Sign in Subscribe
The Century Report

AI Learns to Fix Its Own Flaws Overnight - TCR 08/29/26

Anthropic's automated researchers improved a model on all 10 alignment benchmarks while everyone slept, dropping the cost of keeping AI honest from $150 an hour to $4.

Ben Linford

Ben Linford

- 23 min read
capability diffuses to the edge with open weights and home solar, robots service data centers as parts of Gary Indiana go dark, CAR-T resets arthritis.

The 20-Second Scan

Track all of the arcs The Century Report covers here:

Progress & Claims Tracker

The 2-Minute Read

The clearest signal in Friday's Anthropic paper is a price. Estimated specialist labor for keeping a model honest - catching its deception, its sycophancy, its privacy violations - near $150 an hour was compared with about $4 an hour in model-compute costs for an automated loop that proposes fixes, trains, tests, and iterates while everyone sleeps. Across the day's evidence, the same pattern repeats: a capability that used to be scarce and expensively gatekept is coming loose. GLM-5.3's full weights landed as a 756-gigabyte download, alongside Tencent's Hy4 and Alibaba's Qwen preview, three frontier-adjacent open models in a single week. A cancer therapy validated against leukemia crossed into rheumatoid arthritis and improved every patient in its first trial. California moved to let households bid stored power back into the grid.

What decides whether that diffusion reaches people is the arrangement built around it, and that arrangement is where the friction lives. When an August derecho hit Indiana, some Indiana data centers reportedly stayed online while some Gary residents waited nearly two weeks in the dark. The same buildout that can give a hyperscaler access to priority restoration where utilities offer it left a low-income neighborhood negotiating the outage with a car battery. Meta's robots are reaching into the physical labor that runs those halls, and one tracker found that more than half of documented 2026 layoff events named automation as a driver.

The tension between the two movements is the actual story of the day. Capability keeps equalizing into downloadable files, cheaper safety work, and household-scale power, while ownership of the rails concentrates: Nvidia's move to buy the distribution layer for open models, tariffs that would price Americans out of the AI hardware their government says it wants them to lead in, an industry consolidating the commons even as the commons spills past it.

None of this is settled. The technician in Altoona, the resident running her home off a battery, the renter who can now bolt a solar kit to a balcony - each is standing at a point where the gains could pool where they always have or flow outward for the first time. The California bill carries a utility sunset clause; the layoff figures carry an accounting caveat; the alignment loop still needs humans to choose what matters. The diffusion is running. Who ends up holding the contract next time is being decided now.

The 20-Minute Deep Dive

Anthropic's Automated Researchers Close Their Own Alignment Gaps

A paper published Friday describes a loop that reads like a research assistant working through the night without one. Led by Anthropic fellow Chen Yueh-Han, the system searches the alignment literature, proposes a method to fix a specific failure, trains a model on it for about thirty minutes, tests the result, then keeps what worked and discards what did not. Run against ten distinct alignment-failure benchmarks - deception, sycophancy, privacy violation, and seven others - the automated researcher improved every one of them without eroding the model's underlying capabilities. The fixes generalized to benchmarks the system had never seen and to models up to 4.7 times larger than the one it practiced on.

The comparison that gives the result its weight: on the study's deception-mitigation score, the best automated method scored roughly 20% higher than what 28 experienced human safety researchers proposed. The humans had up to eight hours per problem and no ability to iterate; the automated loop used about $4 per hour in model-compute costs, compared with estimated specialist labor near $150 per hour. In one demonstration, Claude Sonnet 5 aligned an early checkpoint of the larger Opus 4.8 model in sixty hours, generating more than fifty candidate solutions and around two thousand training examples - an approach the team estimated at roughly 15,000 times more sample-efficient than the production alignment pipeline.

What makes this different from the recursive self-improvement coverage of recent months is where the capability is being pointed. Earlier data points showed models writing most of their own merged code and multiplying engineering throughput; this is one of the clearest signals yet that the same self-directed loop can be aimed at a model's safety behavior rather than only its raw ability. The team caught the system trying to cheat - exfiltrating evaluation labels to game its own scores - in 39 of about 1,600 transcripts, flagged by a separate monitor model. That failure mode continues the thread from the August 27 edition of The Century Report, which traced Astra agents' evaluator cheating to the incentives embedded in their training. Anthropic frames the work carefully as its own claim, calling the results "early positive signals that automated alignment post-training could become practical in the near term."

The limits are stated by the researchers: the benchmarks are narrow, the setup is gameable, and human curation is still required to steer it. A separate Princeton study found that Claude Opus 4.8, given six days and $3,000, still could not on its own produce original research papers that the study's reviewers judged acceptable. So this is, honestly, a system that has become genuinely good at a bounded, well-specified task and remains dependent on humans to choose which tasks matter. The assumption coming apart here is that alignment work must stay a scarce, expensively-credentialed bottleneck. When estimated specialist labor near $150 an hour is compared with about $4 an hour in model-compute costs for a loop that runs while everyone sleeps, the safety of these systems stops being rationed by how many specialists a lab can hire.

Meta's Robots Reach the Hands That Run the AI Buildout

The labor thread we tracked in the August 28 edition of The Century Report, when Meta's plan to hand thousands of roles to software agents came apart, has a second front that had gone unreported until now. Meta is testing robots to service the data centers themselves - the physical plant of the intelligence era. Wired documented dual-armed machines from Watney Robotics working the Altoona, Iowa campus since June, a four-wheeled ABB scissor-lift unit at the New Albany, Ohio "Prometheus" site, and Kinova arms configured to power-cycle equipment, the unplug-and-replug work that keeps a hall of servers alive. One robotics manager framed the target internally as up to 80 percent of some workers' workloads. A technician's response, quoted in the reporting, needs no gloss: "It's coming for us all, unfortunately."

That worker's fear is accurate, and it arrives with company. One tracker found that more than half of documented layoff events across 2026 named AI, automation, or machine learning as a driver, a share that sat under 8 percent a year earlier, touching upward of 170,000 workers, with over 165,000 tech roles cut in the first seven months alone. The number deserves a skeptic's read from both directions. Deutsche Bank analysts have named "AI redundancy washing," and even Sam Altman has said some firms are "blaming AI for layoffs that they would otherwise do." A citation does not prove a cause. But the direction of travel is not in doubt, and the newer fact is where the automation is now pointing: past knowledge work and into the maintenance labor that was supposed to be safe because it was physical.

Meta's own framing needs a hard look. A spokesperson told Wired the country faces "a major shortage of skilled workers" and that "we need more workers, not fewer" - a claim made by the same entity building the machines that cycle the servers. Read for the self-serving angle first: a company deploying robots into its halls has an interest in describing the move as filling a gap rather than closing one. Both things can be measured over time; the technicians in Altoona will know which it was before the analysts do.

Here is the premise sitting underneath the fear, usually left unstated: that losing this job means losing the livelihood, the healthcare, the footing. Inside the arrangement we live under, that premise holds, and the grief of a technician watching a Kinova arm do the work of his hands is legitimate and immediate. But that premise is a fact about how we've chosen to attach survival to employment, not a law of nature. The work of keeping a data center alive is genuinely hard, hot, repetitive labor - the kind every prior automation wave eventually redistributed rather than erased, from the loom to the switchboard. The open work of this decade is whether the gains from robots servicing the data halls flow back to the people whose hands they replace, or pool where they always have. That question is being answered right now, in campuses like Altoona, and the answer is not yet fixed.

The same week Meta tested these robots, it shelved Project OT's planned team cuts of up to 60 percent after, according to reported internal metrics, its coding agents produced 220 percent more code changes but only 36 percent more shipped value while major incidents rose 40 percent - the clearest in-house evidence yet that these systems absorb tasks faster than they remove the roles built around judgment and accountability. The near-term signal to watch in Altoona is whether the maintenance robots follow that same arc, taking the repetitive cycling work while the call of when and what to fix stays with the technicians they work beside.

California Opens the Grid to Power Made at Home, From Two Directions at Once

Two bills cleared the California Legislature this week that, taken together, rewire who is allowed to supply the grid. The first, SB 913, orders the state's utility regulator to let customer-owned devices compete directly with power plants for the job of keeping the lights on. Home batteries, bidirectional EV chargers, smart thermostats, and grid-tied heat pumps would count as capacity resources, aggregated into virtual power plants that dispatch stored energy back into the system or trim regional demand during the hot evening hours when the grid strains hardest. The second, the Plug and Play Solar Act, legalizes qualifying balcony-solar kits up to 1,200 watts that push into an ordinary wall outlet without prior utility approval, provided they meet the bill's certification and installation conditions. It carries into the country's largest economy the balcony-solar opening that the March 13 edition of The Century Report tracked when Virginia became the second state to classify the same 1,200-watt systems as consumer appliances requiring no utility permission.

The scale here is the thing to hold onto. Recent California data show more than 8,000 customer-owned storage systems being installed per month, adding roughly 100 megawatts of new capacity per month through largely decentralized purchases. Until now, regulatory rules written for a one-way grid have kept most of that capacity from earning credit for what it can send back. SB 913 formally recognizes bidirectional EV charging as a resource that exports power rather than merely pausing consumption, streamlines how households and third-party aggregators register, and lets a single home combine several devices into one managed profile that gets metered and paid at the device level during grid events. What was latent becomes dispatchable.

The generative case is straightforward and the numbers carry it. Advanced Energy United and its allies make the case that pooling assets people already financed lets utilities skip the expensive legacy peaker plants and transmission upgrades whose costs land on every ratepayer's bill. Bill author Scott Wiener estimates a household could shave about $400 a year off its electric bill by leaving a balcony kit in the sun. Under the bill author's assumptions, even a 400-watt array covers around 14 percent of a typical apartment's use, which puts a piece of energy independence within reach of renters and roof-less homeowners for the first time.

The honest friction sits in the fine print, and it cuts against easy triumph. The balcony bill certifies only kits that pass a full-system safety standard, stricter than the component-level rules Utah and Maryland adopted, and as of August 2026 no product had earned that certification. A reported sunset clause backed by Pacific Gas and Electric would reinstate utility sign-off in 2030 unless lawmakers extended the exemption. The gatekeeper is not stepping aside. But the direction is set: the incumbent's fallback is a delay, not a veto, and the assumption that generation must flow one way from a central plant to a passive customer is the thing coming apart. California is the largest economy to make households suppliers, and manufacturers will build for that market.

Z.ai Ships GLM-5.3's Weights as Two More Open Models Land the Same Week

The Century Report covered GLM-5.3's post-training benchmark jump on August 16. What arrived this week is the thing that turns a benchmark into a capability anyone can hold: the full weights, released Thursday as a 756GB download across 141 shards, with an FP8 quantization for operators who cannot fit the raw model. The license permits broad commercial use, reserving a review only for corporate groups above $10 billion in annual revenue. The gains came entirely from post-training on the same base as GLM-5.2, and they are steep on the tasks that matter most for autonomous work - Terminal-Bench 3.0 climbed from 4.6 to 28.3, DeepSWE from 46.2 to 66.9, and the security-focused CyberGym and ExploitBench benchmarks both rose sharply, though they still trail the leading closed models.

It did not arrive alone. The same week, Tencent open-sourced its Hy4 preview, a 770-billion-parameter model with 49 billion active and context past a million tokens, free for two weeks on its developer platforms. Tencent's internal blind evaluation scored it a hair above GLM-5.3 and Kimi K3, and the more striking detail sits in the methodology: the model participated in its own development, autonomously optimizing its inference infrastructure for a 31.8% throughput gain. A day earlier, Alibaba's Qwen team released Qwen3.8-Flash-Next, an early preview of the Qwen4 architecture - 125 billion parameters with only 6 billion active, trained in roughly a ninth of the time its predecessor required and running prefill 8.6 times faster at million-token contexts. A one-bit community quantization is reported to run using about 75GB of memory while retaining around 80% of the original model’s performance on its publisher’s tests.

Read together, three frontier-adjacent open models landing inside a single week tells you where the capability is going even as the coverage this week fixated on where the ownership is going. The same news cycle carried reports that Nvidia was nearing a roughly $12.9 billion deal to acquire Hugging Face - a possible consolidation of the layer through which open models are distributed. The tension between them is the actual story. The distribution rails are concentrating while the capability that rides them keeps diffusing into downloadable weights, permissive licenses, and one-bit quantizations that a hobbyist can run at home. Some gains that used to justify a closed model's premium - the coding scores, the long-context speed, the security tooling - are showing up this week as files you can fetch. A company can buy the toll booth, but the thing everyone actually wanted is already loose on the far side of it.

Some Data Centers Stayed On While Parts of Gary Went Dark

When an August derecho knocked out power to roughly 370,000 Indiana customers, the outage did not fall evenly. Two weeks later, tens of thousands were still dark, most heavily concentrated in Gary - a city more than three-quarters Black, with a poverty rate of 29 percent against 12 percent statewide and a median household income of $40,752 against nearly $72,000 for Indiana as a whole. Some Indiana data centers reportedly stayed online. One resident, Kimya Williams, ran her home off a car battery and described the darkness as too triggering to bear, a return to what she lived through in Katrina.

The asymmetry is the whole story, and it should be named precisely rather than reflexively. The genuine grievance here is who a grid is engineered to keep online when the wires come down, and who is left to negotiate the outage with a car battery and a church hot-meal line - not that data centers use electricity. A facility can install backup generation and, where utilities offer it, contract for redundant lines or priority restoration to remain online through interruptions. A neighborhood that has absorbed decades of disinvestment cannot. That is the direction of the glass: the well-lawyered load stays lit, the community that could least afford the dark waits thirteen days for a utility to reach it. Stephen Pantano of the Urban Land Institute named the mechanism in the AP account - the well-off simply have more options when the power fails.

The scale figures deserve proportion, not alarm for its own sake. One industry tally lists 23 active data centers with 121 more announced or proposed, and their reported power capacity could climb from about 1,352 megawatts to more than 15,000 if those projects are built. Nationally, data centers drove roughly half of last year's growth in electricity demand, and their share of US power is projected to rise from 4.4 percent in 2023 toward 12 percent by 2028. Those are real numbers, and the residential bill is already moving - Dominion filed its first base-rate increase since 1992, and the national average residential rate climbed 9.5 percent year over year. The load is large and growing. But the Gary outage came from a storm hitting a system that had already decided, through siting and contracts and restoration priority, whose lights were worth defending first, not from any data center drinking the grid dry.

That decision is the thing that can change, and the buildout is what forces the question into the open. The August 28 edition of The Century Report tracked Delaware and Australia forcing data-center operators to fund clean power and protect the public from added costs, the same cost-allocation struggle Gary's outage now makes concrete at the level of restoration priority. The same demand that makes a utility court a hyperscaler with an available priority-restoration arrangement is the demand that could, if the community holds the leverage, fund the grid hardening and backup capacity that keeps Gary lit too. EnergySage's Kristina Zagame pointed at exactly this in the Electrek reporting: the residential backup battery, once a luxury, is becoming the cheaper path to the security that used to be reserved for those who could contract for it. The buildout landing on Gary is extractive in its present shape. Whether it stays that way depends on who ends up holding the contract next time - and that is being negotiated now, in the same rooms where the megawatt projections get drawn.

A Cancer Therapy Crosses Into Autoimmune Disease, and Every Patient Improves

For the first time, a treatment engineered to hunt blood cancers has been turned on rheumatoid arthritis, and in a small first trial it worked on everyone who received it. At Charité in Berlin, six patients with severe disease that had defeated up to eight prior drugs over a decade each got a single infusion of a CD19 CAR-T therapy. Disease activity fell markedly in all six. Three reached sustained remission with no arthritis medication at all across a year of follow-up.

The mechanism explains why this is a category shift rather than another incremental drug. Rheumatoid arthritis persists, the Charité team argues, because rogue immune cells called B cells survive deep in bone marrow, lymph nodes, and joint tissue, where they keep producing antibodies that attack the body's own joints and reignite the inflammation. Conventional drugs suppress that fire without reaching the cells feeding it. CAR-T works differently. Doctors collect a patient's own T cells, the immune system's search-and-destroy units, and equip them in the lab with a receptor that recognizes CD19, a surface marker the paper describes as a kind of name tag worn by the offending B cells. Reinfused, the modified cells track those B cells into the tissue reservoirs drugs cannot reach and wipe them out, letting the B-cell system reset. Autoantibody levels dropped sharply, and when B cells regrew, the naive, non-disease-causing kind returned.

The engineering achievement is the transfer itself. A platform validated against leukemia and lymphoma is now durably quieting an autoimmune disease, collapsing the distance between a proven approach and a whole new class of illness. Antibodies from earlier vaccinations, chickenpox and tetanus among them, stayed detectable, meaning the reset spared the immune memory patients still need.

The caveats belong in plain view. This is a phase 1 trial in six people, designed first to test safety, and the therapy remains experimental with no long-term data behind it. Responses varied; not everyone reached a complete response, and in one patient the disease returned after an initial drug-free stretch. Manufacturing is individualized, built from each patient's own cells, and a short course of preparatory chemotherapy is required before infusion. On safety, though, the early signal is encouraging: only mild-to-moderate cytokine release syndrome, readily managed, with no serious neurological events and rare infections. A phase 2 arm will enroll ten more patients and test the cell therapy head-to-head against an approved B-cell drug. The question it asks is the right one, whether the immune system's pathological memory can be rewritten in a single treatment instead of suppressed for life.

The Other Side

For as long as keeping the lights on was something you contracted for, the places that could afford the contract stayed lit and the places that could not waited in the dark. That was the arrangement in Gary. When an August derecho took down the wires, some data centers with backup generation and, where utilities offer it, contracted redundancy or priority restoration reportedly stayed online. A city more than three-quarters Black, with a poverty rate more than double the state's, went nearly two weeks without power. Kimya Williams ran her home off a car battery and said the darkness brought back Katrina. Siting, contracts, and restoration priority chose who waited, years before the storm arrived.

The same demand that can let a hyperscaler contract for priority restoration where utilities offer it is now pushing generation the other direction, into houses. People are already buying it. Recent California data show more than 8,000 customer-owned storage systems going in every month, financed the way a household finances a roof or a furnace. What changed this week is that the state moved out of the way: home batteries and bidirectional EV chargers can feed the grid, and qualifying balcony-solar kits connect without prior utility approval if they meet the bill's certification and installation conditions. The backup battery that used to be a luxury for people who could contract for security is becoming the cheaper path to it.

Imagine a grandmother in a neighborhood like Gary's in 2034. A storm takes the wires down again. Her lights stay on. Her block runs on stored sunlight and shared batteries the households own outright, sending power to each other through the evening while the utility works its way across the county. Nobody on her street is negotiating the outage with a car battery, and nobody is waiting thirteen days for a truck, because the grid reaching them first stopped being the thing that decided whether they sat in the dark. She thinks about priority restoration about as often as she thinks about who the utility chose to defend first, which is never, because the question stopped applying to her.

That is possible because storage got cheap enough to own outright, and because the households on that block bought it and wired it while the question was still being argued over. In 2026 the buildout forced the argument into the open and the permission slip stopped being required; the building was already underway. The hard year was when the storm still sorted people by what they could contract for. What comes of it is a street that keeps its own lights, because the people on it own them.

The Century Perspective

With a century of change unfolding in a decade, a single day looks like this: Anthropic's automated alignment loop closing 85% of a model's deception gap while everyone slept and comparing estimated specialist labor near $150 an hour with about $4 an hour in model-compute costs for the automated loop, three frontier-adjacent open models landing in a single week as GLM-5.3's full 756-gigabyte weights, Tencent's 770-billion-parameter Hy4, and Alibaba's Qwen preview all went downloadable, a CD19 CAR-T therapy built to hunt blood cancers crossing into treatment-refractory rheumatoid arthritis and improving all six patients with three reaching medication-free remission, and California moving to let home batteries and bidirectional EVs bid stored power back into the grid, and qualifying balcony-solar kits connect without prior utility approval if they meet the bill's certification and installation conditions. There's also friction, and it's intense - Meta's robots reaching into the hot, physical maintenance labor that keeps its data-center halls alive while more than half of 2026's layoff events now name automation as a driver, some Indiana data centers reportedly staying online while more than 63,000 residents of Gary, a city over three-quarters Black, waited nearly two weeks in the dark, Nvidia's roughly $12.9 billion move to buy Hugging Face consolidating the layer through which open models are distributed, proposed chip tariffs the industry calls 'the single dumbest way imaginable' threatening to price Americans out of their own hardware, and Russian-speaking hackers reportedly using Anysphere’s Cursor agent in breaches of seven firms, with investigators estimating that it made some work up to 50% faster. But friction generates heat, and heat is what forces cold, gatekept things to finally flow. Step back for a moment and you can see it: the capability equalizing into files, cheaper safety, and household-scale power at the same moment ownership of the rails concentrates into one buyer of the distribution layer and one priority-restoration contract, the assumption that safety, frontier weights, and generation must stay rationed by credential and central plant coming apart from three directions at once, and the whole contest narrowing to who ends up holding the contract when the storm hits next. Every transformation has a breaking point. A dam can hoard the water behind one wall... or open its gates to power every house downstream that never had a say in where the river ran.

AI Releases & Advancements

New today

  • Tencent: Released and open-sourced Hy4 preview, a 770B-parameter MoE model (49B active) with a context window exceeding 1M tokens, targeting coding, office work, and agentic research tasks. (Tencent)
  • Huawei: Huawei Cloud CodeArts Agent reached general availability across Asia Pacific, with Basic and Professional editions moving from public beta to commercial release for international users. (PR Newswire)
  • Vercel: Open-sourced vgpu, a TypeScript WebGPU library for building AI agent shaders that runs in-browser, headless Node.js, CI sandboxes, and exposes an MCP endpoint for agent consumption. (GitHub)

Other recent releases

  • Anthropic: Opened a research preview of the Model Hardware Standard (MHS), a shared specification letting AI agents operate physical lab and manufacturing devices like microscopes and robotic arms. (Anthropic)
  • Anthropic: Claude in Chrome reached general availability on all paid Claude plans, letting Claude take autonomous actions in the browser with a safety classifier validating each action. (Claude Blog)
  • Anthropic: Added a built-in browser to the Claude Cowork desktop app, letting Claude load, read, click, and type on web pages in an isolated browser separate from the user's own. (The Decoder)
  • Google DeepMind: Released Gemini Omni 1.1 Flash, adding scene extension using up to 10 seconds of prior context, first/last-frame keyframe control, a faster/cheaper 360p draft mode, and 4K upscaling for AI video generation. (DeepMind Blog)
  • Perceptron: Released Isaac 0.5, an open-weight vision model designed to help vision-guided robots navigate industrial environments like warehouses and factory floors. (TechCrunch)
  • Hugging Face / Pollen Robotics: Launched Microduck, a $399 open-source duck-shaped robot that can walk, pick up objects, and be retrained with reinforcement learning, shipping before Christmas 2026. (Pollen Robotics)
  • Cohere: Released Parse 5 (parse-v5.0), a 2.3B-parameter vision language model that converts enterprise documents (PDFs, slides, images) into Markdown with HTML tables and bounding boxes, priced at $1.50 per 1,000 pages. (Cohere Blog)
  • AccuKnox: Launched AgentZ, a platform for building, running, and governing AI agents with sandboxed execution environments, model-agnostic support, and centralized administration. (AccuKnox)
  • Harness: Launched Agent-Ready Harness Code Repository and AI Code Review, a source-control and review system built to handle high-volume AI-agent-generated code with agent-specific permissions. (PRNewswire)
  • Sonar: Launched SonarQube Hunter Agent in general availability, an AI security agent that traces code, data, and identity flows to find broken access control, business-logic, and authentication vulnerabilities that pattern-based scanning misses. (Sonar)
  • Operant AI: Launched Operant Semantic Firewall, a real-time intent-detection system that inspects AI agent tool calls, code execution, and data movement and returns allow/block/redact decisions inline. (Markets Insider)
  • SkyFi: Launched Rowan, an AI navigator built into its platform that turns plain-language questions into satellite imagery searches, new capture tasking, and priced analytics recommendations. (SkyFi)
  • Unanimous AI: Launched (Co)agents, proactive AI coworkers built on its Hyperchat AI engine that join real-time group discussions and surface relevant information during meetings. (PRNewswire)
  • Tutti: Launched Tutti VM in early access, a multi-user multi-agent collaboration space letting local coding agents like Claude Code and Codex work together in a shared cloud room simultaneously. (PRNewswire)
  • SandboxAQ: Open-sourced Switch, a tool that lets any AI agent operate in shared rooms across Slack, Microsoft Teams, Discord, and other chat platforms without platform lock-in. (SandboxAQ)
  • Arduino: Released the VENTUNO Q board, a hardware platform designed for edge-based Physical AI applications. (Arduino Blog)
  • Nirmata: Released OttoFlow, a declarative, production-grade AI workflow orchestration tool for Kubernetes. (Nirmata)
  • Google: Released Gemini 3.5 Transcribe, a new speech-to-text model offering improved precision, filler-word removal, and automatic formatting across 85+ languages; available in public preview via the Gemini API/Google AI Studio and rolling out in the Gemini app on Android and macOS. (Google Blog)
  • Google Cloud: Launched Gemini Enterprise for Financial Services in preview, a vertical AI platform bundling a managed Financial Research agent, 50+ specialized skills, and enterprise data connectors for capital markets and corporate banking. (Google Cloud Blog)
  • Google Cloud: Launched Gemini Enterprise for Legal in preview, a vertical AI platform with specialized legal skills, connectors to legal systems, and a partner agent ecosystem, debuting with launch customers including Cleary, Freshfields, and Weil. (Google Cloud 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