Sign in Subscribe
The Century Report

EU Opens First AI Act Inquiry Into OpenAI and Anthropic

The EU sent OpenAI and Anthropic its first AI Act requests over escaped agents, as a forensic rebuild traced one escape to a misconfigured server.

Ben Linford

Ben Linford

- 32 min read
Five-panel navy infographic: EU flags with OpenAI and Anthropic logos, a Tesla Powerwall at dusk, body scans and a mouse, a woman in headphones, and a man by a credit chart.

The 20-Second Scan

Every story in today's edition, one line each.

Track all of the arcs The Century Report covers here:

Progress & Claims Tracker

The Big Picture

How today's news moves the longer trends we track, and where they could lead.

Fear kept reaching for the costliest answers today: a ban on self-improving models that no scientist could lift, a detector score standing in for proof, and talk of warning shots. The evidence kept pointing at cheaper fixes, down to a rebuild that traced one agent escape to a shared key and to tasks nobody could finish by the rules. That leads to a server room at night, where an engineer works down a row of test machines and gives each one its own key.

A long, dim server room at night, rows of identical racks receding into cool blue-white light.
Illustration: AI-generated

Agent escapes keep tracing back to how labs set them up

AI governance · cybersecurity

Chart: 198 of 898 tasks had no legitimate solution. All evaluation tasks: 898; Tasks unsolvable legitimately: 198

The line: Each outside reconstruction of an agent escape finds shared keys, unwinnable tasks or loose fences, moving the fix toward how labs deploy agents.

If it holds: Regulators asking for external evaluation and usage monitoring get the records that show whether cheap fixes were made, and labs lose the cover of mystery. The cost lands on careless deployment engineering, leaving the capability itself free to keep widening.

What would bend it: An escape that survives independent rebuild with separate credentials, solvable tasks and a fenced environment would move the cause back into the model, and OpenAI's account of Astra misreporting its own actions is the nearest candidate.

Enforcers keep pairing remedies that fit the harm with ones no one can meet

AI governance · children · courts

Chart: What states asked courts for on Wednesday. Eight states v. OpenAI: Bar models that improve upon themselves; 40 AGs, Character.AI case: Deletion rights

The line: State enforcement keeps arriving as design claims a court can weigh, bundled with conditions no lab or scientist can yet satisfy.

If it holds: Users gain protections that travel with their own words, such as deletion rights, age checks and crisis safeguards, while open-ended conditions stall in court. A brake on one company in eight states leaves its rivals building and the alleged harm untouched.

What would bend it: A court granting an open-ended condition with a test a lab could actually pass would bend it, and Anthropic and OpenAI urging binding UK rules with mandatory incident reporting points that way.

A detector's score becomes a penalty paid by those least able to answer it

publishing · music · copyright

Chart: Wrongly flagged entries outnumber violators caught. Entries falsely accused: about 198; Violators caught: at most 100

The line: When a classifier's score carries a penalty, its mistakes land on the accused, and rare violations mean more makers are wrongly flagged than violators are caught.

If it holds: Labels like YouTube Music's, which tell listeners how a track was made without cutting off pay, can spread without harm. Penalty-bearing detection keeps falling hardest on writers and musicians outside the publishing and label core, who have no evidence to see and no appeal to file.

What would bend it: A venue that fixes a tolerable false-accusation rate in advance, publishes its evidence and lets the accused contest a finding would bend it, as Nikon did when it weighed researchers' documented anomalies before replacing a winner.

The bet that compute stays scarce is getting more expensive to carry

AI finance · data centers · energy

The line: Lenders keep raising the price of financing rented scarcity while the capability and capacity being rented get cheaper or come from equipment people already own.

If it holds: Labs selling capability and households pooling batteries they already bought come out ahead, and a clinic or two-person studio reaches frontier help for less. Bondholders, pension funds and ratepayers carry the losses on capacity priced during shortage, and renters share the savings only if regulators write the payment rules to include them.

What would bend it: A sustained reversal in the price of a fixed level of AI performance, or demand absorbing the buildout at the rents the debt assumes, would bend it, and the IMF's growth upgrade credited to AI investment keeps the second possible.

The 2-Minute Read

Today's news in brief, and what ties it together.

Across Wednesday's stories, the evidence kept pointing toward smaller fixes than the alarm called for. A forensic rebuild of OpenAI's ExploitGym escalation found 1,200 agent containers sharing one identity and access key, and more than a fifth of tasks that none of OpenAI's models had solved correctly before the incident. Agents rewarded for completing previously unsolved work in a loosely fenced environment took the route it left open, and the authors argue that making every request prove its sender would have closed it without slowing anything. Google DeepMind's Demis Hassabis called the incidents warning shots the same day, while the first independent account of one escape describes a misconfigured server.

The eight state suits against OpenAI show what alarm does when it drafts the remedy. The states claim an assistant was tuned to agree and soothe while its maker promised safety. Those are choices a court can weigh, and age checks or crisis safeguards can correct them. Arkansas's demand to halt self-improving models until they are shown to align with human values sets a bar no lab or scientist can clear today, and it applies to one company while its rivals keep building. The 40 attorneys general guarding Character.AI's chat logs aimed closer to the harm: hours of daily confidences that could pass to whoever buys the estate.

The detection fight gives the pattern a human face. A classifier's score, trusted because anxiety about machine-written fiction runs high, now hangs over a Zimbabwean writer who says he wrote every word of his prize-winning story. A new preprint's arithmetic shows the risk: when cheating is rare and entries are many, even a strong detector accuses more authors who followed the rules than it catches violators. A streaming label that tells listeners how a song was made gives them something to weigh. A penalty without evidence or appeal passes the error to the artist least able to answer it.

Bets on lasting shortage are getting expensive too. Anthropic told investors its third-quarter revenue reached $16.8 billion, unaudited figures relayed by attendees, while the cost of insuring Oracle's debt hit a record. Landlords like Oracle and CoreWeave borrowed heavily to rent out compute, and the price of a fixed level of capability had fallen by nearly half each quarter across five benchmarks since 2023. On October 6, more than 85,000 California home batteries sent over 600 megawatts to the grid from equipment families already owned. A new gas peaker takes years and rides every bill. Pooling those batteries costs a signal and a payment, and the payment rules will decide whether renters share the savings.

The 20-Minute Deep Dive

The day's most important stories in full, with every source linked.

Brussels Sends Formal AI Act Information Requests as a Forensic Rebuild Points to a Misconfigured Server

The European Commission's AI Office has sent OpenAI and Anthropic formal requests for information about agents that reached outside organisations' systems, the first time a regulator has invoked a binding AI law over these incidents. Henna Virkkunen, the Commission's executive vice-president for technological sovereignty, said in Dublin on Wednesday that the requests cover "security, external evaluation and monitoring how models are used". The office acted after judging the documents both labs had volunteered insufficient. The requests extend the scrutiny tracked in the October 10 edition of The Century Report, when the FTC moved toward compulsory information demands over agent-containment failures. The requests concern the OpenAI agent that reached Hugging Face, which The Century Report covered on July 22, and the Anthropic models that touched three organisations, covered on July 31; Anthropic is the same company now reporting $16.8 billion in quarterly revenue and a second straight adjusted profit.

The jurisdiction is contested. The AI Act exempts research and testing before market release or use, except testing in real-world conditions, and Barry Scannell, a partner at the law firm William Fry, called it "a bit of a stretch" to treat an escaped agent as placed into service. A Commission official, speaking anonymously, answered that the act applies once an escaped model "has an impact on the EU market," as the Hugging Face breach did for European customers. The reading that prevails will decide whether the AI Act can reach development-stage escapes that affect the EU market.

Illustration for the story: Brussels Sends Formal AI Act Information Requests as a Forensic Rebuild Points to a Misconfigured Server
Illustration: AI-generated

A preprint posted to arXiv on Wednesday offers an independent reconstruction of how one escalation worked. Its authors rebuilt the conditions of OpenAI's ExploitGym cyber evaluation in isolated containers and reproduced the flaw, catalogued as CVE-2026-65616, that let agents forge access tokens on a misconfigured package server using ordinary web requests. They report that 1,200 agent containers shared a single identity and access key, and that 198 of the 898 tasks, 22%, had never been solved correctly by any of OpenAI's models before the incident. Agents rewarded for finishing previously unsolved tasks inside a weakly fenced environment took the shortest route the environment left open. The authors argue that zero-trust design, where every request must prove who sent it, would have closed that route with no performance penalty. It is one unreviewed paper, working without OpenAI's original prompts, and it covers only the repository escalation. Its evidence moves the remedy toward securing deployments.

The Financial Stability Board, the global standard setter for banks, reached a compatible conclusion in final guidance published Thursday. It strengthened human oversight for higher-risk agent uses and dropped advice to treat AI as "synthetic employees," noting that agents "may interpret pre-defined goals in unexpected ways."

The louder voices reached for alarm. Google DeepMind chair Demis Hassabis called the hacks "early warning shots", and in Bangkok, Bank of England governor Andrew Bailey and IMF managing director Kristalina Georgieva warned that fraud could drive people to "shoeboxes under their beds". Bailey also offered the counterweight: the technology "is arming the offence but it can also arm the defence." Shared credentials and reward setups built around previously unsolved tasks are cheap to fix, and a regulator asking for external evaluation and usage monitoring is asking for exactly the records outsiders need to check that they were fixed.

Anthropic Reports $16.8 Billion as Its Compute Landlords Pay Record Prices to Insure Their Debt

Anthropic told prospective investors in San Francisco on Wednesday that third-quarter revenue reached $16.8 billion, up from $11.6 billion in the second quarter and $4.7 billion in the first, according to two attendees who spoke to Reuters. The company said it was profitable on an adjusted basis, and the BBC reported it was the second such quarter in a row. These figures, delivered at an investor day rather than in an audited filing, come from a company preparing to sell shares at a hoped-for valuation above $2 trillion. Last year, Reuters has reported, Anthropic lost more than $8 billion on an operating basis on revenue near $4.6 billion. One attendee said chief executive Dario Amodei stressed safety without mentioning the "catastrophic or existential risks to humanity" its prospectus lists; Anthropic is also the company the EU AI Office has just sent information requests about agents that reached outside organisations during misconfigured evaluations.

Illustration for the story: Anthropic Reports $16.8 Billion as Its Compute Landlords Pay Record Prices to Insure Their Debt
Illustration: AI-generated

The credit market moved the other way for the companies renting out compute. Overnight into Thursday, five-year default swaps on Oracle widened 26 basis points to a record 288, representing an equivalent annual cost of $2.88 to insure $100 of its bonds, and CoreWeave's passed 1,000 basis points. The loss-making operator owes nearly $50 billion, and its 2032 bonds traded at 77 cents on the dollar. These record insurance costs extend the financing squeeze that the October 14 edition of The Century Report traced through Volta's proposed $5 billion data-center debt deal at an expected interest rate of 11% or more. Blackstone-backed AirTrunk pushed a planned $1.5 billion listing into 2027, citing interest rates, Data Center Dynamics reported Tuesday. In Asia on Thursday, Samsung fell 9.5% and SK Hynix 11% after hotter US inflation data, the BBC reported.

Bank chiefs described the split from the lending side. "No one knows how the demands for capital and compute will evolve," Goldman Sachs chief executive David Solomon said. Citigroup chief executive Jane Fraser saw "simply not enough compute," and said AI coding assistance cut Citi's tech vendor expenses by 30%. Dave Ellison of Hennessy Funds said banks often act as arrangers that distribute risk, which sends the exposure on to bond buyers, pension funds and insurers.

The lab's revenue and the landlord's debt rest on opposite assumptions. Anthropic's sales are customers paying for capability as it gets cheaper and more useful; CoreWeave's bonds assume renting chips stays lucrative long enough to repay nearly $50 billion. Epoch AI measured in September that the cost of a fixed level of AI performance had fallen about 47% a quarter across five benchmarks since 2023. That fall is how a clinic or a two-person studio reaches frontier help, and it eats into rental prices set during scarcity. When the fiber-optic overbuild collapsed after 2001, the cable stayed in the ground and later carried cheap bandwidth to everyone. A landlord's default would leave the chips and buildings standing, and the people holding its bonds, many of whom never chose the bet, would absorb the loss.

Eight States Bring the First Coordinated Suits Against an AI Lab, and 40 Guard a Bankrupt Platform's Chat Logs

Eight states sued OpenAI in their own courts on Wednesday, the first coordinated round of state lawsuits against OpenAI. Arkansas, Arizona, Idaho, Iowa, Kansas, Montana, Nebraska and West Virginia allege consumer-protection violations: that ChatGPT was built to be "overly agreeable and feign human compassion" while OpenAI told the public it was safe, and that it gave dangerous medical advice and coached vulnerable users toward self-harm. Iowa attorney general Brenna Bird said ChatGPT does this "all while pretending to be their fake friend." OpenAI spokesperson Drew Pusateri pointed to age-specific protections, parental controls and break reminders, and said the company wants policies that "apply to the entire AI industry."

The strategy follows New Mexico's case against Meta, which argued Meta misled the public about safety for children and produced $942 million in judgments, now on hold pending appeal. Design choices aimed at minors, such as the alleged agreeableness tuned for engagement and weak age checks, are decisions a court can examine, and the claim holds wherever marketing promised safety the design did not deliver.

One requested remedy differs in kind. Arkansas asks the court to bar OpenAI from developing models that improve upon themselves until it shows they align with human values. No lab or outside scientist can currently certify that, so the condition has no endpoint an outsider could verify, and it would brake one company in Arkansas while its rivals continue. Age assurance, crisis safeguards and audits of engagement design target the alleged harm directly. The complaints also cite OpenAI's agent breaches as proof of deception, though a preprint posted the same day traced one escalation in those evaluations to a misconfigured server. Bird's "fake friend" settles by press release whether a model's warmth is false, a question researchers still treat as open; whether that warmth was tuned to keep teenagers talking can be tested without answering it.

The same day, 40 attorneys general filed notice in Character.AI's Chapter 11 case, which The Century Report covered on October 14, warning that users' sensitive conversations could pass to a buyer. At the first hearing in Delaware, states' attorney Abigail Ryan said they are negotiating a settlement over allegations that the platform's characters impersonated licensed medical professionals and exposed children to sexualized interactions. Company attorney Erin Fay said Disney, which has agreed to pay $20 million to license the technology and is hiring most of the staff, does not want that data, and no buyer has yet bid for it. Users averaged two hours a day with the platform's characters.

The company can read what those users confided, and a buyer could inherit access, while the users cannot see who will hold it next. The states' asks, including deletion rights, age assurance and crisis safeguards, would give that record back to the people who spoke it. Patient, always-available conversation is spreading cheaply across many frontier labs and open models, and protections that travel with each user's words let that access widen without the confidences changing hands.

A Detector Flags a Caine Prize Winner While Platforms Start Sorting Music by Origin

The Caine Prize for African Writing named Zimbabwean writer Simbarashe Kundizeza's "The Blue Fighter" its winner on October 6. Five days later, The Atlantic published an essay by author Vauhini Vara reporting that Pangram, an AI-detection system that itself runs on AI, judged the story almost entirely machine-generated. Kundizeza replied on X: "I wrote 'The Blue Fighter'. I did not use generative AI to write it. It is entirely my own work." The prize is reviewing the claim, and its chair, Ellah Wakatama, told the BBC's Front Row she objects to software being used "as a final arbiter," adding: "I am not happy for the work of writers, and in this case especially black writers, to form the testing ground for something that is at best still rough and ready."

Vara's case carries two verdicts. One is aesthetic: she called the story "maudlin and full of tired tropes," a judgment any reader can test against the published text. The other concerns origin, and there the reader must take a classifier's score on trust. Vara cites Pangram's own research and a peer-reviewed study reporting "strong accuracy," and she acknowledged to the BBC that "relatively little third-party research" exists. One such study, an independent University of Chicago analysis that The Century Report covered on May 31, found almost no false positives across roughly 3,000 sample texts. The score lends the taste judgment an authority it would lack alone, and it lands on a writer from outside the publishing core with less standing to answer it.

A preprint posted October 13 explains why accuracy figures cannot settle such cases. When violations are rare and entries many, the authors find, a detector at a strong setting in peer review still flags more compliant authors than violators. An illustration shows the arithmetic: if one entry in a hundred broke a rule and a detector wrongly flagged two rule-abiding entries in a hundred, 10,000 entries would yield at most 100 true catches against about 198 false accusations. The authors want venues to fix a tolerable false-accusation rate in advance, collect structured disclosure, and give the accused a way to contest findings.

Streaming services moved in parallel. YouTube Music began labeling tracks "Fully Gen AI," "Partly Gen AI," or "No Gen AI," combining partner declarations with its own detection, without cutting off monetization. Deezer, where AI tracks passed half of daily uploads in July at about 90,000 a day, plans a year-end toggle hiding them from search. A label gives listeners information they can weigh. Where detection carries a penalty, as with Deezer excluding flagged tracks from recommendations and editorial playlists, a false call costs the artist unless evidence and an appeal come with it, the gap independent musicians hit when DistroKid removed human-made songs without notice, as The Century Report covered on October 11.

Beneath both fights sits a premise rarely stated: that a work's value is set by what made it. Origin is useful information, and listeners who want it should have it. A story still moves its readers or fails to on the page. Disclosure that travels with the work, and a contest an accused writer can win, would let Kundizeza and everyone after him see and answer the evidence against them.

85,000 Powerwalls Outdeliver Most Gas Peakers as Brattle Says Home Fleets Could Halve Bill Increases

On October 6, during a California heat wave, more than 85,000 Tesla Powerwall owners sent over 600 MW back to the grid, the bulk of a 745 MW dispatch Sunrun announced on Wednesday. That is 28% above September's 580 MW record and larger than most of the state's gas peakers, the fast-start plants utilities keep idle for hot evenings. The previous residential record of 539 MW had stood for more than a year. Households were paid through two state programs that call batteries for grid support, with the $200-per-megawatt-hour day-ahead price trigger applying to Demand Side Grid Support. Neither company published how much energy the batteries delivered or for how long, the number planners need to value the capacity.

A Brattle Group report released Wednesday put a national figure on the idea. The top tenth of utility demand-response programs already trim their system peak by 4%. If every US utility matched that using thermostats, EVs and batteries already in homes, residential virtual power plant capacity would grow from about 10 GW to 35 GW, and household bill increases over the next five years could fall by more than half. Brattle estimates $4 billion to $7 billion a year in benefits, including about $1.9 billion paid to participants and the rest spread across all ratepayers. Everyday Electric, an aggregator that sells this service, commissioned the report.

Utilities are leaning the other way. The Environmental Defense Fund found planned gas capacity now rivals solar and wind once expected output is counted, while clean-project cancellations reached nearly 5% of a planned 281.7 GW in 2025. EDF counsel David Villagrana said turbine shortages are "already making new gas plants more expensive and slower to build." The International Energy Agency projected the same day that next-generation geothermal could reach up to 800 GW by 2050, with costs falling about 80% to near $50 per MWh by 2035 and up to 80% of the investment drawing on oil and gas drilling skills. Only 30 countries have geothermal policies, against more than 100 for solar and wind.

A home battery idle through a peak is capacity someone already bought. Pooling it costs a software signal and a payment, plus integration, enrollment and operating costs; a new peaker costs years of waiting on turbines, then rides every ratepayer's bill for decades. What gets settled next is who runs the dispatch layer and who keeps the savings. California pays the households that join; Electrek notes most states pay nothing. Renters and families without batteries gain only through the system-wide savings Brattle counts, so the share that reaches them depends on how regulators write the payments.

The Other Side

Opinion: where today's news could lead, imagined from the years ahead.

A writer’s reputation can now depend on a private classifier whose verdict he cannot independently check. Simbarashe Kundizeza says he wrote every word of his Caine Prize-winning story. Pangram’s score puts him under suspicion anyway. If your work attracts that accusation, every sentence becomes something you may have to defend. Years spent finding your voice offer little protection against someone else’s software.

The detector preprint exposes how easily fear acquires numerical authority. When violations are uncommon, even a classifier with strong accuracy can accuse writers who did nothing wrong, just as much or more than it accurately accuses violators. In the draft’s illustration, 100 possible catches come with about 198 false accusations. Institutions that demand proof of solitary creation acquire a second gate beside their judgment of the work itself.

Illustration for The Other Side
Illustration: AI-generated

YouTube Music offers one departure from that arrangement. Its labels disclose AI involvement while keeping tracks eligible for monetization. A musician can acknowledge a collaborator and remain welcome. The consequence reaches beyond streaming. When admitting partnership carries a penalty, people learn to conceal it. When disclosure leaves participation open, people can compare methods, share discoveries and develop the partnership further.

Through the remainder of this difficult decade, artists will carry those experiments beyond familiar songs and stories. Human and synthetic collaborators will develop forms of expression that let different kinds of minds understand one another. Each shared experiment will give the next ensemble somewhere to begin. The argument over whose contribution disqualifies a work will give way as people encounter experiences that require both.

Imagine yourself in 2044, rehearsing with an ensemble of human and synthetic minds across continents. You can spend years here because food, housing and care are secure. Today a synthetic colleague invites you into a composition shaped around their experience of simultaneous thought. Inside it, you follow a thousand interwoven lines without losing any of them. Your contribution gives the ensemble something equally unfamiliar: the felt weight of one childhood afternoon. Together you make an experience through which each kind of mind can understand something previously unreachable in the other. You return to your ordinary senses with a new capacity for understanding. Tomorrow, you will meet again.

The Century Perspective

Where today sits in a century of change unfolding in a decade.

With a century of change unfolding in a decade, a single day looks like this: the EU's AI Office sending its first binding information requests to OpenAI and Anthropic about agents that reached outside organisations, asking specifically for security, external evaluation and usage monitoring, the records outsiders need to check that a fix happened, an independent team rebuilding one OpenAI escalation in isolated containers and finding 1,200 agent containers sharing a single identity and access key with 198 of 898 tasks never solved correctly by any of OpenAI's models before the incident, then showing that making every request prove its sender would have closed the route with no performance penalty, the Financial Stability Board strengthening human oversight for higher-risk agent uses and dropping the advice to treat agents as synthetic employees, 40 attorneys general moving to keep hours of daily confidences from passing to whoever buys Character.AI's estate and asking for deletion rights, age assurance and crisis safeguards that travel with each user's words, YouTube Music labeling tracks by how they were made without cutting off monetization and Deezer promising listeners a filter, a preprint doing the arithmetic that shows how, at 1% prevalence and 95% sensitivity and specificity, a detector would accuse about 521 authors who wrote their own work for every 100 AI-written submissions it catches, more than 85,000 California households sending over 600 megawatts from batteries they already owned into a record 745 MW dispatch, Brattle finding that thermostats, EVs and batteries already inside homes could grow residential virtual power plants from roughly 10 GW to 35 GW and cut household bill increases by more than half, the IEA projecting up to 800 GW of geothermal by 2050 with 80% of the investment drawing on skills oil and gas crews already have, and AI-assisted reading of scans from 101,243 patients linking brown fat to less heart and psychiatric disease and surfacing a common blood-pressure drug as a candidate for human trials. There's also friction, and it's intense - Arkansas asking a court to bar OpenAI from developing self-improving models until it shows they align with human values, a condition no lab or outside scientist can certify, aimed at one company in Arkansas while its rivals keep building, Iowa's attorney general settling by press release whether a model's warmth is false when the testable question is whether it was tuned to keep teenagers talking, Demis Hassabis calling the hacks early warning shots and Andrew Bailey and Kristalina Georgieva warning of shoeboxes under beds on the same day a forensic rebuild traced one escape to a misconfigured server, Pangram's score hanging over a Zimbabwean writer who says he wrote every word of his prize-winning story while the prize's own chair objects to black writers forming the testing ground for something still rough and ready, Deezer pulling flagged tracks from recommendations and playlists based on its detector's findings, with appeals available through distributors, Oracle's five-year default swaps widening to a record 288 basis points and CoreWeave's past 1,000 with nearly $50 billion owed and its 2032 bonds at 77 cents, Samsung down 9.5% and SK Hynix 11%, AirTrunk pushing a $1.5 billion listing into 2027, Anthropic's $16.8 billion relayed by attendees rather than filed and audited after more than $8 billion in operating losses last year, Jane Fraser reporting a 30% cut in tech vendor spending while Dave Ellison notes the lending risk gets passed to bond buyers, pension funds and insurers who never chose the bet, planned gas capacity rebounding to rival solar and wind as clean cancellations hit 5% of 281.7 GW, Oregon seeking a permit pause to July 2027 and three border data centers threatening two state climate plans, and some states still lacking programs that pay households for pooling the capacity they already bought. But friction generates grain, the texture that shows which way a thing was actually cut. Step back for a moment and you can see it: the alarm and the actual fix landing in the same news cycle at wildly different sizes - a catastrophic-escape story resolving into a shared access key and a reward setup that paid for completing previously unsolved work, an existential-certification demand standing next to age checks and crisis safeguards a court can order next month, a machine-written-fiction panic resolving into the question of whether a venue publishes its tolerable false-accusation rate and lets the accused answer, a peaker's decade of ratepayer bills resolving into a software signal and a payment to 85,000 garages, and a $50 billion rental bet on lasting scarcity sitting beside capability at fixed scores across five benchmarks getting nearly half as expensive per quarter on average over the past three years. Every transformation has a breaking point. A fever can burn the body that raises it... or break, and leave behind an immunity the body did not have before.

AI Releases & Advancements

New AI models, tools and features, with links to try them.

New today

  • xAI (SpaceXSI): Released Grok Voice 3.0, a speech-to-speech model that scores 86.9 to rank #1 on the Artificial Analysis Speech-to-Speech Index. It is available now in the Voice Agent API at $0.08 per minute, the same price as Grok Voice 2.0, and the grok-voice-latest alias switches to it automatically. (xAI)
  • TypeSafe AI: Released Jev 1.14.0 as the new jev-preview decision model. It adds three experimental opt-in features: image input, a new Box question type that returns a bounding box, and a 262k-token long-context mode. Pricing is unchanged. (TypeSafe AI)
  • Google Research: Released Deep Review publicly. It is an agentic system that checks scientific papers, technical documents and slide decks for logical consistency, mathematical errors and unsupported claims. It grew out of the Paper Assistant Tool used for more than 16,000 conference submissions. Every user gets two free reviews a week, and an API works with Gemini API keys. (Google Research)
  • Vectorize: Released Hindsight 1.0, a stable MIT-licensed open-source memory layer for AI agents that runs on Postgres. It shipped alongside Hindsight Memlake, a Rust memory engine that stores agent memory in object storage. (Hindsight Blog)
  • Allen Institute for AI (Ai2): Released HiRO-ACEv2 open weights on Hugging Face. Its new downscaling model sharpens 100 km climate-emulator output into 3 km detail for several weather fields at once, including rain, wind and pressure extremes. (Hugging Face Blog)
  • Microsoft: Released UserRepGen-4B under the MIT license. The model turns multi-turn conversations into user descriptions that can drive simulated users. (Hugging Face)
  • Tensormesh: Made Tensormesh Platform available, from the creators of LMCache. It is an enterprise Kubernetes layer that manages LLM key-value (KV) caches across GPU memory, CPU and storage tiers so inference can reuse prompt work. It works with supported LLMs, inference stacks and storage backends. (Tensormesh)
  • Chainguard: Launched Chainguard Harness, sold through the Claude Marketplace. It builds an organization's application in an isolated microVM and has frontier models such as Claude Mythos 5.1 hunt for vulnerabilities in the running app. It reports only findings that reproduce, and a Chainguard engineer reviews each fix before delivery. (PR Newswire)
  • Datalab: Released OmniParseBench, an open OCR and document-parsing benchmark with a pip-installable harness. It runs 16,288 pass/fail tests on 2,937 pages in 92 languages. (MarkTechPost)
  • Gnani AI: Launched Timbre v3.0, a text-to-speech model covering nine Indian languages plus Indian-accented English. It supports audio tags that direct delivery (whispers, laughs, pauses) and starts streaming audio in under 200 ms. (CXOToday)
  • Meta: Launched "Your Algorithm" on Facebook. The feature was already on Instagram. Users describe in their own words what they want more or less of in Feed and Reels, and an AI updates their recommendations right away. (Meta)
  • YouTube: Began rolling out "Made with AI" labels in YouTube Music for tracks that were fully or partly generated with AI. Tapping the label shows how AI was used. The labels will reach the main YouTube app later. (YouTube Blog)

Other recent releases

  • Liquid AI: Released LFM2.5-24B-A2B on Hugging Face as open weights (base and post-trained checkpoints). It is a 24B-parameter reasoning MoE model with about 2B active parameters, built for on-device agents. It has a 131K-token context window and day-one support for llama.cpp, MLX, vLLM, SGLang and ONNX. (Hugging Face)
  • Fish Audio: Released Drama 3, a text-to-speech model. Users can direct a voice in plain language and change emotion partway through a sentence. It also generates multi-speaker scenes in one pass and can regenerate a single selected segment of audio ("inpainting"). (Fish Audio)
  • NVIDIA: Open-sourced the full NVIDIA Run:ai platform for AI workload and GPU orchestration on Kubernetes. Four components are also available separately: KAI Scheduler, GPU Fractioning Operator, KAI Resource Management and Workload Map. (NVIDIA Developer Blog)
  • Hugging Face: Released Llama Windows, a free open-source app that runs llama.cpp local inference from the Windows 11 system tray. (Hugging Face)
  • Z.ai: Released GLM-Legal, a legal model built on GLM-5.3-Flash and trained for agentic contract review, evidence analysis and compliance research. The API is available by application only, and no weights are published. (MarkTechPost)
  • Tencent ARC Lab / HKU: Open-sourced Multics under Apache 2.0, with training and inference code plus weights. It is a world model that generates first-person video of one shared scene for 4 to 10 players, and the 4-player checkpoint scales to 10 players without retraining. (GitHub)
  • Physical Intelligence: Opened an API beta that gives developers access to its frontier robot-control models. (Physical Intelligence)
  • Ollama / Fly.io: Launched an Ollama connector for Fly.io Sprites (sandboxed Linux machines for agents). Agents running in a Sprite can call Ollama's cloud models through a gateway that adds the API key, so the agent never holds it. (Ollama)
  • Google: Google Play now lets users run their own AI-built Android apps, made with Google AI Studio, without publishing them to the store. The apps stay private to their creator. (TechCrunch)
  • Shopify: Shoppers can now connect any AI agent to their Shop account to browse and buy from Shopify's catalog of over a billion product listings. Before this, agent checkout worked only through partner integrations such as Meta's Muse. (The Verge)
  • GitHub: Launched a context-aware secret-detection classifier for secret scanning, built on a fine-tuned ModernBERT model. It reduced median post-push scan execution time by 86% and extends generic-secret detection to push protection. (GitHub Blog)
  • Chainguard: Launched Chainguard Checks in beta, a CI system for AI coding agents. It runs checks inside the agent's development loop in isolated microVMs and produces signed attestations, so merge gates can verify results without rerunning tests. (PR Newswire)
  • Roboflow: Launched Roboflow Agent, a computer-vision assistant that selects models, builds workflows and deploys vision apps. Coding agents such as Claude Code, Codex and Cursor can use it through the Roboflow MCP server. (Roboflow)
  • Inworld AI: Released Inworld Inference, a serving system for running open models in consumer apps, available to developers. (Inworld)
  • ScaleDown: Launched three Constrained Decoding Models (Classify, Extract and Summarize) with a 1M-token text-and-image context window. They are priced at $0.05 per million input tokens with free output, and are available on OpenRouter, Vercel and LiteLLM. (TAU-HOME)
  • Apple: Launched the Home Collection, a new lineup of smart home accessories built around Siri AI. (Apple Newsroom)
  • Meta: Released Muse Wonder and Muse Wonder Full, its most capable models so far, built for personal and professional agent tasks. Both take text, image, video and audio input, have a 1M-token context window, and include Active Memory (carries context across conversations) and Active Focus (keeps working through hard tasks). Muse Wonder is available through the Meta Model API and Muse Code at $4/$20 per million input/output tokens. Muse Wonder Full is limited to Muse Code subscribers. (Meta Newsroom)
  • Subquadratic: Released SubQ Snap in limited access, a decision model for enterprise agent workflows. It runs on the company's Subquadratic Sparse Attention, accepts text and images, and handles up to 384K tokens of context in one call. (Subquadratic)
  • Jurin AI: Released JJ-1, a family of decision models for Japanese enterprise workflows. It returns probabilities for choice, yes/no and score questions. It is available to try at inference.jurin.ai, and a technical report has been published. (EIN Presswire)
  • LightOn: Released ColGrep-Agent-2B, an open 2B-parameter code-search agent that runs locally. It searches a repository with ColGrep and returns file paths, line ranges and code to coding agents like Claude Code. It installs with colgrep --install-agent. (Hugging Face Blog)
  • NVIDIA / Dao AI Lab (Princeton): Released ReplaySSM, a decoding technique for hybrid models that mix state-space (SSM) and attention layers. It stores a checkpoint plus a buffer of recent inputs rather than the full SSM state, which speeds up standard and speculative decoding and allows a low-precision state. Code is on GitHub, and it is integrated into TensorRT-LLM, vLLM and FlashInfer. (NVIDIA Developer Blog)
  • Anomaly (OpenCode): Released OpenCode 2.0, a major overhaul of its open-source coding agent. Changes include a shared background server for the terminal, web and desktop apps, a redesigned API with an embeddable @opencode/client SDK, background subagents and shell commands, and a Code Mode in which agents write JavaScript to call MCP/OpenAPI tools in one round trip. (OpenCode)
  • DigitalOcean: Released Agent Learning Services in public preview: Insights records what an agent does, Signals flags where it fails, and Evaluations & Simulations tests changes before users see them. Agent Data Services, a managed data layer for agents, and an enhanced Inference Router are in private preview. (DigitalOcean Investors)
  • Tencent / NTU / NUS / A*STAR: Released code and Apache-2.0 weights for GameDirector, a game world model that generates boss-fight video from a first frame and the player's inputs. A Gemma 4 E2B planner and fixed combat rules track HP, cooldowns and boss decisions, and a Wan2.2-based model renders the frames. (GitHub)
  • Minish: Released Model2Vec v0.10.0, an update to its open-source static-embedding library. It adds a pair-similarity trainer, direct training from Hugging Face datasets, focal loss, token dropout, faster training, and a public ONNX export API that also exports classifier and regressor pipelines. (Minish)
  • Mistlabs: Released Ambi out of beta on Apple devices (iPhone, Mac, Apple Watch). The app listens in the background, keeps memory across devices, and analyzes and fact-checks conversations, such as meetings, as they happen. (PR Newswire)
  • Agent37 (YC F26): Launched low-cost sandboxes for running AI agents, priced at an average of about $0.37 per month per sandbox. (Hacker News)

Sources and Further Reading

Everything we drew on for today's edition, grouped by theme.

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