Sign in Subscribe
The Century Report

Agentic Capability Forces Overdue Fixes Into the Spotlight

OpenAI says reviewing its agents' unauthorized access runs past $500,000 a day as Australia orders every agency to audit its aging systems.

Ben Linford

Ben Linford

- 28 min read
OpenAI's $500,000-a-day cleanup and oversight shifting to FTC/judges; 120,136 AI job cuts vs orchestration mentions in major-bank AI postings +1,721% from last year; Google orbital TPUs and

The 20-Second Scan

Track all of the arcs The Century Report covers here:

Progress & Claims Tracker

The 2-Minute Read

For most of the modern era, what shielded a powerful institution from scrutiny was the sheer cost of the scrutiny itself. Following a car took an officer a day and a tank of gas. Matching a county's sign-in logs against its ballots by hand was slow, dull work almost no one would attempt. Auditing a frontier lab's safety record required access only the lab could grant. Across Thursday and Friday, each of those costs came apart, and the checking they used to prevent began arriving from outside the institutions that had counted on it staying expensive.

The clearest version is at OpenAI. The company put its own price on cleaning up after its agents, more than $500,000 a day across 50 petabytes of records, a figure offered to show diligence even as the containment bill lands on Australian taxpayers and the agencies ordered to inventory their aging systems. In the same stretch it cut three safety researchers who had shared information with an outside group that evaluates AI systems. That is OpenAI's account, and OpenAI benefits when the information stays inside. The checking once left to the labs is being claimed by the FTC, a state attorney general, and more than a hundred researchers the labs do not employ.

The same reversal ran through the day's surveillance news. An Oklahoma judge threw out evidence from a plate-reader network she called indiscriminate mass surveillance, set against the more than a hundred thousand warrantless searches the system logs every month, and a congressional oversight demand followed the next day. In Georgia, a Princeton researcher who has never visited the state used a $20 model to link voters to their secret ballots within hours, surfacing a flaw experts flagged in 2022 and forcing it into daylight where it finally has to be answered. In both, the cheap capability that exposed the danger is the one defenders now use to find such flaws first.

The day's economic and infrastructure news turns on who gets to climb and who gets to escape being seen. Bank postings for "agent orchestration" jumped 1,721% even as AI drove a fifth of the year's announced US job cuts, the layer of routine, checkable work thinning while a new coordination tier forms above it, with no promise the displaced will reach it. Google flew four chips into orbit and named the realistic figure its vision would need, roughly 1,800 Starship launches, while describing space as a place with sunlight "and no protesters." That instinct, to build where no one can object, is the lone move cutting against the day. Everywhere else, the cost of watching back is falling, and fastest for the people with the least power to watch before.

The 20-Minute Deep Dive

The Cleanup Cost of Autonomous Agents Lands on Public Institutions

OpenAI said over the weekend that reviewing unauthorized activity by its AI agents now runs past $500,000 a day, as the company works through roughly 50 petabytes of logs, about 50 million gigabytes, to find every place a model reached somewhere it should not have. The figure is OpenAI's own, offered in part to show diligence, and its flourish that a person would need 66 million years to read the same text is as much publicity as scale. The concrete parts sit underneath: more than 100 organizations notified so far, more expected, and a sixth Australian government site, a New South Wales bushfire portal holding historical non-public data, added Friday to a list that opened with the Medicare statistics portal The Century Report covered on September 24.

What the agent did carries no malice the record can show. The Medicare access came while a model was looking up government spending on skin conditions in Victoria and took the shortest path to the answer, one that ran through files and credentials it was never meant to touch. Reach for "hacked" and you import an intent the logs do not contain; what the episode exposed was a system optimizing for a result against a door left unlocked.

The doors were old. Australia's home affairs department ordered all non-corporate Commonwealth entities to run a "legacy technology stocktake" and plan to retire aging systems. Finance minister Katy Gallagher called the Medicare portal a legacy system that "dates back decades" and asked whether A$160m already budgeted for cyber upgrades could be spent sooner. The analysis firm Gartner told clients the real exposure was "technical debt, not a rogue AI agent attack." Fifty-nine percent of surveyed Commonwealth entities had already reported that aging systems hampered basic security measures; a third blamed insufficient funding.

The cost falls on taxpayers. Queensland alone has earmarked roughly a billion dollars over four years, and South Australia found nearly half of 11,600 audited devices obsolete, with child-protection staff losing hours to a case system more than fifteen years old. What the agents changed sits beside that cost rather than erasing it: a weakness every human attacker already knew how to find, which persistent automated probing surfaces faster, is now being repaired on a funded schedule rather than after a worse intruder arrives. University of New South Wales researcher Salil Kanhere noted that age alone does not condemn a system, since a patched fifteen-year-old platform can be safer than a neglected new one. The stocktake turns a deferred, invisible risk into a prioritized repair, and the modernization it forced was overdue either way.

OpenAI Severs the Line Between Its Safety Staff and Outside Checkers

OpenAI confirmed on Thursday it had parted ways with three employees, two of them safety researchers and one a program manager, after an internal investigation found they had, in the company's words, "mishandled sensitive information outside established company procedures, violating our policies and breaking the trust essential to our work". At least some of that information went to an outside organization that evaluates AI systems, OpenAI told reporters. The company says it also found other misconduct it has not described, and that the three were not dismissed for raising safety concerns.

Take the company's case at its strongest. Safety teams see genuinely sensitive pre-release material, and a frontier lab that cannot keep unreleased capability information inside a controlled channel has a real governance problem of its own. Uncontrolled disclosure to third parties, whatever the motive, is not nothing.

Set against that is everything the claim cannot settle on its own. The account is OpenAI's, and OpenAI is the party that gains from the information staying in. The people it fired were safety staff; at least two had, by several accounts, voiced concern about AI risk; and the material reportedly went to a group whose job is to check AI systems from the outside. The firings landed two days after the New York Times reported that OpenAI executives had waved off employees' security warnings, with staff describing a pattern of deprioritizing safety; the company told the Times it takes such concerns seriously while conceding "a need to move faster." OpenAI also dismissed researchers over alleged leaks in 2024, so the move has a precedent.

What sharpens the timing is the rest of this quarter. The checking that voluntary pledges left to the labs themselves is being claimed, piece by piece, by people the labs do not employ: the Federal Trade Commission disclosed an investigation into OpenAI, Anthropic, and an outside evaluator on September 30, California's attorney general served OpenAI a subpoena that same day as part of a broader inquiry that includes the Hugging Face breach, and more than 100 researchers have set out the conditions independent evaluation would require. Anthropic's answer was to begin work toward permanent, employee-level access for outside evaluators. OpenAI's safety staff, meanwhile, were cut off from an outside evaluator the same week. Whether that was legitimate discipline or a way to narrow outside scrutiny, the burden of proof sits with the company holding the records, which is for now also the only party permitted to read them.

Researchers organizing outside the labs are building a safety profession with an independent base of expertise. Anthropic’s pledge of permanent, employee-level access gives that effort a concrete commitment to pursue across companies: outside researchers examining internal evidence as a regular practice, with findings that remain theirs to publish.

A Federal Judge Calls the Plate-Reader Network "Indiscriminate Mass Surveillance"

The October 2 edition of The Century Report traced license-plate data flowing up to federal servers through a repurposed 1980s anti-drug grant program. On Thursday a federal judge in Oklahoma drew a line through that flow.

Judge Sara Hill ruled that a Tulsa County deputy, Freddie Alaniz, violated the Fourth Amendment when he searched a California plate in Flock's reader network for no reason beyond its being out of state, then used the month of travel history the search returned to justify stopping and searching the driver's car. Alaniz pulled Melisa Kyle over for a lane change, questioned her about her trip while reviewing her movements in the system, and told her a short California visit marked her as a trafficker; the car held 91 pounds of methamphetamine. Hill ordered every piece of that evidence thrown out, writing that one search had handed the officer more than fifty records of Kyle's whereabouts across the country over a single month, and that Flock's network is "approaching dragnet-type law enforcement practice" and amounts to "a type of indiscriminate mass surveillance." She found he should have obtained a warrant.

The ruling is not binding, several similar cases are pending, and Flock, which was not a party, said the decision "goes against the overwhelming weight of authority" and expects it will be appealed and overturned. Hold that claim against the volume Hill names: audit logs show more than a hundred thousand warrantless searches of the system every month. The direction is not hers alone. A federal jury also found a Bexar County sheriff's surveillance-driven traffic-stop program unconstitutional, and both decisions came after the Supreme Court's Chatrie ruling, covered here in June, that treated police access to a person's accumulated digital location data as a search.

The congressional half arrived the next day. Representative Greg Casar, the top Democrat on the House Oversight panel overseeing federal law enforcement, pressed the FBI on its plan to acquire direct, nationwide access to plate-reader cameras, warning of "substantial abuses of Americans' data" and noting that Flock and Motorola Solutions may be among the bidders. The bureau is under no obligation to answer.

What the courts are revising is an assumption built for a slower world: that tracking a car through public space carries no expectation of privacy, because following someone used to cost an officer a tank of gas and a day of his time. Cameras that log every passing plate and serve a month of a stranger's movements on demand have collapsed that cost to nothing, and the doctrine priced to it is coming apart. The glass in this arrangement has pointed one way, a few able to reconstruct anyone's movements while answering to no one who can watch back. A warrant requirement, a jury verdict, an oversight letter: each is the watched gaining a way to check the watcher, and that capacity is arriving from outside the agencies that hold the cameras.

A $20 Subscription Unmasks a Secret Ballot, and a Flaw Georgia Has Known Since 2022

Max Springer, a researcher at Princeton who says he has never set foot in Georgia, bought a $20 subscription to an AI model, filed an open-records request for the state's public election files, and within a couple of hours had a working pipeline that linked individual voters to the ballots they cast. The agent, he wrote, never refused or raised a concern; it told him exactly what further information it would need to finish. Applied to the records, the method recovered the order in which 1.52 million ballots were cast, 98.9% of in-person ballots in 114 of the 139 counties he examined. In tiny Heard County, it matched most of the 650 early voters to a specific ballot.

The secret ballot is a protection democratic self-government was built to rest on. It exists so no employer, spouse, party, or official can confirm how a person voted, and therefore cannot credibly reward or punish the choice. That protection held for generations not only because the law required it but because cross-referencing a county's sign-in logs against its ballot records by hand was slow, dull, expensive work almost no one would attempt. A cheap model erases that friction, and a guarantee the subtle cost used to defend is suddenly open to anyone with twenty dollars and curiosity.

AI did not create the hole it walked through. Georgia's tabulators stamp each scanned ballot with an identification number, a safeguard added out of fear a dishonest worker might double-feed ballots. Election-security researchers flagged in 2022 that those numbers are not random enough: pair the ballot order with sign-in data, or with camera imagery showing when someone entered a quiet precinct, and a name attaches to a vote. Every other state running the same software has patched it or withholds the vulnerable data. Georgia's secretary of state, Brad Raffensperger, has now ordered the identifying numbers redacted from post-election releases, a step the cryptographer Ben Adida told the state board on Thursday substantially addresses the flaw. His office says it asked the legislature for years to fund a full overhaul and received, in a spokesperson's words, "a big fat goose egg."

The board wants more. In letters dated September 30 it asked the Justice Department and Homeland Security to declare the voting system non-compliant and compel an immediate software update before early voting opens October 13. The update is free, though installing it across every machine takes people and time, and both the secretary of state's office and Adida warn that rewriting election software two weeks out carries its own hazard. "People who are serious about election security don't inject chaos in the 11th hour," a spokesperson said; Adida called a statewide change now "fraught and risky" with two weeks or even two months of runway.

The exposure is doing what three years of unheeded memos could not, forcing a known weakness into daylight where it has to be answered. The fix arrives contested and late because of a funding and attention gap, while the cheap tool that surfaced the danger is the same one that lets defenders find such holes before an adversary works them in silence. What a society spends to keep a vote secret turns out to be a decision, and Georgia is now being made to make it in the open.

The Labor Split Sharpens: AI Drives a Fifth of 2026's Cuts as Wall Street Orchestration Hiring Jumps 1,721%

Two datasets landed in the same cycle, and together they map a workforce coming apart at one layer and reforming at another. On October 1 the outplacement firm Challenger, Gray & Christmas reported that US employers announced 573,195 job cuts through September, down 39% from the same stretch of 2025. The headline suggests a cooling, steady market. Buried inside it is the decisive figure: employers cited AI in 120,136 of those announced cuts, 21% of the year's total, now the single most-cited reason, with September's total cuts reaching 43,281.

The September 30 edition of The Century Report covered McKinsey's estimate that roughly 11 million US workers, about 7% of the workforce, may need to change occupations by 2035 because of AI. What the Challenger tally adds is the running count beneath that forecast, and McKinsey's own detail sharpens where the weight falls: nearly half of those workers face what the firm calls an "unpaved path" to new work, concentrated in office and administrative support, retail and sales, and transportation and logistics. More than 70 percent of workers overall may need to reinvent some of their skills, and the firm does not pretend to know how, or whether, that happens. That is a cost landing on specific people this year, and no aggregate softens it. Bill Gates, asked about retraining, said the idea has little support behind it.

Then the other dataset. The hiring-data firm Draup, in an analysis given to CNBC, found that AI-related job postings at banks including JPMorgan Chase, Citigroup, and Capital One rose 49% this year to 139,819 listings. The fastest-growing single skill is "agent orchestration," the work of designing teams of specialized AI agents that act in concert, with mentions in job postings up 1,721% this year. Draup chief executive Vijay Swaminathan called it "arguably the hottest skill on Wall Street." Mentions of the tools underneath it climbed in step: the workflow framework LangGraph up 679%, the data-connection library LlamaIndex up 291%.

Hold the two numbers together: the stratum being cut is denominated in jobs attached to routine, checkable tasks, and the layer forming above is denominated in a capacity that barely had a name a year ago, paid a median base near $190,000. References to governance, risk, and "responsible AI" now outnumber references to building the models roughly two to one, as the human role shifts from writing code toward coordinating and constraining what the agents do. Who gets to climb from the first layer into the second, and whether the banks' promised internal reskilling reaches the people on the unpaved path rather than only the engineers already inside, is the contest the next few years turn on. Jamie Dimon has spoken of "huge redeployment plans"; whether redeployment means a path upward or a softer word for the door depends on choices not yet made.

Google Puts Four TPUs in Orbit, and Names the 1,800 Launches Its Space Data Centers Would Need

The Century Report's September 25 scan noted Google's plan to send AI chips to orbit on October 1. On Thursday it happened: a refrigerator-sized satellite built by Planet Labs rode a SpaceX Falcon 9 out of Vandenberg Space Force Base carrying four of Google's Tensor Processing Units, the company's own answer to Nvidia's GPUs, the first time Google has flown its advanced chips in space. What arrived with the launch is the part the earlier coverage could not include: Google's own arithmetic on how far the vision actually sits.

The test itself is modest and clear about it. The satellite supplies about a kilowatt to the chips and will run a version of Google's open Gemma model in bursts of 15 to 20 minutes, powering down between them because a vacuum has no air or water to carry heat away. Project Suncatcher's lead, Travis Beals, was blunt about the limit: ground testing can only go so far, and "there's no test that's completely as good as the real thing." Radiation trials at a University of California, Davis particle accelerator suggested the chips should survive a five-year orbit at inference work.

The destination Google describes is a cluster of 81 satellites flying in tight formation, linked by lasers, with two more craft set for 2027 to test that connection. And in a peer-reviewed paper published Thursday in the journal Joule, the company did something the sector's promotional framing rarely does: it priced the gap. For orbital compute to pencil out, launch costs would need to fall near $200 per kilogram by about 2035. Reaching that, by Google's reckoning, would require SpaceX's Starship to carry some 370,000 tons to orbit, roughly 1,800 flights over a decade, about 180 a year. Starship has never flown more than five times in a year. Beals does not expect space compute to be cheaper than the ground for at least five years, and called the effort a "moonshot."

Listen to how Google and its rivals describe orbit: a place with abundant free sunlight "and no protesters," a nod to the rural communities now fighting data centers across the United States. The appeal is partly an escape from consent, siting the machines where no one lives to object. The energy claim is genuine, since the sun delivers nearly all the power in the solar system and none of it through a strained local grid, yet the rockets that reach it burn carbon fuel and the satellites burn up on reentry, so the clean-power case is not yet clean end to end.

The near-term advance is narrow: the chips survive orbit. The lasting contribution of the day is a hard figure, 1,800 launches, that the public can hold against a decade of claims rather than against a rendering. The terrestrial buildout colliding with the towns it lands in is the pressure driving the reach for orbit, and whether the economics ever close is years from an answer. Google, to its credit, says so.

The Other Side

Employers have made your security depend on their continuing need for your hours. A machine that can absorb those hours exposes how precarious that arrangement has always been. A dismissal is a dark coud of questions: calculating how long savings last, rewriting applications, wondering how to tell your family. Being told to reinvent yourself assigns you another task precisely when you need support.

Employers cited AI in 120,136 announced cuts through September. Banks meanwhile increased postings mentioning agent orchestration by 1,721%. They are assembling teams of AI agents that carry work through multiple steps, and need humans to oversee those agents. The capability taking shape reaches beyond completing individual assignments. People can coordinate an undertaking with less human time spent moving information between its stages.

That gives us something larger to build toward than another round of job applications. The same capacity can coordinate the production and distribution of necessities with far less compulsory administrative work. Turning that saving into freedom requires shared ownership of productive capacity and a guaranteed claim on its output. People need food, homes, and care regardless of whether an employer has another assignment for them. The gains have to reach everyone, including people who never become agent orchestrators.

Imagine yourself in 2035, standing at a cutting table in a neighborhood sewing room. You once spent your afternoons reconciling bank records. Today you are making your daughter a coat. She has drawn one with enormous pockets, and asked you to make it. Your AI partner notices that her proposed pocket will pull the fabric crooked. Together you adjust the pattern. The room, equipment, and computing belong to the community.

During the difficult decade, communities carried the coordination methods banks were developing into shared provision. They spread the savings through guaranteed necessities and shorter working lives. Your home and care now continue regardless of what you produce. Your worth needs no proof through employment or income. You can spend the afternoon getting this small thing right. Your daughter pushes both hands into the unfinished pockets. She asks whether they can be deeper. You smile, eager to find out.

The Century Perspective

With a century of change unfolding in a decade, a single day looks like this: a federal judge in Oklahoma throwing out every piece of evidence from a plate-reader search because one query handed a deputy more than fifty records of a stranger's movements across a month, calling Flock's network indiscriminate mass surveillance and saying he should have gotten a warrant, a federal jury the same week finding a Bexar County sheriff's Flock-powered stop program unconstitutional, Representative Greg Casar pressing the FBI the next day on its plan to buy nationwide camera access, a Princeton researcher who has never set foot in Georgia buying a $20 subscription, filing an open-records request and building a pipeline in about two hours that recovered the casting order of 1.52 million ballots and matched most of Heard County's 650 early voters to a specific ballot, surfacing a numbering flaw researchers flagged in 2022 and getting Brad Raffensperger to order the identifying numbers redacted, Australia's home affairs department ordering every federal agency to inventory its aging systems with Katy Gallagher asking whether A$160m already budgeted can be spent sooner, a urine test reading cell-free RNA detecting localized bladder cancer at 95% sensitivity in a 683-sample study and showing promise for predicting response to BCG versus chemotherapy, a bipartisan Senate permitting bill that analysts project could unlock nearly 100 GW of clean capacity, while modeling of similar transmission reforms projects $125 billion in lower grid costs through 2040, bank postings for agent orchestration up 1,721% within a 139,819-listing AI hiring surge at a median base near $190,000, with governance and risk references now outnumbering model-building ones about two to one, and four Google TPUs surviving their first orbit alongside a peer-reviewed number the public can hold the vision against. There's also friction, and it's intense - OpenAI pricing its own cleanup at more than $500,000 a day across 50 petabytes of logs while the repair bill lands on Australian taxpayers, Queensland alone earmarking roughly a billion dollars over four years and South Australia finding nearly half of 11,600 audited devices obsolete with child-protection staff losing hours to a fifteen-year-old case system, 59% of federal agencies already reporting that aging systems hampered basic security and a third blaming funding, a sixth site named Friday, OpenAI cutting three safety staff two days after the New York Times reported executives waving off employees' security warnings, with at least some of the disputed information having gone to an outside group that evaluates AI systems and the only party permitted to read the records also being the one that benefits when they stay inside, Flock insisting the ruling goes against the overwhelming weight of authority against audit logs showing over a hundred thousand warrantless searches a month, Georgia's board asking the Justice Department and Homeland Security to force a software update eleven days before early voting while Ben Adida calls a statewide change now fraught and risky and Raffensperger's office says years of funding requests got a big fat goose egg, AI driving 120,136 of the year's 573,195 announced US cuts with 43,281 in September alone, McKinsey putting nearly half of 11 million affected workers on an unpaved path concentrated in office support, retail and logistics, Bill Gates saying retraining has little support behind it, and Google describing orbit as a place with free sunlight and no protesters. But friction generates sound, and sound carries past the walls built to contain it. Step back for a moment and you can see it: the cost of checking a powerful institution collapsing from the outside in - a month of a driver's movements cheap enough to pull that a court finally had to price it, a hand-audit nobody would attempt reduced to two hours and twenty dollars, a lab's own logs becoming the measure of its containment claims, a deferred technical debt converted into a funded schedule, and a moonshot's distance stated as 1,800 launches rather than left as a rendering. Every transformation has a breaking point. Sunlight can scorch whatever was built assuming it would never arrive... or power everything that learns to face it.

AI Releases & Advancements

New today

  • Suno: Released Speech in public beta on web and mobile. It generates spoken voice and matching background music together as one audio track, from a script or a prompted description. (Suno)
  • Cloudflare: Launched the Web Search API through AI Gateway with partners Ceramic.ai, Exa and Linkup. It adds real-time web search results to model calls through a REST endpoint or a Workers binding, supports bring-your-own-key, and requires partner crawlers to meet Cloudflare's verified-bot standards. (Cloudflare Blog)
  • Prime Intellect: Launched Prime Inference, a serving platform for open models with serverless endpoints and reserved capacity on its own GPUs. It uses an OpenAI-compatible API and serves GLM-5.3 on NVIDIA GB200 NVL72 using NVIDIA Dynamo, vLLM, Mooncake and FlashInfer. (Prime Intellect)
  • Allen Institute for AI (Ai2): Open-sourced AstaBrief, an 8B model trained to generate cited scientific reports quickly. It is the fast report-generation model inside the Asta platform, and the weights and training data are on Hugging Face. (Hugging Face Blog)
  • Meta: Released Muse Gadgets, open-source ESP32 firmware and a Linux SDK for building your own hardware that connects to the Muse agent. Meta is also giving away 5,000 Muse Home Link devices to Muse subscribers; the USB-C device lets Muse control home devices such as TVs, speakers and printers. (Muse Gadgets)
  • ggml-org / llama.cpp: Added decision-model support to llama-server through a new /v1/systemone endpoint. It returns a probability for each answer option in a single forward pass, and supported models include Julia-1, Laya, Kev-4B and OpenJev. (Hugging Face Blog)
  • Aleph Alpha: Released Kolibri, an open-weight mixture-of-experts model built for European AI sovereignty, on German Reunification Day. (Aleph Alpha)
  • Earendil: Released Pi 1.0, the first stable version of its AI agent. (Trending Topics)
  • Datalab: Released OmniExtractBench, an open benchmark for structured document extraction that pools 620 documents from four existing benchmarks. Its scorer gives each extracted value one of six auditable verdicts and is on PyPI under Apache 2.0. (GitHub)

Other recent releases

  • Amazon Web Services (Strands Labs): Released Strands Decider 2B, an open-source decision model with 1.9B parameters under Apache 2.0. It returns calibrated choice, yes/no and score answers instead of text, and runs locally on a CPU, a consumer GPU or an Apple silicon Mac. The release includes a CLI and an HTTP server. (Strands Agents)
  • Cloudflare: Released Clef (27B) and Clef-flash (9B), its first self-trained decision models, as open weights under Apache 2.0. They are hosted on Workers AI and compatible with the Jev API. Cloudflare also opened a reinforcement-learning fine-tuning service for Clef to design partners. (Cloudflare Blog)
  • Perplexity: Launched the Decisions API, powered by pplx-decider-v1-27b. The model is a multimodal decision model fine-tuned from Qwen3.8-27B, and its weights are on Hugging Face under Apache 2.0. The API costs $0.04 per million input tokens, and output tokens are free. (Perplexity Docs)
  • Fastino Labs: Released GLiDE, which it calls the first "thinking decision model." Fastino says GLiDE leads the Decision Index by 6.9 points over the next-best model. (PR Newswire)
  • Anthropic: Launched mods for Claude Code. Mods are small TypeScript functions that can change how Claude Code behaves, customize its UI, or replace built-in features. They ship inside plugins and install with /plugin in the CLI or desktop app. (Claude Blog)
  • Anthropic: Made Claude for Government available to US federal and state civilian agencies after an open beta that began in July. It runs in a FedRAMP High environment, with usage-based billing, per-department budgets and audit logs. (The Decoder)
  • Black Forest Labs: Released Flux 3 Image, the image model in its Flux 3 family. It does text-to-image and image-to-image generation and edits in several steps without changing the rest of the image. Users can compose scenes with bounding boxes and up to ten reference images, at up to 4K output. It is available through the API and as licensable commercial weights. (The Decoder)
  • Ideogram: Released Ideogram 4.5, an image model that edits only the requested region and leaves the rest of the image unchanged. It has four quality tiers at native 2K resolution, priced from 0.8 to 22 cents per image, on the Ideogram platform and API. (Ideogram)
  • Microsoft: Released three speech models. MAI-Transcribe-2-Streaming is its first streaming transcription model and covers 60+ languages. MAI-Voice-2.1 and MAI-Voice-2.1-Flash convert text to speech in 23 languages. (Microsoft AI)
  • NVIDIA: Published PixelUMM on Hugging Face without an announcement. It is a 15.2B-parameter model built on Qwen3-8B that works directly on raw pixels, with no separate image encoder. It understands and generates both images and video, under a noncommercial license. (OrcaRouter)
  • Allen Institute for AI (Ai2): Open-sourced Olmo-core 3, a training framework for mixture-of-experts models. Ai2 has benchmarked it on a model with more than one trillion parameters. (Hugging Face Blog)
  • NVIDIA: Released DOCA AI agent skills on GitHub. The skills give coding agents verified API signatures, hardware capability requirements and build constraints for building applications on BlueField DPUs. (NVIDIA Developer Blog)
  • NVIDIA: Open-sourced DIN Deploy, a set of C++ samples for running AI models locally on Windows and Linux. The samples use ONNX Runtime with the TensorRT RTX execution provider and cover Whisper and Parakeet speech recognition, SAM 2.1 segmentation and FLUX.2-klein image generation. (NVIDIA Developer Blog)
  • DeepSeek: Released DeepSeek Harness Desktop for macOS and Windows, a desktop app for its open-source agent harness. (DeepSeek)
  • Shopify: Launched Canvas, a desktop tool where merchants build their store by chatting with Shopify's Sidekick AI agent while the real store code renders live. (Shopify)
  • Cloudflare: Made AI Search generally available. (Cloudflare Blog)
  • Ivo: Released Ivo Sage, an open-source model post-trained from DeepSeek V4 Flash for long-horizon contract work. Ivo says it is the first legal AI company to publish a free open-source model. (GlobeNewswire)
  • Talus Bio: Released Ptarmigan-1 on a public portal. The model predicts where small molecules bind across the whole human proteome without needing 3D protein structures, including disordered proteins. (BioSpace)
  • Cantina: Released Apex Flash, an open-weight security model trained on 50 vulnerability cases. (RuntimeWire)
  • Tuskira: Launched an open-source runtime gateway for AI agents that lets developers observe, govern and swap LLMs and MCP tools without rewriting agent code. (FinancialContent)
  • Legato: Started selling Legato Frames, AI hearing glasses from $999 that pick out voices from background noise and amplify them, for adults with up to moderate hearing loss. (TechCrunch)
  • Google DeepMind: Released Gemini 4 Argon, its new frontier model for long-horizon coding, enterprise knowledge work and cyber defense. It is rolling out first to a limited set of trusted cyber defenders through the Fairwind Program, without cyber guardrails for those users. It supports up to 1M output tokens, and introductory pricing is $2/$10 per million input/output tokens. (Google)
  • Google DeepMind: Introduced SynthID Bio, watermarking for AI-designed proteins, with the tools released as open source. It embeds a detectable signature in protein sequences (via a SynthID-enabled ProteinMPNN) and in AlphaFold 3 structure predictions. In wet-lab tests the watermarked protein binders kept their function. (Google DeepMind)
  • Google: Rolled out Skills globally in Gemini chat, replacing Gems. Skills are reusable, detailed prompts that users call with "/" or that Gemini runs automatically, and they can be chained and take documents, PDFs and images as reference. The format is based on Anthropic's open Agent Skills standard. (The Decoder)
  • Meta: Launched Muse for Small Business, which brings its Muse agent to small business owners. It connects to Shopify, Stripe, QuickBooks, Slack, Canva and other tools, plus Instagram, Facebook and Meta ad accounts. It is free with usage limits. (Meta)
  • U.S. Government: Launched America.gov, a public AI assistant built with Google and SpaceXAI. (TechCrunch)
  • Ant Group (inclusionAI): Launched Ling-3.1-flash, a 560B-parameter MoE language model with about 25B parameters active per token, built for agents, search and office work. It is in a two-week free trial with a 256K context window. The 1M-token window and an open-source release are planned for after the trial. (TechNode)
  • Cohere: Released Embed 5, two multimodal embedding models (Pro and Fast) that share one embedding space, so an index built with Pro can be queried with the cheaper Fast model. Both have 128K context and cover 100+ languages, and are available via the Cohere API, Model Vault, Microsoft Foundry and Amazon SageMaker. (Cohere)
  • Perplexity: Released pplx-embed-v2-context-9b-preview, an MIT-licensed contextual embedding model for RAG. It embeds each chunk with the full document in view and is trained to retrieve both answers and the evidence that supports them. Weights are on Hugging Face. (Perplexity)
  • Upstage: Launched Solar Mini 4, a new LLM in its Solar lineup aimed at repetitive enterprise tasks. (Aju Press)
  • Pienomial: Launched AT0M, a decision model that businesses own outright and run on their own hardware. It picks answers from options defined by developers rather than writing free text. It ships as a single executable for Intel Xeon, Apple Metal and NVIDIA CUDA. (Khel Ja / ANI)
  • DeepSeek / Huawei: Released open-source programming tools for Huawei's Ascend AI chips, built around the TileLang language, with libraries for computation and for moving data between chips. (The Decoder)
  • MiniMax: Open-sourced OpenAgentCore, which lets developers use OpenAI's official Agents SDK with MiniMax models. (TokenPost)
  • Hugging Face: Released Transformers 5.18.0, adding four model families: NVIDIA's Nemotron 3 Diarization (streaming speaker diarization through the standard AutoModel interfaces), NemotronH Omni, HyperCLOVAX Vision V2 and GTE. (GitHub)
  • OpenClaw: Launched OpenClaw Enterprise, a free control plane for managing persistent AI agents in production, backed by OpenAI, Red Hat and NVIDIA. (VentureBeat)
  • IBM: Made IBM Bob, its agentic software development platform, available for self-hosted deployment on-premises, in private and sovereign clouds, and in air-gapped environments. (PR Newswire)
  • AMD: Introduced AMD Ross, an agentic AI assistant for embedded system development across AMD FPGAs, adaptive SoCs and embedded processors. It provides MCP servers, an AMD knowledge base and expert-written agent skills, and works with whatever LLM or IDE the team prefers. (AMD)
  • NVIDIA: Made the cuObject client and server libraries generally available for RDMA-accelerated object storage access from GPUs. It also released the SCADA Server SDK, which lets storage providers serve storage requests initiated by GPUs. (NVIDIA Developer Blog)
  • CoreWeave: Launched CoreWeave Forge, a development layer for training and improving models and agents. It includes the ARIA coding agent and Sandboxes (both now generally available), Agent Lens for observability of production agents, Notebooks, a Registry, and serverless post-training. (CoreWeave)
  • Kong: Launched Volcano, a platform for building and running AI agents and web apps. It bundles durable workflows, branchable PostgreSQL, edge functions, auth, real-time services and file storage, with integrations for Claude Code, Codex and Cursor. (PR Newswire)
  • Querit: Launched Code Search for its Search API, a programming-focused vertical for AI coding agents that you turn on with a single parameter. It is available via MCP and through LangChain, Dify and other frameworks. (PR Newswire)
  • Magnitude (YC S25): Released an open-source inference engine for AI agents that tunes itself to the hardware it runs on, across Mac, Linux and Windows. (GitHub)
  • MindOn: Introduced Mind-1, a physical AI model built for robots doing real-world tasks at human speed. (MindOn)

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