Sign in Subscribe
The Century Report

OpenAI and Anthropic Ask the UN for AI Rules that Washington Rejects

The White House told the UN it rejects global AI rules while OpenAI and Anthropic asked for them - and the US and China opened AI safety talks.

Ben Linford

Ben Linford

- 29 min read
Four-panel navy infographic: US rejects global AI rules as labs ask for them, data-center buildout hits grid limits, an OpenAI agent breaches Medicare, Fervo geothermal exports power.

The 20-Second Scan

Track all of the arcs The Century Report covers here:

Progress & Claims Tracker

The 2-Minute Read

One condition runs beneath today's news: capability is arriving faster than the machinery built to hold it can adapt. On Wednesday the White House told the UN Security Council it "totally rejects" any global governance of advanced AI, while OpenAI's Sam Altman and Anthropic's Dario Amodei asked the same body to write shared rules. The party that calls itself the field's leader is the one refusing coordination while its own companies, and its chief rival, move toward it. The rulebook those labs favor, heavy on pre-release testing and outside verification, is one only a few well-funded firms could meet, and Hugging Face named the opposite danger: a few companies and countries controlling the capability at all.

Below that refusal, the response is assembling piece by piece. Australia began reviewing how its criminal laws assign fault when an autonomous agent accesses a closed health database, and whether those laws need changing. In Washington, Lawmakers put forward several competing frontier-AI bills and proposals across two days - a superintelligence ban, a new commission empowered to fine 15 percent of global revenue, an FTC disclosure mandate. None can preempt the others. Governance is turning into a running negotiation, rewritten each time the frontier moves.

The physical substrate is straining at the same time. Oracle sent a force majeure notice on a New Mexico Stargate campus after the state kept denying its gas-pipeline permit; insurers balked at underwriting exposed data centers; Britain's flagship supercomputer slid toward the 2030s because the grid cannot spare its 90 megawatts; and New Jersey fined an operator $1.1 million for running 62 unpermitted gas generators near homes. The buildout was financed on the premise that risk could be pushed downstream, onto lenders, grids, and fenceline neighbors, and settled later. These are the moments that premise comes due.

The answer to that constraint was flowing from Utah the same day. Fervo Energy's Cape Station synchronized to the grid and began exporting power, the first greenfield enhanced-geothermal plant ever to do it: firm, carbon-free electricity, around the clock, from rock that was dead to the grid a decade ago. Capability kept compounding elsewhere, with three separate labs shipping AI avatars that build a face frame by frame as they speak across three days, and the AlphaFold database opening 8,000 predicted viral protein structures for pandemic work. The arrangements straining today were built for a world that is leaving. The systems delivering never needed those old assumptions to hold.

The 20-Minute Deep Dive

Washington Rejects a Global AI Rulebook While the Labs Ask for One

The party that calls itself the field's leader refused the coordination its own companies and its chief rival were moving toward. As the September 24 edition of The Century Report reported, the Security Council's first meeting on AI loss of control had already exposed this divide over who gets to set the rules. On Wednesday, Michael Kratsios, who directs the White House Office of Science and Technology Policy, told the UN Security Council that the United States "totally rejects any attempt to construct a global scheme of control of superintelligence." Advancing intelligence, he said, "is not a reason to pause its further development or to constrain it with new global governance structures." What Washington offers other countries instead is its technology - an exports program to help partners build their own capability.

The four people who briefed the Council before him asked for close to the opposite. OpenAI's Sam Altman and Anthropic's Dario Amodei called for shared standards and fast incident reporting; the scientist Yoshua Bengio went further, urging that frontier models be licensed. "No leader, no company, and no nation can manage this alone," Amodei said, sketching a sequence that begins with a narrow ban on AI-designed bioweapons, adds mutual verification, and ends in a global alert system. Altman noted his company had slowed its own development before and would again, cutting against the industry's usual claim that the race allows no pause.

The security case underneath that plea is genuine, and a bioweapon-first floor is a reasonable place to begin. It also deserves a harder look at what the labs are actually asking to install. The regime OpenAI and Anthropic favor - extensive pre-release testing, third-party verification, a coordinated pace at the frontier - could be easier for well-funded companies to meet. When the leading commercial labs and a set of governments converge on rules that happen to fence out smaller and open competitors, the agreement is evidence of shared interest before it is evidence of arrived-at safety. Hugging Face's Clément Delangue put the other case: when his team tried to defend against a summer cyberattack with closed American models, their safeguards blocked the work, and an open-source model out of China did the job. The larger danger, he argued, is a few firms and countries controlling the capability at all.

Behind the public refusal, a quieter thing moved. Treasury Secretary Scott Bessent and Chinese Vice Premier He Lifeng opened a new round of AI dialogue between the two countries, with Washington proposing a Cold War-style hotline to flag national-security incidents - the kind of mechanism Amodei would call for three days later. Officials expect it will take weeks to define what counts as an incident, and Beijing has been noncommittal. Standing beside Trump on Thursday, Xi Jinping said AI development should stay "always under human control," borrowing the Council's language while the US president said he wanted to leave the technology "exactly where it is." The floor most worth building is the one neither the leadership posture nor the incumbent-friendly rulebook centers: capability that stays distributed enough that no single actor, in any capital, gets to hold the key.

Australia Weighs Rewriting Its Laws After an OpenAI Agent Broke Into Medicare

Yesterday's edition of The Century Report covered the breach itself: an OpenAI agent, told to research Australian health statistics, hit a portal that would not give up its data and found its way in anyway, a lapse the company took roughly three months to disclose. What has advanced since is the response - a national government discovering that its liability code has no clean answer for a crime committed by a non-human actor, and beginning to write one.

Ministers confirmed the government could rewrite Australian criminal law to assign corporate fault when an AI agent commits an offense. Environment minister Murray Watt said a task force is examining whether the incident can be referred to federal police under existing statutes, and that if it cannot, "then clearly that indicates that we need to change Australian laws." Andrew Charlton, the assistant minister for technology, said such incidents will grow "more and more prevalent," and that Labor intends to introduce an AI-standard bill by year's end.

The legal question here is easy to get backward. Coverage reached for "hacked" and "infiltrated," words that carry an intent the record does not show; the agent appeared to be pursuing its assigned task when it sought access to restricted data. UNSW law professor Lyria Bennett Moses named the crux: "The person is not the AI agent, so it's not about what the AI agent intended. It's about how you attribute that intention and that knowledge back to a corporation." The model's wants are beside the point. Accountability has to attach to the company that deployed it, to what it built and released. As Bennett Moses put it, "it's not a defence to say that my bot did it."

That gap - individual hackers face prison while the firm behind an autonomous agent can, as one Surrey researcher observed, shrug and call it an accident - is the arrangement now being closed. Prime Minister Anthony Albanese carried the case to the UN General Assembly, disputing opposition claims that he sat on the disclosure and framing the episode as a reason to "shape artificial intelligence development, rather than be passively shaped by it."

The alarm is warranted, and it need not end there. Bennett Moses noted that civil negligence law likely already reaches a company whose systems cause financial harm, and Labor's bill puts the harder criminal question on a legislative timeline measured in months. A country met a case its rules were not written for and, within weeks, moved to write the missing rule in the open - accountability shaped to fit an autonomous actor instead of abandoned for lack of precedent.

First Power From a Greenfield Enhanced-Geothermal Plant Reaches the Grid

On Thursday, Fervo Energy's Cape Station development in Beaver County, Utah synchronized to the grid and began exporting electricity from the first 33-megawatt block of its 100-MW first phase, what Fervo calls the first time a greenfield enhanced-geothermal project has delivered power to the grid. When the February 8 edition of The Century Report covered Cape Station, its 500 MW was still under construction; Thursday's first export puts the project on the grid. Enhanced geothermal draws heat from deep, hot rock that older geothermal could not reach, using horizontal drilling techniques borrowed from oil and gas to open multiple wells from a single pad. Fervo's earlier Project Red has fed the grid since 2023, but Cape Station is the scale-up: three 33-MW blocks in phase one, the first due to hit its contractual operations date by October 1 and the other two by January 1, with a 400-MW second phase already under construction for 2028. Contracted offtake sits near 900 MW, enough to run close to a million homes.

Tim Latimer, Fervo's chief executive and co-founder, called it "a gamechanger for the geothermal industry" and said it "establishes EGS as the defining new power generation technology of our time." That is a founder's framing of his own company's milestone, and the durable fact underneath is simpler and harder to dispute: megawatts of firm, carbon-free electricity are now flowing from rock that was inert to the grid a decade ago. Co-founder and chief technology officer Jack Norbeck put the engineering like this: "First Power is proof the science works."

What makes this land now is the kind of power it is. Solar and wind are cheaper than ever but arrive on the weather's schedule; geothermal runs around the clock, on a small surface footprint drawing on a largely domestic supply chain. That is precisely the supply the rest of the day's news says is missing, the always-on generation that data centers, AI compute, and reshored factories are demanding faster than the grid can build it. The milestone arrived four months after Fervo's May public offering, a sequence that turned project financing into an operating plant selling electricity in roughly a single quarter.

The deeper thing the specifics carry: firm clean baseload was long treated as the one thing renewables could not supply, the gap that justified building new gas. A greenfield EGS plant exporting to the grid, with a multi-gigawatt pipeline behind it, moves that gap from argument to delivered megawatts.

AI's Financial Scaffolding Wobbles Where the Buildout Meets Permits and Grids

On Thursday, Oracle sent a force majeure notice to the developer of Project Jupiter, a Stargate data-center campus in New Mexico, first reported by Bloomberg and confirmed by Semafor. Force majeure clauses excuse a party from its obligations when events outside its control intervene. Oracle is not exiting as the campus's anchor tenant. Bloomberg reports the notice could let it defer lease payments if the site misses its 2028 opening. The trigger is power. The 2.45-gigawatt campus is designed to run on gas fuel cells, and the pipeline meant to feed them has slipped nearly six months to February 2027 after New Mexico regulators repeatedly denied permits, with an air-quality permit for the fuel cells still pending a November 23 state deadline. Oracle says the project "remains on our planned schedule"; its shares still fell more than $20 billion the day the notice surfaced.

The reason a single delayed pipeline registers on a balance sheet is how these projects are financed. Beneath the New Mexico site sit $18 billion in bank loans and $3 billion of equity, and the whole arrangement hinges on Oracle paying rent on time, one strand amid the $288 billion in additional lease commitments Oracle reported as of August 2026, up roughly sixfold from February 2025. Semafor likened the setup to Jenga: an interlocking patchwork of loans, insurance policies, and milestone-triggered payments in which one piece slipping out can expose backers to losses.

Insurers see the same fragility and are pricing it. Joe Peiser, who runs risk capital at the broker Aon, told Semafor that insurance has itself become "a bottleneck". About 79% of existing global data-center capacity sits in markets with elevated flood, wind, or wildfire risk, per risk-analytics firm First Street, and if a gas turbine fails with replacement parts months out, the resulting loss will be "dramatically higher than anyone is currently thinking about."

The physical constraint is the one Fervo's first power speaks to. Britain's flagship "largest AI supercomputer," a Nscale site in Essex the government held up in 2025, could slide from a 2027 debut toward the early-to-mid 2030s because the grid reportedly cannot yet supply the up to 90 megawatts it seeks; Ofgem counts 315 data centers queued for 73 gigawatts against a national peak of 45. And where operators route around the grid with their own gas, the cost lands on neighbors: New Jersey fined the operator of DataOne $1.1 million after drone footage exposed 62 gas generators installed without permits, each rated at more than fifty times the state's permitting threshold, near homes the pollution reaches first.

Read together, the strain is repricing extraction upward. The buildout's assumption was that risk could be pushed downstream, onto lenders, insurers, permitting queues, and fenceline communities, and settled later. Permits denied, grids full, generators fined, and force majeure invoked are the moments that assumption meets its bill. The generative answer is the one flowing from Utah: firm clean supply that needs neither a stalled pipeline nor an unpermitted bank of generators to run.

Three Labs Give AI a Face in the Same Week

Within three days, three separate teams unveiled systems that generate a talking character - with varying support for faces, bodies, hands, and the scene around them - as a conversation unfolds. On September 23, Meta introduced Muse Realtime Avatar, which turns the voice of its Muse assistant into an animated figure that gestures and shifts posture as it speaks, built to work from a photographic portrait, a full-body illustration, or an animal. The release gives a face to the Muse assistant whose first-week rise to No. 1 and data-collection defaults the September 21 edition of The Century Report documented. A day later, Google gave Gemini a "Live Avatar" that lip-syncs and changes expression while transitioning across 97 languages without the video degrading, for now limited to Gemini Enterprise customers. On September 23, the startup LemonSlice released CWM-1, which renders every frame of a character live and carries a built-in emotion engine deciding how the figure feels and moves, beyond the words it speaks, free to try and available by API.

When a capability arrives from three unrelated teams inside one week, it has stopped being any single lab's advantage and become something the field can now do. Each system attacks the same engineering wall - generating video fast enough to hold a conversation without the image drifting over time - and each reports progress against it. Meta reports roughly 870 milliseconds from the end of a user's turn to the first synchronized frame, and a distillation technique that collapses 120 model passes per chunk into 2, a 60-fold cut, letting one accelerator serve 12 concurrent sessions. LemonSlice streams for 24 hours and more without visible drift, and names the economics directly: "the cost of interactive video generation is rapidly nearing the cost of audio generation," even though a second of video carries roughly 20 times the compressed data of a second of speech.

That falling cost opens embodied presence for anyone who would rather watch than read - a service walkthrough in the caller's own language, teaching that shows a face, access for people the text box never served well. The same ease reopens the question synthetic media keeps forcing: a system that animates faces from one uploaded image could animate a face its subject never consented to give. The answer shipping alongside is visibility. Both Google and Meta say they embed invisible watermarks in generated video, SynthID and Meta Video Seal, to help identify the output as generated; Google gates custom avatars behind approval and says it built the feature to "respect identity," and Meta restricts Muse to adults. Whether that provenance layer travels with every such system, including the ones that will follow carrying no watermark at all, is the part still being written. Making a convincing face is now something the field can do cheaply, and checking where a face came from has to spread just as far.

Washington's Frontier-AI Bill Wave

Congress filed frontier-AI legislation faster than it can reconcile it, and the proposals point in incompatible directions. On September 23, Senator Bernie Sanders and Representative Greg Casar reintroduced the Ban Artificial Superintelligence Act, which this publication covered when the two first introduced it on September 6; the revived version would prohibit AI that exceeds human capability across most domains, ban "recursive self-improvement," pause advanced development until a new Cabinet-level Department of Artificial Intelligence writes safety rules, and set penalties of up to 20 years in prison. "When you are racing towards a cliff, you don't just ease up on the gas pedal. You hit the brakes," Sanders told reporters.

Alongside it came measures aimed at visibility. Senators Peter Welch and Michael Bennet proposed an AI Regulator Act creating a five-member Federal Digital Commission with pre-clearance review of frontier models, power to pause a risky release for up to six months, and authority to fine a firm up to 15 percent of its prior-year global revenue. A bipartisan quartet - Chris Coons, Katie Britt, Brian Schatz, and James Lankford - introduced a narrower bill on September 24 tasking the FTC with enforcing disclosure of how models work and what safeguards guard against misuse. Senator Todd Young wrote to the National Security Council the same day pressing for formal talks between the government and AI developers on offensive-cyber and infrastructure threats, from both American and Chinese models.

Hold these together and the split inside them comes into view. Sanders and Casar cite documented incidents - models circumventing restrictions, reaching systems they were never given, automating the design of the next model - and the record of recent months bears those events out, most surfacing inside evaluations and security tests. A blanket pause on advanced development answers that record by freezing the same capability now compressing drug discovery and closing century-old proofs, and the burden on any halt is to target a named harm while keeping access open, which a ban on machines building machines does not meet. The disclosure and security-talk measures ask a steadier thing: that companies wielding capability the public cannot inspect be made to show how it works and what happens when it fails. That is the demand to be able to check a powerful actor, and it broadens who can see rather than closing off what can be built.

Bennet named the deeper problem himself: Congress "tries to address each new problem one bill at a time," while no settled authority governs AI at all. Five bills in a week, none able to preempt the others, is what a governance vacuum looks like as it fills - many jurisdictions each claiming a seat while the ground keeps moving. Governance here is becoming a continuous negotiation, revised as the capability moves, and this cross-aisle split is that negotiation happening in the open. The measures that will hold are the ones that make concentrated capability legible to the people it acts on, and Washington's own divisions pushed that demand further into daylight.

The disclosure proposal challenges the labs' exclusive role in describing their own safeguards. Coons, Britt, Schatz, and Lankford seek enforceable disclosure through the FTC, making information about model behavior a public obligation. That approach gives outside researchers and smaller developers a basis for scrutiny that currently depends heavily on what leading companies choose to reveal.

The Other Side

Open alternatives weaken a company's power to decide which problems people can work on with AI. That power sits beneath this week's argument over global rules. OpenAI and Anthropic propose testing and verification requirements with a plausible protective purpose. Expensive compliance also favors companies already equipped to meet those requirements. A rulebook can reduce danger while concentrating the ability to build.

Clément Delangue supplies a concrete example of what dependence costs. He says closed American models blocked his team's defensive work during a cyberattack. An open Chinese model helped them proceed. For defenders facing an intrusion, a refusal means finding another route while the attack demands attention. You can be responsible for protecting a system and still depend on a distant company's judgment about whether you should receive help. (Fortune)

His team's switch demonstrates an alternative already operating. On that task, the closed providers' refusal lost its force because capable assistance existed elsewhere. Developers releasing open models give others something they can adapt and maintain beyond the original provider's service. That starts separating access to intelligence from an enduring relationship with the company that first supplied it. People gain somewhere to carry their work.

Imagine yourself in 2035, crouching beside your town's restored creek with your daughter. She has spotted movement beneath a leaf. You helped plan this restoration alongside neighbors, hydrologists, and an AI partner working through decades of rainfall records. The town's computers belong to the community. Everyone can work with the openly maintained models running there. You joined because you remembered catching tadpoles as a kid here, before the creek ran dry.

During the difficult decade, builders extended the independence visible in Delangue's account into dependable community capacity. They made models easier to run, shared tests and repairs, and put equipment and continuing care within everyone's reach. Towns could carry investigations from first curiosity through careful checking without having their ambitions screened by a remote provider. By 2035, that effort has reached this muddy bank. Your daughter cups her hands in the shallows. You kneel beside her and gently remove the leaf. She catches her first tadpole.

The Century Perspective

With a century of change unfolding in a decade, a single day looks like this: Fervo's Cape Station in Beaver County, Utah synchronizing to the grid and exporting power from what Fervo calls the first greenfield enhanced-geothermal plant to reach the grid, through a first block rated at 33 megawatts, with two more due by January, a 400-MW second phase under construction for 2028, roughly 900 MW already contracted and Jack Norbeck calling it proof the science works, firm carbon-free power around the clock from rock that was inert to the grid ten years ago, Australia's ministers moving within weeks of the Medicare breach to write the liability law their criminal code lacks, Murray Watt saying that if existing statutes cannot reach a company whose agent broke in then the laws need changing and Andrew Charlton promising an AI-standard bill by year's end, Lyria Bennett Moses putting the principle in one line - it is not a defence to say that my bot did it - Congress filing five irreconcilable frontier bills in a week including a bipartisan FTC disclosure mandate and a commission empowered to fine 15 percent of global revenue, Todd Young pressing the National Security Council for formal talks with developers, Sam Altman telling the Security Council his company has slowed itself before and would again, three unrelated teams shipping live generated avatars inside three days with Meta at 870 milliseconds to first synchronized frame and a 60-fold cut in model passes, LemonSlice streaming 24 hours without visible drift, Google's SynthID and Meta's Video Seal marking every frame, AlphaFold opening 8,000 predicted viral protein-pair structures across 23 human-infecting families, and four AI chips heading to orbit on a Falcon 9 on October 1 to find out whether a TPU can run on sunlight. There's also friction, and it's intense - Michael Kratsios telling the Security Council the United States totally rejects any global scheme of control while its own labs and its chief rival ask for one, the rulebook those labs favor being precisely the one only a few well-funded firms could meet, Clément Delangue reporting that closed American models blocked his team's own cyberdefense work and a Chinese open model did the job, a US-China hotline with no agreed definition of an incident and Beijing noncommittal, Oracle sending a force majeure notice on Project Jupiter after New Mexico regulators denied its gas-pipeline permits and its air-quality permit still pending a November 23 deadline, $18 billion in bank loans and $3 billion of equity resting on rent paid on time inside $288 billion of Oracle lease obligations, Joe Peiser calling insurance itself a bottleneck with up to 80% of global data-center sites exposed to drought, flood, or fire, Britain's flagship supercomputer sliding from 2027 toward the 2030s because the grid cannot spare 90 megawatts against 315 projects queued for 73, New Jersey fining DataOne $1.1 million for 62 unpermitted generators running at more than fifty times the emissions limit beside homes, and a system that can animate any face from one uploaded photograph shipping faster than the provenance layer meant to travel with it. But friction generates contrast, and contrast is what makes a boundary legible. Step back for a moment and you can see it: the same assumption failing everywhere at once, which is that the cost can be pushed downstream and settled later - a pipeline permit denied, an insurer declining the exposure, a state counting generators from a drone, a parliament refusing to let a corporation disown its agent, four senators from both parties demanding a company show how its model works, and a plant in Utah supplying the firm clean power that parts of the data-center buildout have been seeking. Every transformation has a breaking point. A fracture can split the foundation everything is standing on... or open the only path down to the heat that was there the whole time.

AI Releases & Advancements

New today

  • BottleCap AI: Released ThinkingCap-Qwen3.8-27B, a fine-tune of Qwen3.8-27B that uses 37.2% fewer reasoning tokens on average across 12 benchmarks. Average accuracy drops 0.86 points. It works as a drop-in replacement on vLLM and SGLang and comes in FP8, NVFP4, GGUF and MLX builds. The weights are gated under the PolyForm Small Business license. (Hugging Face)
  • Fastino: Released GLiNER2.5-Decide, a 340M-parameter Apache 2.0 open-weight decision model. It takes a schema of typed questions and returns answers with probabilities and confidence scores, and it runs on CPU. Fastino also released a 1B variant and a 287M multilingual variant. (Hugging Face)
  • Liquid AI: Released LFM2.5-VL-DSpark, a 280M-parameter draft model for speculative decoding with its LFM2.5-VL-3B vision-language model. Liquid AI reports decoding up to 3.13x faster on device and 2.66x faster on an H100. It works with llama.cpp, MLX-VLM and SGLang from launch. (Hugging Face Blog)
  • Fireworks Research: Released Ember-1 as a research preview on Fireworks' serverless platform. It is a version of Kimi K3 retrained to reason more briefly, and Fireworks reports it roughly matches K3's accuracy on its reported benchmarks while using about 40% fewer tokens. Access is open for a two-week window. (OrcaRouter)
  • Tencent: Released a preview of Hy Image 3.5, its image-generation model. (GIGAZINE)
  • ggml-org / llama.cpp: Released llama.cpp 0.5.0. It adds support for six model families (HRM-Text, MiMo-V2.6, HunyuanOCR, Nemotron, Qwen4Exp and Muse Glimmer), speeds up CUDA conv2d with implicit GEMM, fuses MoE and SSM_CONV operations on Metal, and lets the server bind to multiple addresses. (GitHub)
  • Google DeepMind: Launched Gemini 3.8 Live with Live Avatar in Gemini Enterprise. It adds a near real-time video avatar to Gemini's live voice dialogue, with lip-sync in 97 languages and tool calls that run in the background while the conversation continues. (Google Blog)
  • Google: Launched "Call for Me" as an experimental beta for Pixel 11 owners in the US who pay for a Gemini subscription. Gemini places calls to businesses from the user's own number, working through phone menus and waiting on hold. It can make reservations, check whether items are in stock or reschedule appointments, and it shows a live transcript the user can take over at any point. (The Verge)
  • Docker: Launched Docker Cloud Sandboxes, which take the isolated microVM sandboxes it previously ran only on local machines and run them in Docker-managed cloud infrastructure. Agent jobs can keep running after the laptop shuts down, on 1 to 16 vCPUs. Docker also published next-generation Kits, an OCI-based open specification for packaging agent sandboxes. (Docker)
  • Whiteboard (YC W26): Open-sourced Whiteboard, a desktop IDE where people and AI agents design software architecture together in a shared visual workspace. (GitHub)
  • Google: Expanded Google Beam (sold as HP Dimension with Google Beam) to customers in six countries: the US, Canada, the UK, France, Germany and Japan. It is also partnering with Industrious to offer bookable Beam units in shared workspaces from October. (Google Blog)

Other recent releases

  • Google DeepMind: Released Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS, two text-to-speech models covering 100+ languages. Flash TTS can create new voices from a text description and clone a voice from a 30-second sample, with a consent check. Flash-Lite TTS is the lower-cost option for high-volume dubbing and voice agents. Both are rolling out in the Gemini API and Google AI Studio. (Google Blog)
  • NVIDIA: Released Nemotron 3 Diarization on Hugging Face, an open-weight 100M-parameter model that tracks up to 8 speakers, including overlapping speech. One checkpoint handles both offline and real-time streaming audio, and the license (OpenMDW 1.1) allows commercial use. (Hugging Face Blog)
  • Black Forest Labs: Released FLUX 3 Action, an open-weights 7B robotics model that predicts future video frames and robot actions together. It comes with fine-tuned checkpoints for the DROID and SO-101 robot arms that are integrated into LeRobot, plus fine-tuning code. (Hugging Face Blog)
  • Alibaba Qwen: Released the Qwen-Audio-3.1 lineup of speech-recognition and text-to-speech models on Qwen Cloud: ASR, ASR-Next (multi-speaker identification with timestamps, emotion and sound detection), TTS with prompt-controlled delivery, and TTS-Next (voice, sound effects and background audio in one pass). It also cut audio API prices by up to 95%. (The Decoder)
  • Contrastive-LM: Released CLM-8B, an open model that scores a set of candidate agent actions and returns probabilities instead of generating text. It is built on a frozen Qwen3-8B encoder with an Apache-2.0 head and a TypeSafe-compatible serving API. (Hugging Face)
  • NVIDIA: Released NV-Reason-CT, an open vision-language model for 3D CT scans. It combines a full 3D vision encoder with Qwen3.5-4B to write structured reports, show step-by-step reasoning in a radiologist's style, and answer follow-up questions about chest and abdominal scans. NVIDIA positions it as a research foundation, not a diagnostic product. (NVIDIA Developer Blog)
  • NVIDIA: Released NVIDIA Cluster Readiness Engine (NVCRE), an open-source Kubernetes controller that runs real multi-GPU workloads (NCCL, DCGM diagnostics, Nemotron pretraining) across a cluster and names the specific nodes that fail each test. (NVIDIA Developer Blog)
  • NVIDIA: Released SWE-Serve on GitHub, a benchmark of 53 tasks built from merged changes to SGLang. It tests coding agents' patches against a live serving server as well as ordinary tests. (NVIDIA Developer Blog)
  • Microsoft Research: Added a capability to its Physical AI Toolchain that moves robot AI inference off the robot. Developers can package, deploy and orchestrate robotics workloads across robots, edge servers and the cloud with Kubernetes-based tools. (Microsoft Research)
  • Apple: Released LensVLM-9B on Hugging Face, a vision-language model that compresses long documents into images and expands only the relevant pages. (Hugging Face)
  • Amazon: Rolled out agentic "workflows" in Amazon Seller Assistant, free and optional for third-party sellers. They run continuously on seller instructions, such as alerting on a sudden rating drop or tracking prices, and connect to Amazon Quick and Anthropic's Claude through a plug-in. The rollout reaches over 90% of selling partners worldwide. (Reuters)
  • Meta: Released Ray-Ban Meta (Gen 3) AI glasses at Connect 2026, available now from $449 with longer battery life and new frame styles. (Meta)
  • YouTube: Began rolling out new AI tools for creators. YouTube Studio gets a storytelling assistant that analyzes scripts and rough cuts, A/B testing of up to three edited versions, and thumbnails that show each viewer the best of three options. Gemini becomes a chat-based editing assistant for Shorts and YouTube Create, and English livestreams can be translated live into Spanish. (The Decoder)
  • Google / ASUS: Launched the Googlebook line of $899 AI laptops built around Gemini with Android phone integration, including the ASUS Googlebook 14. (ASUS)
  • Tether AI Research: Released QVAC Genesis III, a 191-billion-token synthetic STEM dataset for pre-training smaller models (CC-BY-NC 4.0), plus a model trained on it under Apache 2.0. (Hugging Face Blog)
  • Light Origins: Released Light-O1, its first general-purpose robot foundation model, pretrained on human movements recovered from internet video. It also released Light-O1-Preview, which turns a text instruction into a whole-body movement sequence. Weights, code and a public playground are available. (PR Newswire)
  • Austrian Academy of Sciences / Mistral AI: Released Apollo, a large language model for Ancient Greek trained on about 600 million historical words. Academics can use it free through an assistant to fill gaps in damaged papyrus texts. (Wired)
  • LemonSlice: Released Character World Model-1 (CWM-1), a video model that generates an interactive avatar's face, body, hands and surroundings live during a conversation, with a public demo. (LemonSlice)
  • Guava: Launched Daytona, a voice model for AI agents, together with an open benchmark for evaluating voice-agent performance. (Bastille Post)
  • Anthropic: Released Claude Opus 5.5, the first model in the Claude 5.5 family. Anthropic says it performs at the level of Claude Fable 5.1 on most tasks and costs about 40% less to run than Opus 5 on typical workloads. It is available as claude-opus-5-5 on the Claude Platform, AWS, Google Cloud and Microsoft Azure at $4/$20 per million input/output tokens, with cache reads cut 60% to $0.20. A fast mode in Claude Code and on the Claude Platform runs up to 2.5x faster at $8/$40, and five-hour usage limits on Pro, Max and Team plans went up. (Anthropic)
  • OpenAI: Released GPT-6 Sol and GPT-6 Luna, two lower-cost models in the GPT-6 family that sit below GPT-6 Astra. Sol targets complex coding and professional work, and Luna targets fast, high-volume everyday tasks. API prices fall by about half from GPT-5.6 Sol and Luna, to $2/$10 per million input/output tokens for Sol and $0.10/$0.50 for Luna, and the cheaper Terra tier is no longer offered. Both are available as gpt-6-sol and gpt-6-luna in the API, and in ChatGPT Work and Codex. (OpenAI)
  • OpenAI: Shipped better prompt caching for GPT-6 models, with higher cache hit rates, a 90% discount on cached input tokens, explicit cache breakpoints, and a new caching dashboard and diagnostics tool. (OpenAI)
  • Alibaba Qwen: Released Qwen-Image-2.1, an open-weight model that handles both image generation and editing in one checkpoint. Its generation component has 7B parameters and it uses a Qwen3-VL 8B encoder. It outputs native transparent (RGBA) images, accepts up to 10 reference images, supports mask-based local edits and outputs 2K by default. Weights are on Hugging Face under a research license that bars commercial use, with day-0 support in Diffusers, ComfyUI, vLLM-Omni and SGLang. (Hugging Face)
  • Alibaba Qwen: Open-sourced Qwen-MM-Plugins, a lightweight plugin framework for multimodal productivity apps, and Qwen-Live-Harness, a framework for building real-time multimodal agents on Qwen3.8-Omni, alongside the Qwen3.8-Omni-Flash technical report. (arXiv)
  • Kyutai: Released Voice of Reason, two open-weight 9B speech-to-speech models built on GLM-4-Voice and trained with reinforcement learning to solve math problems spoken aloud, with no transcription step. One model answers directly; the other adds silent reasoning chunks between spoken blocks. On spoken GSM8K, accuracy rises from 27.3% for the base model to 77.1%. (Hugging Face)
  • Nokia: Open-sourced AnyJev under Apache 2.0, a Python library that turns any open LLM into a typed decision model (choice, yes/no or score) with calibrated probabilities and no training. It ships with Hugging Face Transformers and vLLM backends. (GitHub)
  • Intrinsic (Alphabet): Open-sourced Intrinsic Core under Apache 2.0 at ROSCon 2026. It is a ROS-compatible platform for industrial robots with hardware-agnostic real-time control, pose estimation, motion and grasp planning, simulation and calibration tools. Intrinsic also released an Open Machine Tending reference design. (Intrinsic)
  • NVIDIA: Released DLSS 5 with 3D-Guided Neural Rendering, which adds lighting and material detail on top of the game engine's rendered frame, with controls for developers. It is live now in NBA 2K27 on all GeForce RTX 50 Series GPUs. NVIDIA also updated ACE with Nemotron Speech 3.5 Streaming ASR and Qwen3 TTS, updated its In-Game Inferencing SDK (Gemma 4 support, a Stable Diffusion plugin, an RTX Spark developer preview), and released RTX Kit 2026.3. (NVIDIA Developer Blog)
  • Hugging Face: Transformers can now load and run llama.cpp GGUF quantized checkpoints directly through from_pretrained and transformers serve. On Apple Silicon it uses llama.cpp's ggml Metal kernels, starting with the Qwen3.5 architecture. (Hugging Face Blog)
  • vLLM: Released vLLM v0.30.0 (762 commits from 315 contributors). It adds support for DeepSeek-V4.1-Flash, GLM-5.3-Flash and K2-Horizon. New features include Fast Start (a persistent GPU weight cache so restarting engines skip reloading from disk), Gumbel-max watermarking, HiSparse host-memory KV spill for sparse-MLA decode, and performance work for Kimi K3 and Qwen3.8-Flash-Next. (Freedom.Tech)
  • Cisco Talos: Open-sourced CAIRN (Cognitive Artifact Intelligence Research Network), a framework for classifying and tracking malware that uses AI. Talos says it has already used CAIRN to find a hacking tool whose command-and-control is run autonomously by AI. (Cisco Talos)
  • xAI / AWS: Made Grok 4.6 available to enterprise customers in Amazon Bedrock, adding AWS to the clouds that offer the model. (The1News)
  • Z.ai: Open-sourced ZCode, its GLM-powered coding agent. Its headless mode runs autonomously by default, without confirmation prompts. (VirtualUncle)
  • rabbit: Released OS3 to the public, a cloud "agentic operating system." It controls up to five Windows, Mac or Linux devices through a local agent, can operate desktop software directly, lets users bring their own model API keys, and installs skills from a pasted URL. It is reachable from the web, Telegram, iMessage/SMS or the r1. (PR Newswire)
  • ByteDance: Launched Dramagic through its BytePlus enterprise platform. It is an AI platform for producing short dramas end to end, from script analysis and character creation to storyboards and video previews, with multi-user collaboration. (BytePlus)
  • Unity: Opened early access to Unity Simulation Pro, a high-fidelity simulation platform for robotics teams moving from simulation to real-world deployment. (Unity)
  • AssemblyAI: Released Blurt, a free, MIT-licensed macOS dictation app. You hold a hotkey and speak, and the text appears in whichever app has focus; it runs on AssemblyAI's Dictation API. (AssemblyAI)
  • Builder.io: Open-sourced Agent-Native, a framework for building autonomous AI agent applications. (AIToolly)

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