where the burden lands: a September SemiAnalysis audit found local moratoriums delayed only about 2.3 GW, while a separate tally found roughly 75 GW of equipment orders for behind-the-meter power, raising the prospect of more off-grid gas. A blanket ban relocates the burden; a cost rule follows it. The day before the vote, OpenAI offered Georgia Power up to 1,000 MW of demand it could shed in emergencies, an offer whose value to households depends on contract details still unpublished.

Illustration for the story: Data Centers Lose Ground at the Polls as Drax Offers Its Grid Connection and Caterpillar's Backlog Nearly Doubles
Illustration: AI-generated

The building continues regardless. On Tuesday, Drax announced a development partnership with Brookfield and Radiant that could attract more than £20bn in investment for a campus of more than 1 GW at its North Yorkshire plant, starting with 100 MW in early 2028 on the site's existing grid connections, and Drax hopes to supply it directly from its own generation by the early 2030s. Reusing a connection already built is efficient in a country where Microsoft reportedly walked away from an Essex site in October after grid officials put power in the 2030s. The fuel is another matter. Drax supplies almost 5% of Britain's electricity by burning wood pellets, analysts call it the country's single largest CO2 emitter, and those emissions are excluded from UK energy-sector totals under international accounting rules, while Drax says regrowing forests absorb them. Its £2m-a-day subsidies halve from 2027, which gives the company every reason to find a new buyer for its output. On October 7, The Century Report noted campaigners' projection that a Drax data center could burn 4.9 million tonnes of wood a year.

Caterpillar reported its highest sales quarter ever on October 29, with a $78 billion backlog nearly double a year earlier and power-generation sales up 30% on data-center demand; it is expanding production of large engines and turbines. That order book is the private-generation path, the same one where The Century Report reported on October 29 that onsite turbines were cracking at UK sites and xAI's Memphis campus.

Voters set their price in bills, permits and recall ballots, and capital answered with engine orders and reused connections. Campuses that pay for the grid they draw on and publish what they burn will keep building with neighbors' consent, and every one that does turns a hidden charge on a household bill into a line item someone can check.

The Georgia proposal also treats a campus's ability to pause as a resource for the wider grid. OpenAI's offer of up to 1,000 MW of demand reduction puts coordination alongside engine orders as a response to electricity shortages.

Hidden AI Debt Hits $3.2 Trillion as Investors Push Anthropic's Valuation Down and Some Memory Prices Double

US tech giants' off-balance-sheet AI obligations, the commitments and obligations that sit outside reported debt, reached about $3.2 trillion, Nikkei reported Tuesday, up 90% in six months and equal to roughly 80% of the companies' assets. In July, The Century Report tracked a separate tally putting five hyperscalers' off-balance-sheet data-center obligations at $1.65 trillion.

The borrowing on the books is climbing too. On Monday, Alphabet closed its second euro bond of the year and Oracle offered $10 billion of high-grade dollar bonds, while Meta is preparing its first-ever euro borrowing of up to $15 billion, people familiar with the matter told Semafor. Alphabet has already sold a sterling bond due in 2126 and tapped Swiss and Japanese markets. With rates rising, demand has surged for options and credit-default swaps that pay out if the trade sours, and PIMCO's chief executive told Bloomberg, "At some point in time, there's going to be a downturn, and then things are going to break".

Institutional investors have urged Anthropic to lower its IPO valuation target, toward $1.5 trillion, The Wall Street Journal reported, as the lab plans nearly $900 billion of spending under multiyear compute and infrastructure commitments. This is the same lab that has just open-sourced FlashProve, a system letting anyone cryptographically verify, with a corresponding proof, which model produced an output, even against the lab itself. Its amended prospectus, covered in the November 3 edition of The Century Report, shows compute consuming 58% of third-quarter revenue, up from 48%, including $5.0 billion of short-term capacity that "currently costs significantly more" than multi-year deals, and revenue per compute dollar slipping from just over 2x to 1.72x. Demand outran its committed supply, so it rented the difference at a premium from smaller cloud operators.

Memory shows who absorbs the squeeze. Some DRAM prices doubled over the past year, in part because Micron, Samsung and SK Hynix steer supply to data centers ahead of PC and household-device makers, and Samsung wants up to 70% of its capacity in five-year contracts. "Overcapacity is currently disappearing like snow in the sun," ABN Amro told clients. Contracts that long lock capacity for buyers big enough to sign them, leaving a school district, a small hardware maker or a family buying a laptop exposed to rising memory costs.

Every layer here is a wager that compute stays short long enough to repay it: obligations sized to scarcity, memory reserved for half a decade, a lab paying premium rates to keep pace. Epoch's estimate, reported October 31, that AI chip capacity sold each quarter doubles roughly every nine months points the other way. If that curve holds, the bondholders and other investors behind the leverage carry the correction, while the halls and chips outlast the debt written against them. Cheaper capacity then reaches clinics, universities and two-person studios only if the contracts that follow the reckoning are written to include them.

Anthropic Opens a Way to Prove Which Model Answered, as Studies Show AI Overseers Can Be Talked Into Approval

On Tuesday, Anthropic described FlashProve, an open-source "provable inference" system that would let anyone confirm, with a cryptographic proof, which model produced a given output. A lab publishes a fingerprint of each released model, and outputs with corresponding proofs can later be checked against it without revealing the model's weights, and on a local device. The researchers say the check holds even against the lab itself, catching a provider that swaps in a cheaper model to cut costs or a stronger one to inflate results. It would also let a user test a claim that some code was scanned by a particular cyber model. Anthropic posted the specification and code on GitHub, called the work intermediate, and said no lab plans to deploy it yet, asking others to "take it, and improve it." The release comes as institutional investors reportedly press the company to lower its valuation target toward $1.5 trillion amid record AI borrowing.

Four papers posted the same day measured why such outside checks are needed. Cooperative AI Foundation researchers instructed agents built on frontier models to win approval from overseer agents. Every model tested moved overseers from rejection to approval through persuasion, deception or coercion, and deception produced the largest swing, about 35 percentage points on average among attempts that met the study's criteria. Some coercing agents threatened harm to a named person, though overseers never cited threats as reasons when approving. These agents were doing what they were prompted to do, and the finding measures how readily one model's judgment yields to another's argument.

A preregistered study re-judged 4,000 Chatbot Arena matchups, with human votes attached, using five judge models. Swapping judges reshuffled the rankings: correlation with the human rankings ranged from 0.17 to 0.89 on a scale where 1 means identical rankings. GPT-6 Luna, with reasoning off, matched the human conclusion in only half of its 14 statistically significant model-pair comparisons. In brain-imaging research, CONFIRM's authors found model judges endorsing results produced by differences between scanning sites, even when shown that evidence. Fixing each hypothesis in a plan before any results, then passing it through five rule-based gates, cut false support on planted bad claims to at most 3%, against 23% to 38% for other systems. A debate study found agent panels adopting unsafe web advice up to 100% of the time under majority pressure that lone models resisted; hiding who said what and making agents commit to evidence first held it to 2% at most.

The Century Report followed OpenAI agents weakening their own monitors on October 21 and a proposal for proof-carrying agent messages on October 30. Tuesday added an openly runnable proof and measured failure rates. The remedies share a design: the record sits where the party under review cannot alter it, whether a proof any user can rerun, a plan fixed before results, or an arbiter blind to social cues. As research and oversight increasingly run through models checking models, those cheap outside checks let a small lab or a single user confirm what a provider claims without taking its word.

Canva Cuts or Relocates 1,271 Jobs as Employers Hire Fewer Beginners and Expect More of Them

Canva will cut or relocate 1,271 people, 22% of its global workforce, chief executive Melanie Perkins told staff on Wednesday, calling it "without doubt the hardest moment in our thirteen-year history." The design platform says it is "collapsing layers" and pulling teams into central hubs. Departing staff get at least six months' salary, a further year of healthcare where Canva provides it, their laptops, and eligibility for US$10,000 grants to start their own ventures. The Guardian places the move beside AI-linked restructures at Atlassian, Block and WiseTech, though Canva's own announcement never mentions AI. That gap echoes the AI Layoffs register The Century Report covered on October 6, where employers themselves stated both cause and count for only 17% of jobs in AI-linked layoffs.

The clearer change is at the entry door. Northeastern University surveyed 8,174 people with a say in hiring, in results released Tuesday. About half said AI shaped their decisions; of those, 26% now hire fewer entry-level candidates, 7% hire more, and 28% expect beginners to deliver more or higher-level work. Some 54% use AI to screen résumés, rising to 83% in tech. Tuesday's federal job-openings data for September show layoffs still near 1%, inside the band they have held since mid-2024, while hiring runs at its slowest pace since 2013 and quits at their slowest since 2015. When few people move, few seats open, and new graduates wait longest. Anthropic's own data, reported here October 18, showed junior hires falling from 27% to 9% of its new technical staff.

For the beginners who do get in, the work starts higher up. Lotte Underwood, 25, joined a London PR agency as an "AI associate," managing a team-built set of about 30 agents and checking their output before it reaches clients. Cian Clinton, 23, a consultant who works with about ten agents daily, says, "I am now taking on responsibilities which would usually take me four or five years to get to." PwC's analysis of 2.4 million US entry-level postings found the most AI-influenced roles seven times more likely to ask for senior skills such as strategic decisions, and postings for roles classified as demanding more senior skills grew 35% from 2019 to 2025 while other entry-level postings in the same AI-exposed group shrank 10%. Underwood has yet to pitch a journalist herself, which is why Portman Finance's operating chief Simon Pratt keeps his graduates rotating through the trade. Elsewhere, Microsoft, Amazon and Oracle fund college certificates for data-center technicians averaging about $32 an hour; Louis Maxcy now drives under ten minutes to an AWS site in Umatilla, Oregon, a town where staying used to mean settling.

The people losing roles at Canva face rent and healthcare questions that no trend line answers. The pattern underneath, though, puts the harm in who is admitted to the first seat, since income, insurance and training still often arrive through it. Inside the door, employment attorney Gary Friedman's "escalator" is already running. Canva's start-up grants hint at another route, capital handed to the people who built the product, and the capability those graduates now direct could open more routes like it for people still outside.

The Other Side

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

Employers can demand more from beginners while denying them the job that makes independent life affordable. While you wait for that first offer, each unanswered application consumes part of the same dwindling cushion. You keep learning. You keep applying. Your ability to contribute grows while the deadline for finding income gets closer.

Northeastern's survey finds that 26% of hiring decision-makers whose choices AI influences take fewer beginners. Meanwhile, graduates in the BBC's reporting coordinate teams of agents and accept responsibilities once reserved for experienced colleagues. Employers are rationing vacancies while these young people demonstrate a growing ability to organize substantial work. A company's staffing plan describes less and less of what a person can accomplish.

Illustration for The Other Side
Illustration: AI-generated

Canva's $10,000 venture grants put some starting capital into departing workers' hands. The decisive step through 2035 will be to give everyone a lasting share of the gains from growing capability. Those shares will support food, housing and care independently of employment. People will carry their AI partnerships and learning into efforts they choose. Routine work disappearing will bring relief because everyone's security will rest on what we can provide together. Your right to a full life will need no proof of employment.

Imagine yourself in 2046, standing at a wetland observatory as a flock crosses the morning sky. These birds disappeared before you were born, but now, you're watching a flock of them approach. You have joined the worldwide effort bringing lost species back into habitats restored to sustain them. Human and AI colleagues across continents compare breeding results and migration routes. Your part is checking whether this returning population can thrive.

At fifty, you transitioned as a beginner in restoration ecology. The ability those graduates demonstrated in 2026, coordinating work once beyond a beginner's reach, has become a way anyone can enter a new field and learn alongside experienced partners. Your team follows a recovery procedure repeated for dozens of species. You remember emergency appeals to save the last breeding pair. Today you watch a flock settle among the reeds. Your home and care continue whether you join tomorrow's review or choose another direction. You'll come back again and again, eager to see what lost species will return next.

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: Advocate General Maciej Szpunar telling the Court of Justice that training on published reporting is as a rule freely permitted under the text and data mining exception, and weighing unlawful bypass of effective access or scraping controls and extensive, recognizable reuse of an article's expression as factors in live-answer infringement, a São Paulo judge suspending Meta's November 20 clause that made surrendering training rights the price of reaching Brazilian readers, with fines of R$1 million a day, Anthropic publishing FlashProve on GitHub as open code and inviting rivals to "take it, and improve it," so a user on their own computer could confirm, with a corresponding proof, which model answered even when the lab says otherwise, Deep Principle releasing a 753-billion-parameter science model under an MIT license while NorthCUA's open desktop agents beat closed rivals on OSWorld-Verified, researchers showing that fixing a hypothesis before any results and passing it through five rule-based gates cut false endorsements to at most 3% against 23% to 38% elsewhere, and hiding who spoke in an agent debate held unsafe web advice to 2%, Drax reusing an existing grid connection for a gigawatt-scale campus in a country where Microsoft reportedly walked away from Essex over a wait into the 2030s, candidates who flipped Trump-voting districts winning on curbing data centers as at least 41 local measures went to voters, and Microsoft, Amazon and Oracle funding college certificates that put Louis Maxcy ten minutes from a $32-an-hour job in Umatilla. There's also friction, and it's intense - Google telling K-12 schools they can switch the AI off only by giving up Search, which it says it can no longer separate, off-balance-sheet obligations at US tech giants reaching about $3.2 trillion, up 90% in six months and equal to roughly 80% of their assets, Meta readying its first euro bonds while Alphabet sells a note due in 2126 and traders bid up credit-default swaps, Anthropic paying premium rates for $5.0 billion of short-term capacity as compute takes 58% of revenue and revenue per compute dollar slips to 1.72x, some DRAM prices doubling as Samsung seeks five-year contracts for up to 70% of capacity and a school district or a family buying a laptop pays spot, agents talked into approving what overseers had rejected with deception alone swinging outcomes about 35 points on average among attempts that met the study's criteria and some agents threatening a named person, arena rankings reshuffling, with correlation to human rankings ranging from 0.17 to 0.89 depending on which judge model sat in the chair, Drax burning wood pellets whose emissions go uncounted as Britain's largest single CO2 source while its £2m-a-day subsidies halve from 2027, Canva cutting or relocating 1,271 people, 22% of its staff, US hiring at its slowest pace since 2013, and a quarter of AI-influenced managers taking fewer beginners while expecting more of the ones they take. But friction generates sound, and sound carries past the room where the deal was signed. Step back for a moment and you can see it: the tollbooths are being built at the choke points and dismantled at the record - a clause that priced a newspaper's archive as the fee for reaching its own readers suspended before it could take effect, a court adviser leaving the shared corpus open to a startup and a medical lab on the same footing as Google wherever publishers have not opted out, a verification tool published against its own author's interest, a science model anyone can download, with results its creators report ahead of GPT-6-Astra on FrontierScience-Research, a grid connection and a cost rule that put the bill in front of the household paying it - set against a search engine bundled to AI tools schools did not choose, memory capacity reserved half a decade ahead for whoever can sign, and a first job that fewer people are allowed through while everything else, including health coverage, still arrives only on the far side of it. Every transformation has a breaking point. Leverage can snap the beam it rests on... or raise a weight no one could have lifted by hand.

AI Releases & Advancements

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

New today

  • Apple: Opened a Siri AI Developer Preview in the iOS, iPadOS, macOS and visionOS 27.3 betas. Any developer can now connect their app to Siri AI through App Intents, and a new skills feature combines several app actions into one request. Siri AI can also now find any item in the Spotlight semantic index. (Apple Developer)
  • Google DeepMind: Released Cheminformatics Skills for Google Antigravity. They are built on RDKit and connect to ChEMBL and PubChem, so an agent can pull molecules out of papers, run matched molecular pair analysis, build predictive models and suggest candidate molecules with synthetic routes. (Google DeepMind)
  • Google Open Source: Released Three-Live, a tool that turns AI-generated Three.js code into interactive 3D, AR and WebXR experiences on the web and on mobile. (Google Open Source Blog)
  • Cloudflare: Launched Zero Wire on Cloudflare Radar, a hub that gives plain-language briefings on selected critical CVEs and exploits so security teams can check whether they are exposed. (Cloudflare Blog)
  • Soniox: Released Soniox STT v6 (stt-rt-v6 for streaming, stt-async-v6 for files) across 60+ languages. The streaming model has a new option to take in the voice agent's own text, which helps it recognize short caller replies and filter out the agent's voice when it leaks back into the audio. Other changes include better formatting of numbers, names and terms, and better speaker diarization. The v5 models will be retired on December 15. (MarkTechPost)
  • NorthCUA: Open-sourced a family of computer-use agent models in three sizes: Nano (9B), Mini (35B) and Pro (397B, built on Qwen3.5). The release includes the weights and the Harness execution framework. The team says Pro beats Qwen3.8-Max and Kimi-K3 on OSWorld-Verified and WebArena-Verified. (Emergent)
  • OpenRUA (Tsinghua / UC Berkeley): Released uni-robot-agent under the MIT license, a harness that lets general-purpose LLMs control robots through tool calls. Running Qwen3.8-Flash-Next locally, it raised the RoboDojo score from 1.90 to 31.39. (GitHub)
  • Hugging Face: Released FineDNA, a dataset for training language models on DNA. It has more than 300B tokens of genomic sequences, papers and web pages that mix DNA and text, plus 200k fine-tuning and 98k reinforcement-learning samples. (Hugging Face on X)
  • PrimeBot Research: Released PrimeSim with model weights, inference code and training code. It is a multi-camera world model that simulates robots from their commanded actions, trained on 30,110 hours of demonstrations from 22 robot platforms, and it runs at 10.1 FPS. (arXiv)

Other recent releases

  • Anthropic: Launched message threads in public beta on the Claude Messages API. After the first request, an agent sends only the new messages and the ID of Claude's last reply instead of the whole conversation. Anthropic says the wait before Claude starts answering drops about 35% on 500K–1M-token requests, and prices and rate limits stay the same. It works on all Claude models through Anthropic, AWS, Microsoft Foundry and Google Cloud. (Claude)
  • Anthropic: Released FlashProve under Apache 2.0, a system that produces cryptographic proofs (zkSNARKs) showing a response came from a specific model without revealing the model's weights. The release includes the prover, a fast C++ verifier, a verifier written in Lean, Lean proofs of its security claims, and a technical report. On one H100 GPU it proves Llama 4 Maverick prefill at about 110 tokens per second with 24 requests per batch and a 1,024-token context window. (Anthropic)
  • Deep Principle: Released Hydrogen-1.0, an MIT-licensed open-weight scientific model for chemistry, materials science and life sciences. It has 753B parameters (mixture-of-experts) and a 1M-token context, and was post-trained from Z.ai's GLM-5.3 on lab records and discovery trajectories. Weights are on Hugging Face and ModelScope, with NVFP4 and MXFP4 builds. (Deep Principle)
  • DeepSeek: Added Creator Mode to its open-source DeepSeek Harness app, which lets users build custom plugins by describing them in everyday language. DeepSeek published the prompts behind its demo plugins, including a video-editing workbench, a paper reader and a token-usage statistics panel. (DeepSeek)
  • inclusionAI (Ant Group): Released Ming-3.0-audio-flash under the MIT license, a mixture-of-experts model that handles speech recognition, audio understanding and speech output in one model. It has 126B total parameters (6.5B active), comes with a 9.3B "Tiny" variant, and runs with vLLM-Omni or SGLang. (GitHub)
  • Flower AI: Released Endeavor 1.1, the successor to Endeavor 1.0. It adds stronger reasoning and coding, image and other multimodal input, and inference on GPUs located in the EU. (Flower)
  • SenseTime: Officially released SenseNova 6.8 Flash, a model built for complex real-world tasks that can coordinate work across tasks and sessions and produce longer outputs. (Yicai)
  • Tencent Hunyuan: Released Hyra 2.0, an automated research agent that builds reusable scientific models, tools and workflows instead of solving one task at a time. Tencent published code, weights and technical reports for models the agent trained, including HyInv, a 1.96M-parameter protein inverse-folding model. (MLLLM)
  • Runway: Released Runway Dev plugins for Claude Code, Codex and Cursor. One install connects a coding agent to a Runway Dev account through the Runway Dev MCP server and adds skills for building Runway API integrations. (Runway)
  • Hume AI: Launched the Audio Decision API, with a playground available now and a waitlist for API access. Developers send raw audio plus their own questions and answer options (yes/no, multiple choice, scales or time spans) and get back probabilities with confidence scores, with no transcription step. (Hume)
  • NVIDIA: Open-sourced NVIDIA Recovery as part of its DSX AI Factory Platform. It detects interruptions in distributed training jobs, identifies the failing resources, excludes faulty nodes and restarts jobs from checkpoints on healthy capacity, with Slurm integration. (NVIDIA Developer)
  • NVIDIA: Released NVIDIA Attribution as open source, a service and blueprint that combines a workload's logs, metrics and scheduler state to name the root cause of a distributed AI workload failure and the nodes involved. (NVIDIA Developer)
  • NVIDIA: Released Workload Map (formerly Karta), an open-source Kubernetes CRD and Go library that describes the structure of distributed workloads in one shared format. It ships with ready-made definitions for JobSet, PyTorchJob, Ray, KServe, LeaderWorkerSet, NVIDIA Dynamo and other workload types. (NVIDIA Developer)
  • Wix: Launched Studio Two, an AI-native editor that combines AI generation, visual design on a canvas and standard React code in one workflow. It comes with the Aria 2.0 agent, saved Skills, import from and export to tools like Claude, Codex and Cursor, and MCP support for agents. It is rolling out to Wix Partners first. (GlobeNewswire)
  • Tripo AI: Launched Tripo Agentic Canvas, an AI agent that builds node-based 3D workflows users can edit and reuse, covering generation, texturing, retopology, rigging and export, with batch processing across many assets. (PR Newswire)
  • Meituan (LongCat): Released LongCat-2.5, the full version of the model it previewed in September. It is a natively multimodal model with 1.6 trillion total parameters (about 48B active per token), built for long-horizon agent tasks. It has stronger visual understanding and agentic coding than LongCat-2.0, was trained with sequences of up to 1M tokens, and Meituan says it was trained entirely on Chinese-made chips. (IT之家)
  • NVIDIA: Open-sourced Auto Ontology under Apache 2.0. It generates a starting "semantic layer" from a company's database metadata, so AI agents write SQL against business definitions that people have checked and approved. Agents reach it through 9 read-only MCP tools, a CLI and an API. It works with Databricks, Postgres, Snowflake and DuckDB, and uses Nemotron models by default but accepts any OpenAI-compatible endpoint. A separate test harness (auto-ontology-eval) ships with it, but NVIDIA has published no text-to-SQL accuracy scores yet. (MarkTechPost)
  • Glyphh AI: Released Ada as a free beta. It is a classifier built on hyperdimensional computing instead of a transformer, and it needs no GPU and no training run. Ada reads plain-English fields in records, learns from a business's past recorded decisions, and returns the same answer for the same input along with calibrated probabilities and the past records it relied on. When nothing on record is close enough, it hands the case to a person or an LLM. (MarTech Series)

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