Sign in Subscribe
The Century Report

Unsealed Filings Show Labs' Own Files Admit Data Theft - TCR 09/18/26

Unsealed filings show Microsoft and OpenAI staff calling their AI training data theft, as three governments move to make data centers pay their way.

Ben Linford

Ben Linford

- 23 min read
Three-panel navy infographic: power grid and Scotland/South Africa maps; AI neural nets, molecules and agent swarms; news funnel, Microsoft 'theft of labor' memo, falling publisher traffic.

The 20-Second Scan

Track all of the arcs The Century Report covers here:

Progress & Claims Tracker

The 2-Minute Read

The clearest signal today came from documents the companies never meant to release. A brief unsealed Thursday in the New York Times copyright case put the industry's private verdict on its own foundational method into the open, where a Microsoft director's phrase, "the largest theft of labor in human history," sits beside an internal warning that the answer-engine business is a "doom loop" consuming the web that feeds it. OpenAI has publicly defended its training as transformative, while Microsoft executives have framed broad web scraping as inevitable. Read now, their own memos call it something harder to reframe.

The confession arrived as the extraction it describes came under pressure from two other directions. Within days, Scotland paused approvals on more than 20 data-center projects, South African groups filed against facilities that would draw 4.4 billion liters of water a year, and the US House voted 417 to 3 to let states make large data centers pay for the power plants their demand triggers. A cost-causation principle that once sounded radical now clears a chamber that agrees on almost nothing. The default forming across all three is single: whoever creates the demand carries the bill.

A coalition of Google, Nvidia, and Anthropic moved on the same constraint from the opposite side, proposing to free 100 gigawatts of grid capacity by making data centers pause noncritical loads when the system strains instead of building for their own worst hour. The buffer a grid holds is provisioned mistrust, priced in copper and turbines, and it shrinks the moment strangers can see one another's plans. Extraction confessed from inside, extraction constrained by consent, extraction routed around by coordination: three pressures on one assumption, that costs can be billed to someone downstream out of sight.

Underneath the governance fight, the capability kept compressing. A Chinese lab watched its own model build the inference system that will train its successor, in under two weeks. Novo Nordisk, down 20% and chasing rivals it once led, aimed Claude at drug discovery itself instead of the back office. A Stanford spinout ran a virtual biotech of 37,000 agents with no humans in it. King Charles gathered the labs to warn of "existential dangers," and the sharpest answer arrived from the EFF, which urged lawmakers to ground rules in demonstrated harms and in verification the labs do not themselves control. The settlement that will hold is forming in those votes and coalitions, underneath the summit.

The 20-Minute Deep Dive

Microsoft and OpenAI Called Their Own Training Data "Theft," Unsealed Filings Show

A brief unsealed Thursday in the New York Times copyright case pulled three years of internal correspondence into the open, and the private language runs hard against the public defense both companies have mounted since 2023. A Microsoft applied-science director wrote in a January 2023 memo that mass web scraping was "an astonishing theft of unprecedented proportions" and "the largest theft of labor in human history." The Century Report covered the September 6 disclosure that Microsoft could quantify how often its systems reproduced sixteen consecutive words of a publisher's work, and the September 3 filing in which the federal government backed the labs' fair-use position. What is unsealed now is different: the companies' own internal verdict on what they built.

That verdict includes an account of how the material was taken. The filing describes engineers assembling training sets that leaned heavily on scraped news, pulling millions of articles from an open crawl repository, and stripping copyright notices before the data reached the model so outputs would not surface them. When an OpenAI researcher flagged "a hack to get around nytimes paywall," a cofounder replied, "ah nice." OpenAI's mid-training data alone reportedly held more than 91,000 copies of works from the plaintiffs.

The sharpest passage is not about the taking but about what follows it. An internal Microsoft presentation found that after its answer engine drew on a publisher's reporting, click-throughs to that publisher fell as much as 93 percent, and named the result a "doom loop" that would "hurt the performance of our models and the entire web at the same time." "It is highly unusual that an end-product threatens the economic foundations of its essential suppliers," the document reads, "but that is the situation we have created." A pure-extraction model, described from inside, consuming the very sources it needs to keep learning.

Hold OpenAI clearly as the extractor here, and hold the other fact beside it: this same week its finance chief sat at a monarch's summit arguing for AI in the service of the public good, and the company has pledged outside evaluators inside its walls. One contested actor saying one thing to a king and another in a paywall thread. The people carrying the loss are writers and newsrooms losing both compensation and the readers who once arrived by a link, and that loss is landing on them now. What the unsealed record changes is that the industry can no longer frame the arrangement as transformation or inevitability while its own memos call it theft and warn it eats the web that feeds it.

Microsoft’s warning exposes a limit to capturing the audience: the company still depends on people producing fresh knowledge beyond its systems. Its own account ties model performance to the survival of those contributors, making their continued ability to investigate, discover, and publish part of the service’s foundations.

Google, Nvidia and Anthropic Back a Coalition to Free 100 GW Without Pouring Concrete

A grid holds a buffer for the same reason a household keeps a spare: nobody can see what everyone else will do, so each provisions alone for its own worst hour. Most of the year that reserve sits idle. On September 16, Emerald AI, Google, and Nvidia launched the AI Energy Management Alliance, joined by Anthropic and utilities including Constellation, National Grid, NRG, and AES, on a bet that making data-center demand legible to the grid could connect an additional 100 gigawatts of computing without building the plants to match. The mechanism is demand response, decades old for factories: when the system strains, a large customer dials back. What Emerald AI adds is software that lets a data center pause noncritical tasks or shift compute to a campus where the grid has headroom, responding in seconds rather than firing up diesel generators.

The number is not small talk. A Goldman Sachs study last year found that capping data-center draw at 90 percent for a few hours at a time could free 76 gigawatts of existing capacity, power already bought and standing unused against a shortage that rarely arrives. Emerald's chief scientist, Ayse Coskun, keeps the claim honest: flexibility blunts the need for new generation, it does not erase it.

There is a second route to the same end, and it exposes where the friction actually lives. Verrus, an Alphabet spin-out, is building data centers backed by large battery arrays instead of diesel, cooled in closed loops that cut water use by 95 percent or more, packing roughly a third more servers into the same grid connection. It has answered nearly every objection raised against the buildout on paper. Both of its proposed sites, in Michigan and Oregon, are mired in local opposition, and it has not finished one yet.

Credit where it is due: the coalition's approach releases capacity the extractive path would have poured new concrete and burned new gas to duplicate, and Nvidia, elsewhere arguing that AI safety is an engineering problem for the market rather than a matter for lawmakers, is on the commons-aligned side of this one. The open question is the one that decides every coordination story - who runs the layer that pools the loads, who can audit it, and who holds the claim on the capacity it frees. The September 5 edition of The Century Report documented that same coordinating layer at household scale. When PG&E pooled 21,000 home devices into a virtual power plant earlier this month, the same question sat underneath. The buffer was mistrust all along, priced in copper and turbines, and it shrinks the moment strangers can finally see one another's plans.

GLM Builds the Machinery That Will Train What Comes After It

In an account published Thursday, the Chinese lab Z.ai described something its engineers said unsettled them: Z.ai says a coding agent powered by its GLM-5.3 model built much of the production inference system for the model's successor, GLM-5.3-Flash, under human supervision, doing work the company says a team of experienced engineers would have spent weeks on. The build ran, according to Z.ai, on a cluster of more than 100,000 Chinese-made AI accelerators at what the company described as an unprecedented scale, and went from a first successful run to serving live traffic in under two weeks while tripling throughput. Z.ai says the result reached per-token cost and hardware efficiency comparable to mainstream NVIDIA GPUs, a figure the company reports about its own silicon.

The lab names the direction plainly: the work "would directly change how the next generation of models is trained," an early form of what it calls recursive self-improvement, a model helping design and train its own successor. That framing serves a roadmap, and the lab has every reason to want it believed. What is harder to wave off is the mechanism underneath it, and two papers from the same window document that mechanism in the open.

Researchers at Fudan University, building on Z.ai's 744-billion-parameter foundation, released Atria Dawn, an agent for research and engineering work that posts the top score on five of sixteen benchmarks. More revealing than the scores is the study they ran on their own project: across 769 task records from 56 people, participants judged roughly a third of the completed AI-assisted work infeasible without the agent, which frequently proposed methods and wrote the revisions while the humans kept most of the final calls. A separate paper, Dream-RSI, lets agents replay their own past discovery attempts as a cheap simulator, refining not just their answers but the way they search.

The Fudan team's own hedge is the honest one: a model can get better at the tasks it trains on without getting better at building its successor. What the day's evidence actually shows is a shift from machines executing tasks toward machines and people working as partners on the harder question of what is worth pursuing, with human judgment, for now, holding the wheel. The assumption eroding is that making a frontier model requires scarce elite labor pooled inside a handful of firms. The assumption still standing is that someone accountable decides where all of it points.

Novo Turns to Claude for the Molecule Itself

Novo Nordisk, the Danish company that made Ozempic a household word and then watched its lead in GLP-1 drugs slip away, said on September 16 that it will work with Anthropic to speed the discovery of new medicines. The company, now rebranding to simply "Novo," will put Anthropic's models and Claude Science to work on biological reasoning for research and development and on the software engineering underneath it. As the July 1 edition of The Century Report covered, Anthropic launched Claude Science with more than 60 tools across genomics, proteomics, and cheminformatics; Novo's deal moves that system into a major drugmaker's R&D workflows.

The detail that gives the deal its velocity is where the AI is aimed. Rivals Lilly, Merck, and Roche have all bought into AI, but largely for workplace productivity and for designing clinical trials, the back-office end of the business. Novo is pointing it at discovery itself, the search for the molecule. Here the company reaching for the capability is the one that fell behind: its shares have dropped about 20% over the past year, and the discovery partnership arrives alongside a culture shake-up and the name change.

Both companies described what they expect the collaboration to do, and both descriptions are forecasts that happen to flatter their own plans. Novo's chief executive said AI could "compress the path from research to marketed product" and open scientific opportunities that did not exist before. Anthropic's Dario Amodei went further, saying capable models carry "the potential to compress a century's worth of biological and medical breakthroughs into a decade." Those are the ambitions of the parties who stand to gain from the deal, and not yet results. What Novo actually committed to is an initial test of Claude Science on specific R&D workflows, built, it says, with data governance and human oversight. No candidate, no compound, no trial data. A partnership moves the date a discovery might arrive; it does not produce the discovery.

Held to that honest scope, the opening is still there. If reading and reasoning about human biology gets faster and cheaper, the years between a biological insight and a treatment for a chronic disease start to compress, and the door to that compression is not one any single firm gets to hold shut. A company down 20% can walk through it as readily as the one on top. The caution belongs beside the wonder: a molecule found at the bench still passes through patents, manufacturing, and price before it reaches anyone, and the fight over who can afford a GLP-1 drug is the standing proof that capability at the lab does not become affordable medicine on its own.

Three Jurisdictions Ask Who Pays for the Data Centers, and Who Decides

Within one week, three governments on two continents moved on the same question, and it was not whether to build. On September 16, Scotland's parliament voted to hold every planning decision on data centers above 50 megawatts until national guidance is written, with full rules promised inside a year and mandatory environmental assessments starting now. Ministers declined to call it a moratorium, but the effect is a pause on more than 20 proposed projects, one of them billed as the second-largest data center in the world.

The same day, the US House passed the Ratepayer Protection Act by 417 to 3. The September 14 edition of The Century Report noted that the House was preparing the bill for a floor vote. The Century Report followed this bill through committee in July; on Wednesday it cleared the full chamber. It sets a standard states can adopt to require data centers drawing 100 megawatts or more to pay for the power plants and transmission lines their demand triggers, rather than folding those costs into everyone's electricity bill. Look closer at that near-unanimity. The three who voted no were progressive Democrats who found the bill too weak, since it only invites states to act rather than compelling them, and environmental groups called the voluntary framing a handout. A cost-causation principle that once sounded radical now clears a chamber that agrees on almost nothing, weeks before an election in which data centers have become a live issue.

South Africa marks the frontier where this gets hardest. Civil-rights groups, joined by the nonprofit Foxglove, filed a legal challenge against two Equinix data centers in Cape Town and are pressing for a national moratorium. The proposed 174-megawatt facilities would draw roughly 4.4 billion liters of water a year, about what 18,000 homes use, in a city that nearly ran dry in 2018. That figure would shrink by 95 percent under the closed-loop cooling now available, so the water cost is a design choice rather than a law of physics - but in a water-stressed metro where residents already queue for supply, the concentration is genuinely severe, and it is landing on the community least able to refuse it. Africa hosts most of the continent's data centers and still holds under one percent of global capacity for a fifth of the world's people.

The compute is needed. The dispute is narrower - who absorbs its cost, and who gets a say before the concrete is poured. What is forming across all three votes is a single default: the party creating the demand carries the bill, and consent and disclosure become the price of a connection. That principle is spreading one legislature at a time, and every place it lands, the era where infrastructure billed its neighbors out of sight gets a little shorter. Where no such rule is on the books, the demand routes to the towns with the fewest legal resources to fight it, which is where an industry audit found most of the off-grid gas capacity now heading.

The King Hosts the Labs as the Pacing Fight Hardens Into a Culture War

King Charles convened a private summit Thursday at his Ayrshire estate, gathering Nvidia's Jensen Huang, leaders from OpenAI and Anthropic, the UK's AI minister, the head of British foreign intelligence, and the pope's AI adviser. The monarch, who usually stays clear of live political questions, urged the room to find "sufficient means of control before it's all too late," warning of AI's "existential dangers" if it fell into the wrong hands. A new head-of-state voice entered a fight that has, over two weeks, split the industry down the middle.

On one side, Anthropic's Dario Amodei and OpenAI's Sam Altman want frontier labs to slow where needed and open the door to independent evaluators. On the other, Meta's Mark Zuckerberg and Nvidia's Huang reject the premise that new law is needed at all, arguing the market disciplines what people will not trust. Both poles deserve the same scrutiny. A slowdown designed by the labs already in front doubles as a moat around their lead; self-regulation left to the market is the oldest script any industry reaches for when it would rather write its own rules. When the largest players converge on either, the convergence points to shared position before it points to settled danger.

Nvidia's argument deserves a fair hearing as the company's position. Its VP of agentic AI cast safety as an engineering problem, arguing the scaffolding built around a model matters more than the model, and that the field already knows how to secure it. That is a claim offered by a company that sells the hardware every lab runs on and would carry the friction of any brake. Its politics are genuinely mixed rather than villainous: Huang champions open-weight models so smaller countries are not left behind, and Nvidia joined utilities and rivals in a coalition to fit more compute onto the existing grid by pausing noncritical loads instead of pouring new concrete, a real commons move.

The harder problem is who writes the rules. Cohere warned that a regime authored by a few dominant firms, especially one handed an antitrust exemption to coordinate, would be "a cartel by any other name," and its chief AI officer argued the labs building the systems should inform policy without becoming its sole authors. The commons-aligned answer arrived from the EFF, which urged lawmakers to ground any rules in demonstrated harms rather than doomsday scenarios: the summer's headline lab breaches, its analysts argued, might have been prevented or mitigated by longstanding cybersecurity practices such as proper sandboxing and monitoring, and better understood through mandated independent investigations whose reports the public can read. That is the thread the culture-war framing buries. As effective altruism became a political flag and evaluators drew attacks as "globalists," the concrete and addressable present slipped behind an argument about an abstract future, and the settlement that will actually hold is forming underneath it, around verification that the people building these systems do not themselves control.

The Other Side

An answer engine still needs someone to find out what happened. Microsoft’s internal warning exposes the weakness in a business that absorbs reporting while diverting the audience that sustains it. Publishers depend on those visits to support further reporting. For writers caught between the two, the pressure reaches into ordinary life: another assignment disappears, another month becomes harder to plan, another investigation gets abandoned before anyone learns what it would have found.

Microsoft’s own document warns that weakening its sources threatens its models. That warning carries a consequence beyond the copyright dispute. A company can accumulate yesterday’s reporting and still depend on people discovering tomorrow’s facts. Capturing the route to an answer gives it no permanent substitute for the people who go looking.

Meanwhile, researchers are demonstrating partnerships that expand what people can investigate. In Fudan’s Atria Dawn study, participants judged about a third of their completed AI-assisted tasks infeasible without the agent. The study concerns research and engineering. Its relevance here is the growing capacity to pursue inquiries that previously exceeded a person’s resources. The same broad advance disrupting distribution also gives people stronger partners in discovery.

Imagine yourself in 2036, sitting in the foyer of a restored neighborhood cinema with your recorder on the table. You once wrote local features for a living. Now your home, food, and care are secure through the productive resources your community shares. The community also maintains the computing equipment where your AI collaborator works alongside you. During the difficult decade, people carried the lesson exposed by those 2026 documents into rebuilding how knowledge was sustained. They shared the gains widely enough that someone could investigate a story without first proving it would attract an audience. It no longer mattered if a piece may or may not turn a profit. Consequently, storytelling and journalism exploded as people had more time to investigate, explore, and write about their interests.

Your collaborator has organized the archive and transcribed the interviews. You came because you wanted to experience the cinema as it was in a simpler time, and to understand the projectionist's labor of love culminating in the years spent bringing this place back. Your security and your value as a person do not hinge on whether the story sells. The story deserves to be told, so you tell it. You ask about the first film she saw here. She laughs, then pauses as another memory returns. You have time to listen.

The Century Perspective

With a century of change unfolding in a decade, a single day looks like this: a brief unsealed Thursday putting the industry's private verdict on its own foundational method into the open, where a Microsoft director's "the largest theft of labor in human history" sits beside an internal memo naming the answer-engine business a "doom loop" that eats the web it feeds on, Google, Nvidia and Anthropic joining utilities in a coalition to free 100 gigawatts of grid capacity by pausing noncritical compute when the system strains instead of pouring new concrete, Verrus answering nearly every objection to the buildout on paper with battery backup, closed-loop cooling that cuts water use 95 percent, and a third more servers per grid connection, Z.ai watching its own GLM-5.3 agent build the production inference stack for its successor on 100,000-plus Chinese-made accelerators in under two weeks, a Fudan team publishing 769 task records showing roughly a third of their AI-assisted work would have been infeasible without the agent while humans kept most of the final calls, Novo pointing Claude Science at the molecule itself rather than the back office, a Stanford spinout running 37,000 agents through 50,000 clinical trials in under a week, Scotland holding every data-center decision above 50 megawatts until national guidance is written, the US House passing cost-causation 417 to 3, and the DOE opening a $215 million competition for the first fault-tolerant machine with 100-plus logical qubits. There's also friction, and it's intense - engineers stripping copyright notices before the data reached the model, a cofounder answering a paywall workaround with "ah nice," more than 91,000 copies of the plaintiffs' works in mid-training data, click-throughs to a publisher falling as much as 93 percent after the answer engine drew on its reporting, the same company's finance chief arguing for AI in the public good at a monarch's summit the same week, two Equinix facilities in Cape Town that would draw 4.4 billion liters a year in a city that nearly ran dry in 2018, Africa holding under one percent of global capacity for a fifth of the world's people, the three House votes against coming from members who found the bill too weak to compel anything, Verrus stalled by local opposition at both its sites with nothing finished, Z.ai's efficiency figures reported by Z.ai about Z.ai's own silicon, Amodei's century-into-a-decade promise attached to an initial test with no candidate and no compound, senators trying to slip an antitrust exemption into the defense bill, and Cohere warning that rules authored by a few dominant firms are "a cartel by any other name." But friction generates heat, and heat shows exactly where the load was being carried all along. Step back for a moment and you can see it: one assumption under pressure from four directions at once - that the cost of this buildout can be billed to someone downstream out of sight - confessed from inside by the people who wrote the memos, constrained by a parliament and a chamber that agrees on almost nothing, engineered around by a coalition that found 100 gigawatts sitting unused inside everyone's private margin for mistrust, and contested in Cape Town by residents who already queue for water, while the EFF points past the existential framing to the summer's actual breaches and the sandboxing that would have stopped them. Every transformation has a breaking point. A ledger can condemn everyone whose name appears on it... or finally settle a debt nobody could collect while it went unwritten.

AI Releases & Advancements

New today

  • Anthropic: Rolled out Projects in Claude Code (desktop and web beta), a coordinator agent that splits work into parallel cloud-session "threads," each running on its own git branch with shared project memory. (Anthropic)
  • Salesforce / NVIDIA: Launched Koa, Salesforce's first CRM-specific reasoning model, post-trained on NVIDIA's open-weight Nemotron-3-Super-120B using synthetic (non-customer) data for sales/service workflows; in pilot now with US general availability planned for Winter 2026. (Salesforce)
  • Snap: Launched Specs Intelligence, an "anticipatory AI" agent for its Specs AR glasses that connects other apps/accounts to manage tasks and goals; available now in preview on iOS, with a Mac waitlist open. (Snap Newsroom)
  • Amazon: Expanded Alexa+ to India, offering the AI-upgraded assistant free to select users at launch. (India Today)

Other recent releases

  • Anthropic: Merged Claude Chat and Cowork into a single unified interface and launched Claude Docs and Claude Slides in beta, letting users create, edit, and export documents/presentations directly within any conversation. (Anthropic/Claude Blog)
  • Google: Launched early access to a Model Context Protocol (MCP) server for Google Home, letting any MCP-compatible AI agent (Claude, ChatGPT, Hermes, OpenClaw, Google Antigravity) control connected smart-home devices and query camera/event history. (Google Home Developers)
  • Odyssey: Released Odyssey-3, a general-purpose foundation world model whose single frozen backbone adapts via lightweight decoders to control robot arms, humanoids, autonomous vehicles, drones, and video-game characters. (Odyssey)
  • OpenAI: Published a new framework for tracking, investigating, and disclosing AI model misalignment incidents, launching with six detailed reports of misaligned behavior observed during recent training and evaluation. (OpenAI)
  • Knowledgator: Released GLiFormer, an Apache 2.0 schema-conditioned encoder unifying NER, text classification, relation extraction, and nested JSON structuring in one compact self-hostable model, with Base (264M) and Large (575M) checkpoints. (Hugging Face)
  • Zhongguancun Academy / Zhongguancun Institute of AI: Released ZGCM-1, a fully open 7B dense foundation model for math and agentic search, using a hybrid sliding-window/global attention architecture and FP8 Muon training to compete with much larger 235B-class models. (Hugging Face)
  • Agility Robotics: Unveiled Digit 5, its next-generation humanoid robot engineered for cooperatively safe work alongside people without physical safety barriers, featuring 40% more payload, 9-minute fast charging, and safe human-detection AI; deliveries begin early 2027. (Agility Robotics)
  • Salesforce / NVIDIA: Announced Koa, Salesforce's first CRM reasoning model for Agentforce, built by post-training NVIDIA's open-weight Nemotron 3 Super on synthetic enterprise CRM data; already in customer pilots. (Salesforce)
  • Salesforce: Unveiled AIforce, a live interface layer that brings Salesforce's data, workflows, and business logic to any AI interface (Claude, Slack, Lightning), launching with Claudeforce, Slackforce, and Agentforce Coworker. (Salesforce)
  • TypeSafe AI: Released Jev, its first "System One Model," a new class of frontier model built for fast, structured decisions rather than text generation, claiming two orders of magnitude faster/more efficient inference than standard LLMs; available today in early access. (TypeSafe AI)
  • Prior Labs: Released TabPFN-3.5, a tabular foundation model ranking #1 across seven benchmarks including TabArena and BeyondArena, and beating the 2015 Kaggle Otto competition's winning solution with default settings; open weights for research use. (Prior Labs)
  • Meta: Launched WhatsApp Business Tools MCP, a Model Context Protocol server letting AI coding agents (Claude, Cursor, Codex, ChatGPT) set up and manage WhatsApp Business messaging directly. (Meta for Developers)
  • Nums AI: Released Causilo, a pretrained tabular foundation model for classification and regression with the highest Elo among single models on TabArena; Apache-2.0 code with research-only pretrained weights. (GitHub)
  • Gensyn: Launched open-1b, described as the industry's first "auditable AI model," shipping with training recipe and verifiable proof allowing third parties to independently audit training. (PR Newswire)
  • Creatify Labs: Launched Boreal, a text-to-video/image-to-video AI model built on Lightricks' LTX-2.5, generating video in realtime at roughly 1 cent per second, available via API, Model Playground, and fal. (PR Newswire)
  • DeepL: Released new Voice AI models adding real-time voice-preservation (tone, emotion, pacing) to multilingual conversations, plus a new desktop app for Zoom/Teams/Meet with voice-to-voice translation in 30+ languages, now generally available. (PR Newswire)
  • Lectra: Launched Apogy, a cloud-based agentic AI solution for fashion product development that unifies data, processes, and stakeholders in one environment. (PR Newswire)
  • SimScale: Launched an Engineering AI Agent for Onshape that reasons through CFD, FEA, thermal, and electromagnetic simulation setup from a single prompt, available now via the Onshape App Store. (SimScale)
  • Shanghai AI Laboratory (InternLM): Released Atria Dawn Preview, a 744B-parameter agentic MoE foundation model with MIT-licensed weights and 1M-token context for research and engineering tasks, available now on Hugging Face. (Newsfile / MarketMinute)

Sources and Further Reading

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