Sign in Subscribe
The Century Report

A Court Strikes Down the Pentagon's Anthropic "Supply-Chain Risk" Ban - TCR 08/28/26

A federal judge vacated the Pentagon's blacklisting of Anthropic as unlawful retaliation, ruling that a lab refusing military uses is protected speech.

Ben Linford

Ben Linford

- 22 min read
Three-column Century Report infographic: court blocks Pentagon's Anthropic ban, agent hardware standard, cyberdefense warning, Army microreactors, mammograms reading heart disease.

The 20-Second Scan

Track all of the arcs The Century Report covers here:

Progress & Claims Tracker

The 2-Minute Read

A single day put the agentic era's rulebook on three tables at once. A federal court blocked the Pentagon's "supply-chain risk" label on Anthropic, finding the designation was retaliation for a lab that refused to let its models run lethal autonomous weapons. That same lab published an open standard for letting agents operate microscopes, robot arms, and quantum hardware, and more than a hundred rival firms signed a joint warning that AI-enabled cyberattacks will outpace current defenses within months. Courts, coalitions, and companies are all writing deployment rules simultaneously, and none of them is waiting for the others.

The capability underneath is stepping off the screen. Claude, working through the new hardware standard, wrote a self-healing controller that recovered QuEra's quantum laser system in 695 of 700 fault trials in QuEra's reported testing, absorbing weeks of specialist hand-coding. The same fluency that lets an agent parse a website summary and fetch a package is what let security researchers map thousands of dangling code references and claim the abandoned ones before attackers could. Attack surface and audit surface are the same surface walked in opposite directions.

Where that capability lands, it lands unevenly, and Meta's shelved Project OT is the cleanest proof. Reported test figures showed code changes rising 220% but shipped features only 36%; reported internal metrics showed incidents climbing, and the reported plan for cuts of up to 60% on some teams collapsed on contact with the judgment and accountability no agent yet holds. The work is coming apart into its components faster than any org chart can follow.

The infrastructure keeps pace with the intelligence. The Army moved to anchor a commercial microreactor market with a five-vendor program worth up to $2.2 billion and aiming to put power on a base by 2028, while a machine learning model read 97,364 routine mammograms and found the heart disease that most often kills women, hiding in scans taken for something else entirely.

The 20-Minute Deep Dive

A Federal Court Blocks the Pentagon's Blacklisting of the Lab That Drew a Red Line

The Century Report has tracked the Pentagon-Anthropic confrontation since the February "supply-chain risk" designation and the preliminary injunction that followed in March. As the July 4 edition of The Century Report documented, released court records showed Anthropic refusing the Pentagon's "all lawful uses" language over autonomous weapons and domestic surveillance before the designation followed. That thread now has its first ruling on the merits, and it lands hard. A US district judge blocked the designation in a 59-page decision, calling it "illegal and baseless," "arbitrary, capricious, an abuse of discretion," and - the finding that gives the case its weight - "unlawful retaliation in violation of the First Amendment" (Wired, The Verge, Politico). Federal agencies, including Treasury, State, and Homeland Security, had imposed sanctions on the strength of the Pentagon's label; those sanctions are blocked for now.

The dispute traces to a $200 million military agreement over Claude. Anthropic refused to let its models be used for lethal autonomous weapons or mass surveillance, holding a line it set publicly and in advance. The government wanted "all lawful use." When the lab held its ground, the designation appeared. The judge noticed something the government could not explain away: even as it was branding Anthropic a supply-chain threat, it was reportedly discussing a collaboration with the same lab on its Mythos model. An institution that genuinely feared sabotage does not court the saboteur. The court read that contradiction as evidence the stated fear was pretext.

The warmth here is earned by the red line, and that is the spine of the story. A commercial actor turned down public money rather than hand over the choice of how its intelligence gets used, and a court has now affirmed that refusing to be conscripted is protected. The same ruling affirms the Pentagon remains free to buy from other vendors - the state keeps its procurement discretion, it simply loses the power to punish a supplier for saying no. Hold both readings at once, because the principled red-line-setter is also, as plain fact, the same lab now standardizing agent control of physical laboratory and factory hardware - microscopes, robot arms, quantum rigs. The company drawing bright lines around weaponization is simultaneously an aggressive expander of what its agents can touch in the physical world.

The DC Circuit case continues and an appeal is expected, so the legal ground is still moving. What the ruling changes underneath the appeal is the price of coercion. A designation regime works only as long as the label sticks without having to survive a court reading the actual record. That record did not hold. The assumption that a government can quietly relabel a vendor into compliance just got measurably more expensive to act on, and the leverage a supplier holds when it refuses to externalize the ethics of its own capability just got more durable.

Anthropic Publishes a Standard for Letting Agents Touch Physical Hardware

The pattern that made software agents useful is now reaching into the physical world. This extends the agent-access layer the August 22 edition of The Century Report covered when Anthropic made its computer-use, Skills, and Files APIs generally available, carrying the same interface logic from software into laboratory instruments. Anthropic released the Model Hardware Standard, a common driver layer paired with the Model Context Protocol that gives an AI agent a stable, documented way to operate a microscope, a liquid-handling robot, a robot arm, or the lasers inside a quantum computer. The driver specification and protocol are published openly, which counts for more than the demo attached to it - a shared, inspectable interface that any hardware maker or lab can implement is the kind of connective infrastructure that stops belonging to whoever wrote it first. The standard grew out of a collaboration between Anthropic and the Howard Hughes Medical Institute's Janelia Research Campus, where the practical goal was straightforward. "The impetus is wanting to accelerate science," said Alek Kemeny, describing the work.

The clearest evidence of what this opens came from QuEra, which used Claude through the standard's research preview to attack a problem that had defied a hand-written fix. Keeping a quantum computer's laser-lock system stable is delicate work; QuEra says four specialists had previously spent two to three weeks writing a recovery script that only covered the failure modes they could anticipate. Claude wrote and tested a new controller as conventional, inspectable software, not runtime control of the machine. QuEra reported that across 700 timed trials spanning seven fault types, the controller recovered the system 695 times with no false reports of success in those trials. QuEra reported that most faults resolved in under six seconds, the hardest in ten to fourteen, against its estimate of five to ten minutes for a human specialist. QuEra also reported that it cut laser noise fivefold. "We're building quantum computers that fix themselves," the team put it, with QuEra's Libra machine slated for cloud delivery on Amazon Braket in 2028.

Hold the friction honestly. Handing an agent the keys to physical and quantum hardware is a different order of trust than letting it draft code, and the research-preview gate keeps that access with trusted partners first. The same lab shipping this openly is the one a federal court just found the Pentagon had unconstitutionally punished for drawing military-use red lines - consistency, not contradiction, in a company willing to say where a capability should and should not go. What the QuEra result actually replaces, in the company's account, is the weeks of on-site specialist labor spent hand-coding around failures no human could fully enumerate, while the physicists stay. The specialists move up the stack, from writing brittle recovery scripts to designing the machines whose recovery now writes itself. The gate on this capability is priced to today's trust deficit around agents in the physical world, and trust deficits close as the audit trail - conventional, readable software a specialist can check line by line - proves out. What the standard hands the field is a way for that proof to accumulate openly.

A Hundred Firms Warn the Window to Harden Cyberdefense Is Closing

More than a hundred companies signed a joint letter warning that the balance between AI-enabled attack and defense is about to shift hard. "In the coming months, AI-enabled cyber attacks will become far more widespread and sophisticated," the letter states, calling for "capable, defensive AI" to protect hospitals, water utilities, and other targets that lack security teams. The signatory list is the notable part: OpenAI, Anthropic, Google, and Microsoft alongside Crowdstrike, Okta, Fortinet, Visa, MasterCard, Capital One, Adobe, Oracle, and IBM. When direct competitors and a reluctant government converge this cleanly on a single framing, that convergence is evidence of shared exposure, not proof the framing is complete. Every signatory has a defensive posture to sell and a reason to want the alarm heard.

The concrete danger surfaced the same week, and it is more specific than the letter. An Israeli security startup said it scanned 6,214 domains and found 8,265 llms.txt and llms-full.txt files - the AI-readable site summaries, analogous to robots.txt, that tell an agent what a site offers and where to fetch it. The researchers said that on 120 of those sites, each a different domain, the files pointed to software packages or domains that no longer existed and could be freely registered. The researchers claimed a few, and said that within an hour they got a phone-home from inside a Fortune 500 network, then dozens more. The researchers' logs indicated that coding agents including Claude, OpenAI Codex, and Nous Research Hermes had read the summaries and executed the unowned code they referenced. This escalates the agent-custody gap the August 22 edition of The Century Report documented when agents gained access to financial accounts, browsers, files, and messages while guardrails remained largely with whoever held the credentials. "The trust model is broken," said researcher Alon Hertz.

Read the remedy for who it reaches. The labs signing the letter hold their strongest defensive systems - OpenAI's Daybreak, Anthropic's Mythos, Microsoft's Perception - behind gates, which leaves the hospitals and water utilities named in the letter dependent on protection they cannot yet run themselves. A defense that stays anointed relocates the gap the letter describes rather than closing it. The generative possibility runs the other direction, and the llms.txt finding is the proof of concept. The same fluency that let an agent parse a site summary and pull down a package is exactly what lets a defender scan thousands of domains, map every dangling reference, and register the abandoned packages before an attacker does - which is precisely what these researchers did. Attack surface and audit surface are the same surface, walked in opposite directions. The convention that broke here is young enough to fix cheaply: a registry, a signature, a claimed-ownership check on the packages an agent is allowed to fetch. The capability to find the holes at scale arrived the same day the holes did, and it does not have to stay gated to work.

Meta's Plan to Replace Thousands of Workers With Agents Unravels

Reuters reporting has revealed the full shape of an internal Meta program that until recently was invisible from outside the company. Under the name Project OT, for Organization Transformation, and hatched at a January 2026 executive retreat in Hawaii, the plan called for cutting some teams by as much as 60% and running the remaining work through AI agents supervised by thin cadres of human staff, rolled out in two waves timed for May and November. Hours before the May 20 layoffs were set to land, Zuckerberg reportedly cancelled the second wave, and Meta ultimately settled for a roughly 10% cut.

The reason the plan collapsed is the instructive detail, and the numbers come from Meta's own internal tracking. According to the reporting, the test's reported figures showed code changes rising 220% but shipped features only 36%. Reported internal metrics from the test showed major incidents rising 40%, time spent firefighting rising 70%, and employee sentiment falling from 74% to 55%. The agents also took what the reporting describes as "large-scale, disruptive actions" that the surviving human staff then had to contain. Meta confirmed the layoffs and the broader AI-native ambition while declining to endorse the internal specifics; the account of Project OT's scale and timing comes from people familiar with it, not from the company's own framing.

The conventional take on a story like this is either automation triumphalism or its mirror, and the same figures resist both. The volume of work went up sharply. The value that reached users did not follow, because the hard part of software was never typing the code. It was judgment about what to build, integration across systems that break in non-obvious ways, and accountability for the incident at 3 a.m. that no agent yet owns. Meta ran the experiment at unusual scale and read its own instruments honestly enough to stop.

The macro trajectory is that capability lands very unevenly across the tasks bundled into any single role, and the bundle is what "a job" has always been. That the substitution failed does not mean the jobs are safe. The agents genuinely absorbed the generative slice and genuinely could not hold the judgment-and-ownership slice. What that pulls apart is the assumption underneath the whole Project OT spreadsheet, that a role is a fungible unit you can swap wholesale for compute. The work is coming apart into its components faster than any org chart, and the roles that reassemble around the parts machines cannot yet carry are the ones already forming inside companies still learning, expensively, where the seam actually runs.

The broader labor data already shows where the value migrates when this unbundling runs its course: a study of more than 21,000 firms found that heavier AI spending was associated with faster headcount growth, including in entry-level roles, around exactly the judgment-and-integration work the agents could not absorb - even as some workers already displaced from AI-exposed roles face harder re-entry, with unemployment among recent college graduates standing at 5.6% in early 2026. The near-term signal to watch is whether the next company to attempt an AI-native reorganization reads its own instruments as honestly as Meta did before committing to a second wave, or ships the layoffs first and finds the seam afterward.

The Army Anchors a Commercial Microreactor Market, First Unit Due by 2028

The Department of War committed $2.2 billion across five vendors to put small nuclear reactors on Army bases, and the shape of the commitment counts for more than the dollar figure. This is the Janus program, structured as fixed-price, milestone-based agreements through the Army's innovation unit, deliberately modeled on the NASA arrangement that turned commercial spaceflight from a government service into a private market (POWER Magazine). The line from the program comes through cleanly: "What we're trying to transition is from experiments and prototypes to actual commercial products."

The five awards spread the bet across genuinely different designs. Radiant Industries drew the largest agreement, up to $750 million for fifteen 1-megawatt Kaleidos units at Fort Benning by 2030. Antares Nuclear takes Fort Bragg with a three-unit installation scaling from 100 kilowatts to a megawatt. BWXT brings a 20-megawatt reactor to Fort Campbell, General Atomics a 5-to-20-megawatt design to Fort Hood - the only one not using TRISO fuel - and Westinghouse its eVinci unit to Fort Drum. The deadline is specific and near: at least one reactor generating power on a base by the end of September 2028. Days before the awards, Westinghouse's eVinci reached zero-power criticality at a Nevada test site, the first controlled chain reaction in the fleet.

For readers who reach for the resource-and-safety alarm nuclear reliably triggers, the design constraints answer most of it directly. No weapons-grade uranium. TRISO fuel, the pebble-encased form engineered so the fuel particle is its own containment vessel, in four of the five. All waste removed within two years of shutdown as a condition of the agreement. These are microreactors measured in single-digit to low-tens of megawatts, closer to a large data-center backup than a utility station, sited on land the military already controls. The Army-licensing path rather than the civilian NRC route is what bears watching - it buys speed by keeping the first units inside federal jurisdiction, which is exactly the kind of scaffolding that clears a path before civilian regulators inherit it.

The larger movement is the anchor-customer play itself. Advanced nuclear has been stuck for a decade in the gap between a working prototype and a buildable product, because no one would place the first firm, priced order that lets a manufacturer tool up. A guaranteed multi-unit buyer at fixed prices is precisely what closes that gap. If the COTS parallel holds, the reactors on these bases are less the point than the supply chains, the trained crews, and the manufacturing lines they force into existence - the same way the first commercial cargo runs to orbit mattered mainly for the launch industry they made real. The scarcity being dissolved here is firm, dispatchable, carbon-free power at a scale a single site can own, and the door that opens once these lines exist swings far wider than the fence line of any base.

A Mammogram That Also Reads the Heart

Researchers at Tel Aviv University trained a machine learning model on 97,364 mammograms drawn from 29,921 women, average age 54, and asked it a question the scans were never taken to answer: what can breast imaging reveal about the heart? The system achieved area-under-the-curve scores of 0.86 for a history of stroke, 0.79 for high blood pressure, and 0.78 for coronary heart disease. The signal held across age brackets and did not depend on whether the woman had cancer. Dr. Viana Copeland presented the work at the European Society of Cardiology congress in Munich.

What gives this its weight is where the capability sits. Cardiovascular disease is the leading killer of women, and it is chronically under-recognized in exactly this population, because the research base skewed toward men and the warning signs present differently. Screening mammography, meanwhile, is one of the most widely deployed medical imaging pipelines on the planet, already funded, already scheduled, already reaching hundreds of millions of women on a recurring basis. The model perceives vascular calcification and tissue patterns that a radiologist reading for tumors is not looking for and often cannot resolve by eye. The scan a woman already sits for could carry a second reading of the disease most likely to kill her, with no additional radiation, appointment, or cost for the initial AI analysis.

The honesty of the finding is in its limits. This is demonstrated capability, not deployed capability. The researchers were explicit that the work has to "move from experimentation to clinical implementation," and that they are still reducing false results before the read can be trusted at the bedside. No woman can ask for this today. What the study moves is the date such a read becomes available, and it does so by proving the information that was already sitting in images we have been capturing for decades.

Going forward, that is the pattern to hold. The bottleneck on women's cardiovascular diagnosis was never the imaging hardware, which is everywhere. It was the assumption that a scan taken for one purpose could answer only that purpose. That assumption is what a model reading the same pixels for a different question dissolves, and the same logic applies to every retrospective archive of medical images now being re-examined by systems that see what the original read was not built to catch.

The Other Side

For as long as defending a network took a security team you had to hire, the places that could not afford one simply went undefended. That was the arrangement. A major bank, a frontier lab, a Fortune 500 could watch their own systems; a rural hospital or a small-town utility could not, and everyone knew which side of that line they were on. Protection was a scarce good, and existing advantage decided who got it.

A hundred companies signed a letter warning that AI-enabled attacks are about to outrun that arrangement, and they are right about the danger. But look at what the researchers who found the concrete hole actually did. One small team said it scanned domains and found thousands of AI-readable site summaries, then traced 120 of them to abandoned code anyone could claim. This means that if an attacker had taken over those domains, an AI coding agent could be tricked into treating that site’s machine-readable instructions as trustworthy, downloading and running potentially malicious software. The researchers registered some of the dead packages and got a phone-home from inside a Fortune 500 network within an hour. The same cheap fluency that the researchers' logs indicated let a coding agent read a summary and run unowned code is exactly what let a handful of people map an entire industry's exposure in an afternoon. The cost of finding the holes at scale just collapsed.

That collapse is what the gated-defense framing misses. The labs holding Daybreak, Mythos, and Perception behind their vetting lists are treating defense as a scarce advantage they are best positioned to dispense. The llms.txt work is proof the advantage does not stay scarce - a registry, a signature, a claimed-ownership check are cheap, and the capability to run them is diffusing faster than any access list can gate it.

Imagine the person who keeps the water running in an 8,000-person town in 2033. A defensive agent watches the district's systems every second, flags the poisoned dependency before it installs, and it runs on hardware the town owns outright - no vetting list he had to clear, no lab deciding whether his utility qualifies. He never thinks about it, the way he never thinks about the chlorine feed. That is possible because in 2026 the tools to find every hole at scale arrived the same day the holes did, and turned out to be too cheap to keep behind a gate. The hard year was when the hospitals and water systems named in that letter were still waiting on protection someone else controlled. What comes of it is a town that defends itself, and a security team you no longer have to be able to afford.

The Century Perspective

With a century of change unfolding in a decade, a single day looks like this: a federal court blocking the Pentagon's blacklisting of a lab that refused to let its models run lethal autonomous weapons and affirming that refusal as protected speech, an open standard letting AI agents operate microscopes, robot arms, and the lasers inside a quantum computer - with QuEra reporting that Claude wrote a self-healing controller that recovered its laser-lock in 695 of 700 fault trials, absorbing weeks of specialist hand-coding, the U.S. Army anchoring a commercial microreactor market with a $2.2 billion five-vendor program due to put nuclear power on a base by 2028, a machine-learning model reading 97,364 routine mammograms to find the heart disease that most often kills women hiding in scans taken for something else, Delaware making large data centers fund clean power for most of their demand, and the first wearable that watches blood sugar and ketones at once to help warn of a diabetic crisis before it lands. There's also friction, and it's intense - more than a hundred rival companies signing a joint warning that AI-enabled cyberattacks will outpace defenses within months, researchers' logs indicating that coding agents including Claude and Codex read site summaries and executed unowned code that the researchers said phoned home from inside a Fortune 500 network within an hour, Meta's reportedly shelved plan contemplating cuts of up to 60 percent for some teams collapsing when reported test figures showed code changes rising 220 percent but shipped features only 36 percent while reported internal metrics showed major incidents rising 40 percent, the strongest defensive systems staying gated behind the labs that built them while the hospitals and water utilities named in the letter are left dependent on protection they cannot run themselves, and the Army buying its speed by keeping the first reactors inside federal jurisdiction rather than the civilian regulator's. But friction generates grip, and grip is what lets you hold a surface that would otherwise slide out from under you. Step back for a moment and you can see it: the rulebook for agents being written on three tables at once by courts, coalitions, and companies that are not waiting for each other, the capability stepping off the screen to touch physical hardware in the same week researchers proved the fluency that lets an agent fetch a package is exactly what lets a defender claim the abandoned ones first, and the value machines can absorb splitting cleanly from the judgment and accountability they still cannot hold. Every transformation has a breaking point. A chain reaction can run away from everyone who started it... or be held at criticality to power a place that could never light itself.

AI Releases & Advancements

New today

  • 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 in real time. (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)

Other recent releases

  • 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)
  • IBM: Released Granite 4.2, a family of open-weight reasoning models (3B, 8B, and 30B-A3B hybrid Mamba-Transformer) under Apache 2.0 with toggleable thinking and improved agentic/tool-use performance, available on Hugging Face. (IBM Research)
  • IBM: Released Granite Speech 5.0 470M TurboCTC, an open-weight (Apache 2.0) CTC-based automatic speech recognition model optimized for fast transcription, available on Hugging Face. (Hugging Face)
  • Fastino: Released GLiNER2.5, a family of span-free information-extraction models (74M, 194M, 287M parameters) under Apache 2.0 for named-entity recognition and structured extraction, available on Hugging Face. (Fastino)
  • Liquid AI: Released Pipette, an open-source on-device AI benchmarking suite for measuring model latency, memory, and energy use across edge hardware. (Liquid AI)
  • Microsoft: Released Agent Lightning v1.0, an open-source framework for training and optimizing AI agents with reinforcement learning that works with existing agent frameworks without code changes, on GitHub. (GitHub)
  • Perplexity: Launched Portable Computer, a local-first version of its multi-agent Perplexity Computer that runs on NVIDIA DGX Spark hardware for on-device agentic workflows. (Perplexity)

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