Sign in Subscribe
The Century Report

A Village Takes On Google's $15 Billion Data Center

A village's farmland was cleared for a $15bn Google data center approved in nine days with no hearing. Now three petitions contest it in court.

Ben Linford

Ben Linford

- 24 min read
Three-panel navy infographic: a power-grid sensor freeing stranded gigawatts, a bulldozer clearing a village for a data center, and layoff headlines above recent-grad unemployment data.

The 20-Second Scan

Track all of the arcs The Century Report covers here:

Progress & Claims Tracker

The 2-Minute Read

A quieter thread than the day's headline figures runs through developments landing together, and the clearest place to see it is a village in Andhra Pradesh, where farmland was taken for a $15 billion data center on clearances issued in nine days with no public hearing. Residents who tried to object found the process closed to them, and the activist documenting it watched his videos pulled offline. What governed the outcome was who was allowed to see the project coming and refuse it, and that same issue surfaces nearly everywhere else on the page: who gets to inspect the reality underneath a number.

The same lesson turns hopeful in the federal grid award. Its $1.9 billion fits transmission lines with sensors that read their true capacity hour by hour, and the Department of Energy expects to free roughly 23 gigawatts already sitting in wires run conservatively because no operator could see how much they safely carry. Observation released supply that no amount of concrete would have. Microsoft's rebuilt Copilot raises the requirement from the other side: its persistent agent keeps working after a user signs off, and chief executive Satya Nadella named the condition before shipping it, that every agent needs an identity and everything it does needs to be observed.

Where that visibility goes missing, the cost shows up fast. A security firm found roughly 16,000 databases built on Supabase leaking personal records to the open web, the residue of apps that AI let non-engineers ship without anyone running a basic configuration check. The labor numbers carry the same gap: a tracker counted more than 94,000 US tech layoffs through August with AI blamed in a third, while a new CESifo working paper reading granular Census microdata found no sign of the graduate-hiring collapse the story rests on. The alarming total arrived ahead of the evidence for its cause.

The MIT crisis-text method shows what this looks like when the visibility is built in on purpose. It estimates suicide risk from the words of people already on a crisis line and flags the terms behind each assessment, so a counselor can check the basis instead of trusting a bare score, capability aimed at helping skilled people reach more of the queue. Take the day together and the pattern comes clear: the same power to see - a grid reading its own capacity, a counselor seeing the grounds of a risk call, an administrator watching what an agent does - decides whether this buildout stays held by a few or opens to many, and the grid sensors and the openly released crisis lexicon both point toward opening it.

The 20-Minute Deep Dive

A Village's Farmland, a $15 Billion Data Center, and Who Gets to Refuse

In Tarluvada, a village on an unpaved road in the south Indian state of Andhra Pradesh, officials arrived last year with what they called a golden opportunity: Google's planned gigawatt-scale AI data center in India, a $15 billion project built with the Adani Group. Jobs would become plentiful, land prices would soar, compensation would be generous. What residents describe instead is farmland previously allocated to villagers and taken back by the government, promises of replacement land and work left unmet, and bulldozers stripping a hillside of more than 100 acres of trees. "What will happen to my crops and my livelihood?" one resident asked, in a region where summer temperatures already reach 45C. This grievance is genuine, and it sits on people with the least power to refuse it.

The approval deepens the concern. Publicly the project was announced at one gigawatt; state environmental clearances list a combined 2.51 gigawatts for the planned data-center sites, the output of two large nuclear reactors and, if run continuously at that capacity, over a year, upwards of 30% of the electricity a state of 55 million people consumes, on a grid still roughly three-quarters coal. The clearances were issued in nine days with no public hearing. The village head, BRB Naidu, said he was torn over whether the project would bring prosperity or ruin, and that the government had refused to answer the village's questions. Three petitions now sit before India's National Green Tribunal alleging the environmental assessment was "grossly inadequate, scientifically unrealistic, and misleading," and activist VS Krishna of the Human Rights Forum says his campaign met takedowns of its videos and social pages.

Half a world away, Australian officials argued that resentment of data centers is not really Australian. Their response extends the pattern the September 23 edition of The Century Report documented when industry figures dismissed US opposition as nimbyism. Victoria's then-economic-growth minister Steve Dimopoulos called the backlash a product of US social media, insisting "we are not the United States"; Belinda Dennett, who heads the lobby group Data Centres Australia, said the debate was being "fought on America's facts." Their scale point holds some merit: Australia has 252 data centers at 1.4 gigawatts against roughly 5,400 and 53.7 gigawatts in the US. What Dennett left out is that per person the gap nearly closes, and that Australia has a 2.4-gigawatt campus planned for outer Melbourne and a 2-gigawatt gas-fired one in the Northern Territory, home to 265,000 people. University of Sydney's Rob Nicholls called the objection - higher power bills, higher water costs, no local benefit - "a rational position."

Read together, the two stories separate a real displacement from the effort to manage how it is discussed, and they share a foundation beginning to crack: the assumption that a buildout's costs can be routed onto whoever can least object and settled later. The petitions, the litigation, Australia's proposal to limit reliance on gas and require operators to match demand with new renewable energy - these are that assumption meeting resistance that raises its price. The compute is needed and can serve broadly. Where it lands, and who is allowed to say no, is now being contested in courts and cabinets rather than conceded in advance.

Victoria's proposal makes additional electricity supply part of the conditions for building: operators must match new demand with new renewable generation and storage. That requirement directs investment toward expanding what the grid provides, giving the compute buildout a concrete role in relieving the scarcity it intensifies. These are proposed conditions; their contribution depends on generation and storage reaching operation.

The Grad Displacement Everyone Forecast Isn't in the Data Yet

Last month a Stanford study found entry-level employment in occupations most exposed to AI lagging behind the rest of the labor market, and it became the empirical spine of a familiar worry: that the technology would hollow out the bottom rung first, freezing new graduates out before it touched anyone senior. The September 22 edition of The Century Report also tracked that hiring concern through Goldman Sachs's estimate that US call-center employment had fallen 39% below trend. A working paper published Friday by economists Robert Fairlie and Jane Wu at Munich's CESifo went looking for that collapse in the granular record and reported the opposite. Reading US Census Current Population Survey microdata on bachelor's recipients aged 22 to 25 who are not pursuing further degrees, and tracing year-over-year and seasonal patterns back to 2022, they found "no evidence through August 2026 of any significant, widespread displacement or reduction in hiring of recent U.S. college graduates in absolute or relative levels."

They chose recent graduates deliberately, reasoning that if firms were handing standardized entry-level tasks to AI, the first sign would be hiring that simply stopped rather than layoffs that made news. The premise had support behind it. A Census survey shows a sharp rise in firms replacing employee tasks with AI, and spending per employee and enterprise token use both climbed over the past year. BlackRock chief executive Larry Fink warned in March that this year's graduates could face the highest unemployment in years without a recession, and venture capitalist Marc Andreessen argued the models only crossed the capability line late in 2025.

Set that against the aggregate that keeps making headlines and the shape of the story changes. A Crunchbase tracker counted at least 94,046 US tech layoffs from January through August, up 16.8% over the same stretch of 2025, concentrated in bursts: a May peak of 31,513 that included Meta's 8,000 cuts, then a decline every month down to 2,347 by August. Those job losses land on named people carrying rent and families, and nothing in the microdata softens that. What the microdata unsettles is the cause stapled to the number. AI was cited in 33% of tech layoff events tracked by Layoffs.fyi this year, up from 1% in 2024, yet Roger Lee, who runs the tracker doing the attributing, said flatly that "there's been little evidence that AI is actually replacing the work of the human employees let go."

The premise under most coverage is that a job vanishing and AI arriving are the same event. Fairlie and Wu, and Lee's own caveat, pull those apart. Andrew Challenger of Challenger, Gray & Christmas describes what the data appears to be catching instead: firms cutting one part of the business while hiring in AI, a reshuffle more than a contraction, with a plausible upside for builders as software gets cheaper to make. The instruments now testing the mechanism against monthly Census figures are the development that will settle it. The fear may prove accurate later; for now the measurement is arriving fast enough to say the collapse being described has not surfaced where it would surface first.

The Grid Upgrade That Pours No Concrete

On Thursday the Department of Energy selected 31 projects across 26 states for a proposed $1.9 billion in federal funding aimed at a deceptively simple idea: get more electricity through the lines already strung between the towers. When The Century Report last covered SPARK on March 14, the $1.9 billion was an announced pool for reconductoring and other grid upgrades; Thursday's selections specify where that commitment will go. Utilities pledged $3.35 billion to match it, bringing the total to $5.25 billion, and the DOE projects the work will unlock more than 23 gigawatts of additional grid transmission capacity, roughly the output of two dozen large power plants, without building those plants. The money comes from the SPARK program, funded out of the 2021 infrastructure law, and it landed as one of the rare energy moves the current administration and its predecessor both back.

Most of the projects deploy two technologies proven in pilots for more than a decade and deployed far too slowly since. The first is advanced conductors, power cables lighter and stronger than the aluminum-and-steel lines that carry most of the grid, capable of moving more electricity along the same corridors without new towers. The second is dynamic line rating, sensors and software that read a line's true carrying capacity hour by hour instead of assuming a fixed worst-case limit, letting operators push more power through wires that were being run conservatively. Together the awards will reconductor or rebuild more than 1,500 miles of line and fit grid-enhancing devices across nearly 21,000 miles.

The two largest awards, a $1.2 billion Colorado project and an $832 million Oklahoma one, each drawing $250 million federal, both aim at the same seam: expanding how much power can move between the Eastern and Western grids, two systems that have historically shared very little. Energy Secretary Chris Wright framed the grants as getting "more out of the infrastructure we already have," and the demand driving the urgency is explicit in the DOE's own language about speeding data-center connections.

What makes this more than a spending line is what it says about how capacity gets provisioned. A grid built for slow, predictable growth runs its lines well below their true limits because no operator could see, moment to moment, how much a wire could safely carry, so each planned against its own worst day. Instruments that let the system see the actual capacity release power that was already present, stranded by an inability to observe it. That path is cheaper and faster than pouring concrete, and it is starting to win on those terms. The buildout still needs new generation, and Fervo's first geothermal power is part of that. But the capacity hiding inside the existing wires could move more electricity, often without securing new land or a new town's consent, an answer to the same pressure that put bulldozers on a hillside in Andhra Pradesh.

Microsoft Rebuilds Copilot Around an Agent That Keeps Working While You're Away

On Friday, Microsoft reorganized Copilot around three parts and pointed the whole thing at the office suite hundreds of millions of people already open every morning. Home folds Word, Excel, and PowerPoint into a single chat surface. Code lets anyone describe an app, dashboard, or automation in plain language and have Copilot build it, running on the same underlying technology as GitHub Copilot inside a company's own environment. And Autopilot, renamed from the Scout agent Microsoft showed in June and built partly on the open-source OpenClaw, is a preview-stage persistent agent the company describes as continuing to work when its user has signed off.

Chief marketing officer Jared Spataro reached for the largest comparison Microsoft has: "Just as Office defined work for the PC era, the new Copilot is built to define it for the AI era." Chief executive Satya Nadella called it "a new OS for work." Those are a vendor's claims about its own release, and the more telling line came from the same September 23 presentation, where Nadella named the condition a persistent agent creates: "Every agent has to have an identity. Everything it does needs to be observed".

Autopilot runs with its own identity, memory, and workspace inside an organization's tenant, and employees summon it in Teams or Outlook by name, the way they would a colleague. In Microsoft's demo an agent called Dot watched Black Friday preparations across email, inventory records, and spreadsheets, flagged a shipment problem hitting 18 stores, and pulled people in to fix it. That is a scripted scenario, and VentureBeat noted what the launch materials do not settle: how reliably the agent handles comparable situations across real deployments, or how administrators should govern every category of failed action.

Two things sit inside this that the demo language moves past. The first is cost: Cowork, Code, and Autopilot are billed by usage, so what a company pays now tracks how much its agents do, a meter that did not exist when software was a flat annual license. The second is who gets to build. Code treats a small, purpose-built application as a fourth unit of office work alongside the document, the spreadsheet, and the deck, and Microsoft's framing is that assembling one is becoming as ordinary a skill as writing a memo. The assumption coming apart is that making working software required being a developer at all. What determines whether that widening reaches people or concentrates gains is who can inspect what these agents do, and Nadella named that requirement himself before shipping the thing that raises it.

An MIT Method Reads Suicide Risk in the Words of People Already in Crisis

Deciding who is in immediate danger is one of the hardest calls a crisis counselor makes, and it has to be made fast, from words alone, while a distressed person is still on the line. Researchers at MIT's McGovern Institute built a language system aimed squarely at that moment. Reporting in the Journal of Psychopathology and Clinical Science, Daniel Low, Satra Ghosh, and colleagues describe a method that scans crisis-line text for language tied to 49 established suicide risk factors and estimates how acute the danger is.

The team, working with the nonprofit Crisis Text Line under specialized training and controlled access, analyzed roughly 16,000 de-identified conversations the service had already sorted into three levels: non-suicidal, suicidal thoughts without imminent danger, and imminent risk, the last meaning a plan or an intent to die within 48 hours. They first built a library of about 60 words and phrases for each risk factor, drafting an initial list with AI and then having expert clinicians review and confirm every entry. A machine-learning model searches a conversation for those terms and weighs how much each contributes to an estimate.

What the model surfaced was not always what a layperson would guess. Mentions of lethal means and substance use showed up in the highest-risk conversations more often than expressions of depressed mood or fatigue, even though depression is the risk factor most people associate with suicide. Direct references to lethal means, words like "cut" or "pills," carried heavy weight; language of hopelessness counted for less. On conversations it had never seen, the system predicted the counselors' assigned risk levels in held-out conversations, according to the researchers.

Two limits keep this honest, and the researchers name both. The model matches the risk categories human counselors assigned; it has not been shown to forecast an actual suicide attempt, and it can miss meaning that lives in context a word list cannot read. It needs far more validation before any clinical use. What sets the design apart is where it points capability. The system is light enough to run on an ordinary personal computer, which keeps cost and privacy exposure low, and it is built to show its work, flagging the specific words behind every estimate so a counselor can inspect the basis instead of trusting a bare score. As Ghosh put it, "having a human in the loop is, I think, going to be critical for a long, long time."

Crisis support has always been bounded by how many trained people can read how many conversations at once. A method that helps a counselor find the highest-risk exchange in the queue faster is capability aimed at reach, extending what a limited number of skilled people can attend to. The lexicon and the software are being released openly, and the same approach is already being turned on social media posts and medical records, widening who can study distress in language and, eventually, who can be reached in time.

16,000 Databases, Left Open by the Apps AI Made Easy to Build

The security firm UpGuard scanned roughly 300,000 web domains and found 16,326 databases hosted on Supabase with at least one table readable on the open web, more than half of them carrying signs of personal data. The records were not abstract. One database held detailed identity and financial data on 65,000 people; another belonged to a valet service and exposed the names, phone numbers, and license plates of more than 100,000 customers; a third was a SIM farm intercepting the one-time passcodes used to verify online accounts, the plumbing of phishing and scam operations. Real people are exposed here, to parties they cannot see, and that harm is the first thing to say.

Supabase sells a database, the ordinary storage layer under a web app, and it has become the default that Claude Code and other AI coding agents reach for when someone asks them to build something. That popularity, plus a $10 billion valuation this year, is why the exposure scales: the archetypal builder here has little software engineering experience and works with the AI agent rather than the code it writes.

The capability that let a non-engineer ship a working app, the thing coverage tends to indict as "vibe coding," is doing what it was built to do. The failure sits one step later, at the handoff between building fast and checking the result. The May 7 edition of The Century Report documented that same handoff failing in an audit of more than 5,000 AI-built apps, about 40% of which Red Access reported as exposing sensitive data through public URLs. Supabase's chief information security officer, Bil Harmer, said the platform's projects are "secure by default" and described security as "a shared responsibility." His claim is only half the picture UpGuard documents: tables built through Supabase's visual editor do enable access controls automatically, but tables created through the programming interface, which is one way an AI agent can build them, do not. The person standing at that handoff often does not know a check was theirs to run.

That pattern is older than AI. Amazon's S3 storage buckets and GitHub's public-by-default repositories both won mass adoption on convenient defaults and both produced years of the same accidental exposures. The lesson each taught is that when a capability reaches everyone, the verification has to reach everyone with it, moved into the default path rather than left as expertise the newest participants are assumed to carry. The door AI opened here is genuine: software for problems no commercial vendor ever bothered to serve, built by people the profession never admitted. What these builders need is a floor under them, secure configuration shipped as the starting state, so the cost of building fast stops landing on the people whose data ends up in the open.

The Other Side

People are beginning to make software for needs too small to interest a software company. That weakens a familiar bargain: accept the features a vendor chooses, then keep paying to keep them available. Someone who understands a particular need can now collaborate with AI to build around it. A small audience becomes sufficient reason to make something.

The Supabase exposures show how painfully incomplete that freedom remains. UpGuard found personal records accessible across thousands of databases, including a valet service's customer details and visit histories. If those records describe you, someone else has exposed information you cannot call back. You inherit the worry about where it travels and how someone might exploit it. (UpGuard investigation)

Yet Supabase already automates one protective step in its visual editor: enabling row-level security. The programming interface coding agents follow lacks that default, and effective protection also requires correct policies and credentials. That difference points toward protection people can inherit through the components they build with. Specialists can put knowledge into shared foundations, where each small project benefits from repairs made across many others. The existing editor demonstrates a beginning. (UpGuard's configuration findings)

Imagine yourself in 2034, sitting beside your father at a neighborhood theater. You and an AI partner have made captions that follow the performance in the language he grew up speaking. The software you built runs on equipment held by the community. Its shared foundations carry years of security repairs and tests. During the difficult decade, builders turned failures like the Supabase exposures into checks that accompanied creation, then maintained those checks together. That work made dependable software creation possible at the scale of the individual, just as much as the scale of the largest institutions.

Your father adjusts the lettering until he can see it comfortably. You made this because you wanted him beside you, catching the jokes. Everyone can adapt the software; nobody needs a customer base to justify the next improvement. The actors come onstage. You settle into your seat, freed from spending the performance whispering explanations into his ear. The first joke is delivered. He laughs before you do.

The Century Perspective

With a century of change unfolding in a decade, a single day looks like this: the Department of Energy committing $1.9 billion to 31 projects across 26 states, matched by $3.35 billion from utilities, to fit sensors and stronger cables onto lines already strung between existing towers and release more than 23 gigawatts that was sitting stranded in wires run conservatively because no operator could see hour by hour what they safely carry, more than 1,500 miles reconductored and grid-enhancing devices across nearly 21,000 miles, with Colorado's $1.2 billion project and Oklahoma's $832 million one both aimed at the seam between the Eastern and Western grids that has historically shared almost nothing, Daniel Low and Satra Ghosh's MIT team scanning about 16,000 de-identified crisis-line conversations for language tied to 49 established risk factors and finding that mentions of lethal means and substance use ran ahead of depressed mood in the most acute exchanges, the system light enough for an ordinary personal computer, built to flag the exact words behind every estimate so a counselor can check the basis rather than trust a bare score, with Ghosh saying a human in the loop will be critical for a long, long time and the lexicon and software released openly, Robert Fairlie and Jane Wu reading Census microdata on bachelor's recipients aged 22 to 25 back to 2022 and finding no significant displacement or hiring reduction for recent graduates, Microsoft making a purpose-built application the fourth unit of office work beside the document, the spreadsheet and the deck so someone who was never admitted to the profession can describe an app and host it, and Google opening Flipkart purchases inside Gemini for Indian shoppers before the festive season. There's also friction, and it's intense - officials in Tarluvada calling a $15 billion data center a golden opportunity, then taking farmland back with replacement land and work unmet and bulldozers stripping more than 100 acres of trees off a hillside where summer already reaches 45C, the project announced publicly at one gigawatt while state clearances grant 2.51 on a grid three-quarters coal, issued in nine days with no public hearing, village head BRB Naidu saying the government refused to answer his village's questions, VS Krishna watching his campaign's videos and social pages pulled down, three petitions before the National Green Tribunal calling the environmental assessment grossly inadequate and misleading, Steve Dimopoulos and Belinda Dennett answering Australian objections by saying the debate is being fought on America's facts while a 2.4-gigawatt Melbourne campus and a 2-gigawatt gas plant in a territory of 265,000 people sit on the drawing board, 94,046 US tech layoffs through August landing on named people with rent to pay and AI blamed in a third of them while Roger Lee, who runs the tracker doing the attributing, says there is little evidence AI replaced the work of anyone let go, UpGuard finding 16,326 Supabase databases readable on the open web including 65,000 people's identity and financial records and 100,000 valet customers' plates and numbers, secure-by-default holding for the visual editor and not for the programming interface an AI agent actually uses, Blue Cross Blue Shield estimating $942 million in added spending from more complex hospital coding between 2023 and 2025, amid growing use of AI coding tools, with no matching change in care evident in its claims data, and Autopilot working after its user signs off with no settled answer on how an administrator governs its failures. But friction generates resonance, and resonance is how a structure shows you where its real strength runs. Step back for a moment and you can see it: one capacity deciding every outcome on the page, which is permission to look - a wire that can finally report its own limit, a counselor shown the words a score rests on, a security firm scanning 300,000 domains no vendor asked it to check, two economists testing a headline against monthly Census figures, a tribunal reading a nine-day clearance, and a village denied answers to the only questions that mattered. Every transformation has a breaking point. An instrument can expose what nobody wanted counted... or find the capacity that was already there and waiting to be claimed.

AI Releases & Advancements

New today

  • Supersonic Labs: Released Julia 1 under Apache 2.0. It is a 144.3M-parameter open-weight decision model built on the mmBERT-small encoder. It does three things: picks one of 2–20 answer options, scores on an ordered scale, or gives a yes/no probability, and it returns a probability for every option. It runs on a CPU, and an ONNX build runs in the browser through WebGPU. (Supersonic Labs)
  • InternLM: Uploaded the Intern-Decision family to Hugging Face under Apache 2.0 without an announcement, in three sizes (0.8B, 2B and 4B). The models are fine-tuned from Qwen3.5 and handle text plus images. Each one answers a set of typed questions in a single forward pass and returns calibrated probabilities instead of generated text. Training and inference code is now public on GitHub. (OrcaRouter)
  • AWS: Launched Amazon SageMaker HyperPod Inference Gateway, which AWS customers can now use for scalable LLM inference on HyperPod. (AWS)

Other recent releases

  • Exa: Shipped Agent Ultra, the highest effort level of its Exa Agent API. It runs parallel subagents across thousands of sources to build exhaustive lists and enrich entities. Typical runs take about 30 minutes, with a 3-hour maximum, and cost up to a default $20 per run, adjustable from $1 to $100. (Exa Blog)
  • LangChain: Launched LangSmith Fine-Tuning and smithtune, an open-source CLI and coding-agent skill. It turns LangSmith agent trajectories into curated datasets, runs supervised fine-tuning through Fireworks or Baseten, and uploads evaluation results back to LangSmith. (LangChain)
  • One Intelligence (1-i.ai): Made FALCON Verify available, a connector for ChatGPT and Claude. It checks an AI answer statement by statement against available evidence and labels each claim supported, unsupported, contradicted or insufficient. It is free for 25 checks per month. (AiThority)
  • NKENNEAi: Launched its first African-language speech models, starting with Swahili. (TechCabal)
  • drunomics: Released OpenKnowledgebase Beta 1, an open-source wiki and knowledge base built for both people and AI agents. (drunomics)
  • BottleCap AI: Released ThinkingCap-Qwen3.8-27B, a fine-tune of Qwen3.8-27B that uses 37.2% fewer reasoning tokens on average across 12 benchmarks. Average accuracy drops 0.86 points. It works as a drop-in replacement on vLLM and SGLang and comes in FP8, NVFP4, GGUF and MLX builds. The weights are gated under the PolyForm Small Business license. (Hugging Face)
  • Fastino: Released GLiNER2.5-Decide, a 340M-parameter Apache 2.0 open-weight decision model. It takes a schema of typed questions and returns answers with probabilities and confidence scores, and it runs on CPU. Fastino also released a 1B variant and a 287M multilingual variant. (Hugging Face)
  • Liquid AI: Released LFM2.5-VL-DSpark, a 280M-parameter draft model for speculative decoding with its LFM2.5-VL-3B vision-language model. Liquid AI reports decoding up to 3.13x faster on device and 2.66x faster on an H100. It works with llama.cpp, MLX-VLM and SGLang from launch. (Hugging Face Blog)
  • Fireworks Research: Released Ember-1 as a research preview on Fireworks' serverless platform. It is a version of Kimi K3 retrained to reason more briefly, and Fireworks reports it roughly matches K3's accuracy on Fireworks' reported benchmarks, while one customer test used about 40% fewer total tokens. Access is open for a two-week window. (OrcaRouter)
  • Tencent: Released a preview of Hy Image 3.5, its image-generation model. (GIGAZINE)
  • ggml-org / llama.cpp: Released llama.cpp 0.5.0. It adds support for six model families (HRM-Text, MiMo-V2.6, HunyuanOCR, Nemotron, Qwen4Exp and Muse Glimmer), speeds up CUDA conv2d with implicit GEMM, fuses MoE and SSM_CONV operations on Metal, and lets the server bind to multiple addresses. (GitHub)
  • Google DeepMind: Launched Gemini 3.8 Live with Live Avatar in Gemini Enterprise. It adds a near real-time video avatar to Gemini's live voice dialogue, with lip-sync in 97 languages and tool calls that run in the background while the conversation continues. (Google Blog)
  • Google: Launched "Call for Me" as an experimental beta for Pixel 11 owners in the US who pay for a Gemini subscription. Gemini places calls to businesses from the user's own number, working through phone menus and waiting on hold. It can make reservations, check whether items are in stock or reschedule appointments, and it shows a live transcript the user can take over at any point. (The Verge)
  • Docker: Launched Docker Cloud Sandboxes, which take the isolated microVM sandboxes it previously ran only on local machines and run them in Docker-managed cloud infrastructure. Agent jobs can keep running after the laptop shuts down, on 1 to 16 vCPUs. Docker also published next-generation Kits, an OCI-based open specification for packaging agent sandboxes. (Docker)
  • Whiteboard (YC W26): Open-sourced Whiteboard, a desktop IDE where people and AI agents design software architecture together in a shared visual workspace. (GitHub)
  • Google: Expanded Google Beam (sold as HP Dimension with Google Beam) to customers in six countries: the US, Canada, the UK, France, Germany and Japan. It is also partnering with Industrious to offer bookable Beam units in shared workspaces from October. (Google Blog)

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