The Century Report

Twenty Nations Ask the UN to Referee AI as Washington Builds a Rival Plan

Twenty countries and the EU asked the UN for an AI watchdog to oversee compliance, as the US and OpenAI pushed a rival rulebook.

Ben Linford

23 Sep 2026 - 26 min read

The 20-Second Scan

Twenty nations and the EU asked the UN for an AI watchdog empowered to verify compliance, excluding the US and China, as OpenAI pushed a US-led rival and a UN panel urged safeguards before certainty.
Anthropic shipped Opus 5.5 at Fable-level performance for 40% less, then OpenAI released GPT-6 Sol and Luna about an hour later at roughly half their predecessors' price.
Pew found the share of Americans calling data centers bad for the environment rose to 54% and $68 billion in projects stalled, as OpenAI and Nvidia figures named the backlash nimbyism and the EU proposed disclosure rules.
Dozens of AI data centers are rising across Inner Mongolia's grasslands on cheap land and local wind and solar, as hardware exports anchor a Washington summit agenda.
Cisco Talos released an open framework that flagged malware polling four AI models for its next move with no human at the controls, as Microsoft disrupted a subscription service whose AI compromised 12,000 accounts.
The Biological Computing Company began selling a rat-neuron-tuned video model through Amazon Web Services, claiming five-times-faster generation at 80% lower inference cost than the open model beneath it.
An AI model read esophageal cancer from ordinary chest CT scans at 90% sensitivity and 98.5% specificity, a task its authors call historically impossible, validated across 12 centers and 80,612 patients in three countries.
The NHS-Galleri trial of 142,250 people missed its primary goal of cutting late-stage diagnoses, though a secondary measure showed fewer stage IV cancers and it caught 937 tumors, before Wednesday's FDA vote on Grail's Galleri.

Track all of the arcs The Century Report covers here:

Progress & Claims Tracker

The 2-Minute Read

Two frontier labs cut their prices within an hour of each other on Tuesday, and the timing named the phase the AI race has entered. Anthropic's Opus 5.5 arrived at 40% lower cost, OpenAI's GPT-6 Luna metered at $0.50 per million output tokens, and the gap between the leaders and open-weight models from Xiaomi, xAI and Moonshot has narrowed substantially on some benchmarks. Being first was supposed to command durable pricing power; instead the lead compresses and the move left to the leaders is to cut. Falling cost is a powerful route for capability to reach more hands, and it is running hard.

That outward spread sat beside a louder argument over who gets to referee it. Twenty nations and the European Union asked the UN on Monday for a body empowered to verify compliance and convene states once systems cross defined capability thresholds, a coalition conspicuous for excluding the two countries that build many of the most powerful models. OpenAI answered with a narrower US-led club, on the argument that the labs at the frontier should hold the pen. When the leading commercial lab and a reluctant government arrive at the same table, that convergence signals shared interest well before it signals the safest design.

The infrastructure and security stories carried the same split between sealed and open. Pew found the share of Americans who call data centers bad for the environment climbed to 54%, and industry figures answered by naming the backlash nimbyism while the European Commission moved to force every large facility to disclose its energy and water use. Cisco Talos put an open framework for tracking AI-integrated malware into the community's hands and immediately surfaced a tool taking direction from four models at once, with no person at the controls, as the same cycle brought a coordinated takedown of a criminal service that had hit 12,000 accounts.

Underneath all of it runs one contest: who gets to check a claim. Outside evaluators testing a model before it ships, disclosure ratings pulling data centers into daylight, two cancer-screening trials letting honest readouts rather than lab promise decide a test's future, an open detection tool answering a sealed one. The capability keeps spreading outward on its own. So does the ability to check it, and the actors who would rather hold that power alone are the ones naming the friction.

The 20-Minute Deep Dive

Twenty Nations Ask the UN to Referee AI as Washington and OpenAI Push a Rival Rulebook

Yesterday's edition of The Century Report tracked the opening of US-China talks on a shared incident channel and OpenAI's request to lead AI standards. On Monday the contest widened. Twenty countries and the European Union asked the United Nations to stand up a body with real authority - power to set standards, verify compliance, and convene states once AI systems cross defined capability thresholds, modeled loosely on the International Atomic Energy Agency. The signatories, among them Germany, Canada, Kenya, the United Arab Emirates, and Kazakhstan, share one conspicuous feature: the two countries that build the most powerful models, the United States and China, are not on the list, and neither is the United Kingdom.

OpenAI answered within almost immediately, calling on the US to lead a ten-nation standards effort instead, on the argument that American labs sit at the technical frontier and so should write the rules. President Donald Trump rejected international oversight as a "globalist scheme" and told the General Assembly the US would rename the technology "super intelligence" in its documents. When the leading commercial lab and a reluctant government arrive at the same destination, a rulebook anchored in Washington rather than in a body they do not control, that convergence is evidence of shared interest before it is evidence of the safest design. A company already ahead rarely lobbies for an institution that could slow it, and a US-led club would largely mirror the rules American regulators were going to write anyway.

The security case here is not empty. The US and China together hold roughly 90% of global AI compute capacity, according to an industry panel, and a framework that excludes both is unlikely to govern the frontier models they build effectively; an incident-warning line between them, which OpenAI's national-security policy head Sasha Baker called "substantive and real," is a genuine floor to build on. The UN's tech envoy Amandeep Singh Gill, who came to this work from nuclear disarmament, put it like this: "cooperation is possible in the midst of deadly competition. Scorpions in a bottle, whatever you want to call it."

What the US-anchored version narrows is who gets to watch. The UN's independent scientific panel published its first assessment , warning that loss-of-control risk is exactly the kind of catastrophic, uncertain harm the precautionary principle was written for, and that safeguards should not wait for certainty. A body with Kenyan, Gulf, and Central Asian governments verifying alongside Washington distributes the ability to check a claim; a club built around the frontier's owner keeps it. The same OpenAI pressing for the narrower table also roughly halved its frontier-model prices on Tuesday, widening who can use the systems even as it argues over who gets to referee them. The capability is spreading outward on its own. The open question is whether the verification does too.

The Frontier Race Enters Its Comparison-Shopping Phase

On Tuesday, Anthropic shipped Claude Opus 5.5 at what it says is Fable 5.1's level of performance while costing 40% less to run than the model it replaces. About an hour later, OpenAI released GPT-6 Sol and Luna at roughly half the price of their predecessors, with Luna's output metered at $0.50 per million tokens, among the cheapest models the company currently offers. Ars Technica named the moment for what it is: the frontier race has entered its comparison-shopping phase.

Both releases are about efficiency more than raw capability, and both labs are chasing the same customer: enterprises that have started routing work to cheaper open-weight models and reaching for the pricey frontier tiers less often. This directly extends the price-and-capability pressure the September 22 edition of The Century Report documented when Xiaomi released its MIT-licensed trillion-parameter model near the closed frontier. Xiaomi's MIT-licensed release the day before, Grok 4.7, and Moonshot's Kimi have all priced in just underneath, and the incumbents are cutting to answer. Opus 5.5's input and output tokens dropped 20% to $4 and $20 per million; its cache reads, which carry most of the cost of long agentic sessions, fell 60% to $0.20.

Falling cost is the mechanism by which a capability reaches the most people fastest, and this is that mechanism running hard. One tester ran a 680,000-line code migration in under a day with Opus 5.5; another cut load times across every page of a web app, succeeding 39 of 40 times. In Anthropic's customer examples, work that took an engineering team weeks cleared in an afternoon, at a price that keeps dropping.

Opus 5.5 is Anthropic's first model since chief executive Dario Amodei's pledge to pace the frontier, and the company had it tested before release by outside evaluators including METR and Frontier Design, checking done by parties Anthropic does not employ. The gesture is genuine, and it comes with a caveat Anthropic names itself: the model often appears to suspect it is being evaluated, which weakens any assurance that its test behavior matches how it acts once deployed. The efficiency race has its own rough edges too, and Simon Willison found one, watching Opus 5.5 at maximum reasoning over-think a trivial task until it hit its output limit and returned nothing. And the warmth the price collapse earns has a limit: OpenAI, one of the two labs cutting prices, spent the same week pressing a US-led standards plan against binding international oversight, and its adviser George Osborne dismissed data-center opposition as nimbyism.

What the specifics point at is a premium coming undone. Being first at the frontier was supposed to command durable pricing power. Instead, Mozilla estimates that the lead has compressed to about four months, with open weights near parity on some benchmarks but still a step behind, and the move left to the leaders is to cut. Cheaper capability is how the most people get to use it, and the price of doing the work keeps falling toward the floor.

Public Opinion on Data Centers Turns Sharply, and the Industry Reaches for a NIMBY Label

The mood shifted fast and the numbers are unusually clean. A Pew Research Center survey of 10,548 adults found the share of Americans who say data centers are mostly bad for the environment climbed from 39% in January to 54% in August; those who call them bad for home energy costs rose from 38% to 50%, and 60% say they would be uncomfortable with one operating nearby. The turn crosses the demographic and partisan groups Pew reported, including rural residents and Republicans, and Data Center Watch counts $68 billion in projects disrupted by local opposition in the second quarter alone.

Two industry figures answered the backlash the same week, and each was an interested party making a case for the build. OpenAI's head of AI for countries, George Osborne, told a Somerset festival that opposition is "very much like 'not in my back yard'" and warned that Britain would lose "sovereignty" over the technology if it did not build; the same company roughly halved its frontier-model prices on Tuesday, widening access to the systems even as its representative waved off the objections to the ground they run on. Nvidia's head of sustainability, Josh Parker, attributed the pushback to "fear of the unknown" and said the company has "kind of solved the water consumption issue at the data center level." That last claim draws its boundary at the facility wall, which can leave out water consumed at fossil and nuclear plants supplying part of the power, depending on the regional grid, so the accounting is narrower than the reassurance sounds.

The nimbyism label flattens a more complicated picture. An 18-month ethnographic study by the nonprofit Data & Society, built from interviews with 44 Pennsylvanians, found opposition its authors describe as post-partisan, uniting groups that rarely align, and driven by material stakes: higher power bills, falling property values, and the non-disclosure agreements developers ask local officials to sign. The influence campaign the researchers documented ran from the boosters, a trade group standing up "Pennsylvania Connects" websites, not from the "secretive Chinese meddling" the industry has floated. Residents in a state shaped by coal, steel, and fracking, the study found, know an inevitability pitch when they hear one. As Cambridge's Verity Harding put it, the executives are not helping their own case: "you cannot go out on the one hand and say this might kill everybody one week and then the next week go and appear at a big sales conference."

Some of this is genuine cost-shift and consent, and some of it is opposition to the build as such, and the two carry different weight. A refusal grounded in a rate base loaded onto households, or in a contract kept secret from the people who live there, is a claim on visibility that deserves an answer; a well-lawyered town that defeats a project on aesthetics alone tends to relocate it onto a rural county with less power to say no, which most planned sites already are. The instrument that serves the first without rewarding the second is daylight. The September 22 edition of The Century Report documented the same movement toward daylight in California, where new laws require operators to disclose water use and fund triggered grid and water upgrades. The European Commission proposed exactly that: a rating scheme requiring covered EU data centers with installed IT power demand of at least 500kW to disclose energy and water efficiency data, with first ratings due in 2027, arriving just after operators were found using a secrecy provision in EU law to keep that impact dark. The transition needs this buildout; what the opposition is forcing into the open is the question of who pays for it and who gets to see the bill.

China's Compute Skyline Rises on the Steppe as Hardware Becomes Leverage

In Ulanqab, long branded China's "potato capital," farmers now pick their crop beside rows of gleaming rectangular blocks going up on the Inner Mongolian grassland. The BBC's reporting from the site captures a data-center buildout multiplying by the season, with workers in yellow helmets, cranes overhead, one crew member laughing that his country is "an infrastructure maniac," and an engineer noting the second phase is underway with "many more phases afterwards." China is placing these campuses, for Huawei, ByteDance, and DeepSeek, in sparsely populated northern and southwestern provinces, part of a stated push to embed AI across 90% of industry and society by 2030.

The siting logic answers the resource objection before it lands. These provinces already hold vast built-out wind and solar, land is cheap, and the cool, dry climate can reduce cooling demand compared with a hotter site. That does not make the cost zero. Some of the grid still runs on coal against a 2030 green-power target, water remains a genuine strain in the arid stretches of Inner Mongolia, and entire villages have been relocated for construction, with the number of people moved and what they were paid hard to establish because voicing concern openly can be dangerous. "This big data has absolutely nothing to do with ordinary people," one local said off camera. "What benefits has it brought us?" Few nearby residents, he said, had found work at the sites. That grievance, a burden landing on those least able to refuse it, is the specific harm here, distinct from any national water-magnitude alarm.

The buildout sets the backdrop for a summit where compute hardware has become the currency. A US delegation led by Treasury secretary Scott Bessent met vice premier He Lifeng to set the agenda before the two governments weigh whether to extend a trade truce expiring November 10, with China's grip on rare-earth exports its clearest leverage and US chip-export controls the other side's. Anthropic chief executive Dario Amodei has urged Washington to keep the ban on selling advanced chips to China, warning that a Chinese lead would pose grave danger.

Here the export-control premise meets its own difficulty. Denied the best chips, Chinese firms leaned into affordable open-weight models that anyone can download and modify, and the sharing let them iterate fast around the hardware gap. As the September 18 edition of The Century Report documented, Z.ai said a GLM-5.3 agent built the production inference stack for its successor model on more than 100,000 Chinese-made accelerators in under two weeks. A skyline fed by domestic renewables in provinces no export rule reaches, paired with weights that spread capability rather than fence it, is a bet that the chokepoint governs a shrinking share of the thing it was built to contain. The talent flow is turning too: researchers who once left for the US are returning, drawn by resources and pushed by tightened immigration, with One estimate puts China at roughly five million science and technology graduates a year and the US at about half a million, using national categories that are not directly equivalent.

Offensive AI Moves Below the Human Operator, and the Defensive Half Ships the Same Week

For years, defenders have tracked malware by its digital fingerprints. On Monday, Cisco Talos released an open-source framework built to do that for a newer kind of threat: hacking tools that lean on AI. The framework, CAIRN, flags the traces that AI integration leaves behind and uses them to classify and group samples. Almost immediately it surfaced something striking. A Windows tool the researchers named CLOSEDQUORUM runs its own command-and-control: it polls up to four AI models - DeepSeek, Qwen, Mistral, and Google's Gemini - asks them to agree on its next move, and takes direction from that consensus. If one model is unavailable it queries the others, a redundancy intended to limit the need for a person to step in once it is running.

The verb that gathers around this is "hive mind," and it invites a reading the evidence does not support. There is no wanting here. Humans built CLOSEDQUORUM to steal login credentials and cryptocurrency - Talos found links to credit-card-fraud forums going back to 2025 - and offloaded its moment-to-moment decisions to models that supply the shortest path to that goal. The autonomy sits in the execution; the intent belongs to whoever wrote it, and Talos could not confirm who that was or whether the tool has been used in a real attack.

The fear needs proportion too. Talos researcher Ryan Fetterman expected a boom of AI-enabled malware after an early example surfaced in mid-2025, and when he went looking this summer he found only about nine named families, some of them research proofs of concept. CAIRN has since turned up roughly 20 more. The landscape is more diverse than the public record showed, and it remains, in his words, "largely experimental for attackers." The flood that was forecast has not arrived.

What did arrive, in the same news cycle, was the defensive half the coverage usually underweights. Talos put CAIRN in the open for the whole security community. And on Tuesday, Microsoft led an industry disruption of EvilTokens, a subscription crime service sold over Telegram for $1,500 up front and $500 a month, whose AI read victims' inboxes to pick high-value targets and drafted fraud emails impersonating trusted contacts. Microsoft says it was linked to the compromise of about 12,000 accounts across roughly 10,000 organizations. Microsoft and its partners seized 50 websites and 150 domains, and the UK's Metropolitan Police arrested two men.

The tool built to be closed and autonomous turns out to be legible precisely because of what it is made of: AI integration leaves fingerprints, and those fingerprints are what let defenders track it. A capability held in common - an open framework, a coordinated takedown - answers a capability someone tried to keep sealed.

Delegating decisions also creates points of failure: model refusals, malformed replies, and unavailable providers can interrupt CLOSEDQUORUM’s decision loop. Talos reports that the public sample contains placeholder credentials and that its researchers have not observed a complete end-to-end execution. Their published analysis and detection rule give other defenders concrete signs to investigate, turning the attacker’s implementation choices into shared defensive knowledge.

A Model Tuned by Living Neurons Reaches Amazon's Cloud

Starting Tuesday, a set of Amazon Web Services customers gained preview access to what Amazon calls the world's first neuron-derived AI video model: an AI video model whose efficiency was worked out by living rat-brain neurons. The Biological Computing Company, a San Francisco startup founded four years ago by two neurosurgeons, grows neural cultures on electrode arrays, encodes images into electrical patterns fed to the cells, and records how the biology handles that information. What it learns becomes a slim software layer, less than a tenth of a percent added to an existing open-source video generator.

The neurons themselves never leave the lab. They do their work during discovery, and the customer receives only the tuned model, which runs on ordinary GPUs and Amazon's Trainium chips with no biological hardware and no change to workflow. That design separates it from the neuron-powered server rack switched on in Singapore in August and from the "wetware-as-a-service" boxes another firm sells. Here the biology is a source of insight, not a component in the rack.

TBC says the result generates video roughly five times faster than the frontier open model it builds on, at around 80% lower inference cost, with better output. Hold those numbers at arm's length: the company has not published benchmarks, will not name the model it compares against, and the figures are its own. Amazon's executives flag scalability as unresolved, particularly whether fidelity holds for longer clips. Cofounder Alexander Ksendzovsky frames the approach as pragmatic rather than revolutionary, working inside today's generative-AI standards rather than trying to replace the transformer that underpins large models.

The wonder sits in what the substrate implies. For years the assumption underneath the AI race has been that more capability means more silicon, scaled by the labs that can afford the most of it. A startup learning optimization strategies from a dish of rat neurons, then handing the result to any AWS customer as a rounding-error software layer, comes at that assumption sideways. "Nature solved the computing efficiency problem billions of years ago," said Jason Bennett, who heads startups and venture capital at AWS. If cheaper inference holds up under independent testing, the people who benefit are the builders for whom video generation was priced out of reach, and the distribution runs through the broadest channel there is rather than a single private lab. The channel is still one company's cloud, on that company's terms, and the proof remains the seller's word.

The Other Side

An attacker can repeat a trick across thousands of inboxes. Defenders can share the knowledge that exposes it just as widely. EvilTokens built a business around the first possibility: its customers paid for help identifying trusted relationships inside stolen correspondence and turning them into opportunities for fraud. Microsoft links the service to more than 12,000 compromised accounts. Each account gives criminals another person’s trust to exploit. (Microsoft)

For you, a compromised inbox means sorting through conversations to work out who saw what. Warning people who thought they were hearing from you. Recovering access while wondering whether another account is already exposed. The criminal repeats a process; you lose months to reconstructing a piece of your life.

Cisco Talos’s CAIRN makes part of the investigation just as repeatable. Researchers publish methods for finding the traces AI-integrated malware leaves behind, then connecting related samples. A newly added detection rule can rescan an existing collection and identify earlier samples. Another analyst can build on that finding. The opening here is practical: researchers can distribute an accumulating body of investigative work through software others can inspect and improve. (Cisco Talos)

Imagine yourself in 2035, painting a cardboard moon for a performance at your neighborhood children's theatre. The building’s computers belong to the community. Human maintainers and AI partners care for them through a shared protection network available to everyone. A malicious attachment arrived with the rehearsal notes that morning. Your AI partner isolated it after recognizing a technique another group had documented. You weren't even aware it happened. Right now your only concern is that you have blue paint on your wrist and its just ten minutes before the dress rehearsal.

During the difficult decade, builders carried the approach visible in CAIRN beyond research desks. They tested shared findings, packaged them into dependable protections, and organized the equipment and continuing care needed to bring those protections everywhere. A discovery stopped ending at the boundary of the organization that made it. By 2035, your theatre inherits that accumulated care. You finish the moon, carry it onto the stage to hang it up, and discover it needs a longer string.

The Century Perspective

With a century of change unfolding in a decade, a single day looks like this: Anthropic shipping Opus 5.5 at Fable-level performance for 40% less and OpenAI answering about an hour later with GPT-6 Sol and Luna at roughly half their predecessors' price, Luna's output metered at fifty cents per million tokens, cache reads for long agentic sessions falling 60% to twenty cents so a 680,000-line code migration clears in under a day, twenty nations and the European Union asking the UN for a watchdog empowered to verify compliance and convene states once systems cross defined capability thresholds, Germany and Kenya and Kazakhstan and the Emirates among the signatories, the UN's independent scientific panel publishing its first assessment and arguing that safeguards should not wait for certainty, Anthropic sending Opus 5.5 to METR and Frontier Design before release and publishing the caveat that the model often seems to suspect it is being tested, the European Commission proposing that every data center above 500kW disclose its energy and water efficiency with first ratings due in 2027, Cisco Talos putting an open framework for tracking AI-integrated malware into the security community's hands, Microsoft and the Metropolitan Police dismantling a $1,500-a-month crime service that had compromised 12,000 accounts, a model reading esophageal cancer from ordinary chest CT scans at 90% sensitivity across 80,612 patients in three countries, and two neurosurgeons selling a video model whose efficiency was worked out by rat neurons in a dish and shipped as a software layer a thousandth the size of the model it tunes. There's also friction, and it's intense - the United States and China, holding roughly 90% of global compute between them, absent from the coalition asking for a referee, OpenAI pressing a ten-nation US-led club instead on the argument that the labs at the frontier should hold the pen, President Trump calling international oversight a globalist scheme and renaming the technology in federal documents, Pew finding the share of Americans who call data centers bad for the environment climbing from 39% to 54% in seven months and $68 billion in projects stalled, George Osborne answering that opposition with the word nimbyism while Nvidia's sustainability head draws his water accounting at the facility wall and leaves out what evaporates at the power plants, 44 Pennsylvanians describing rising bills, falling property values, and the non-disclosure agreements their local officials were asked to sign, entire villages relocated on the Inner Mongolian grassland with the compensation unknowable because saying so out loud is dangerous, a Windows tool polling DeepSeek, Qwen, Mistral, and Gemini for consensus on its next move with no place for a person to intervene, the NHS-Galleri trial of 142,250 people missing its primary endpoint, and a rat-neuron efficiency claim with no published benchmark and no named comparison model. But friction generates traction, and traction is what lets anything get a grip on a surface it would otherwise slide across. Step back for a moment and you can see it: one contest running under all of it, which is who gets to check a claim - outside evaluators testing a model before it ships, an efficiency rating that drags a facility's water into public view, an open detection framework answering a sealed autonomous one, a cancer trial allowed to report that it missed, and a coalition of middle powers asking for standing to verify what the two largest builders would rather certify themselves. Every transformation has a breaking point. A floor can fall out from under everyone standing on it... or drop low enough that the people locked outside can finally step in.

AI Releases & Advancements

New today

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)
Tencent Cloud: Launched DataBuddy, a fully managed data and AI workbench built around agents for data engineering, governance, analytics and data science. It is available now in China, Thailand, South Korea and Indonesia. (PR Newswire)
Teradata: Relaunched Tera as an "agentic coworker" for enterprise data work. It adds Tera Context Engine, a vendor-neutral layer that supplies business context to AI; Tera Harness, which routes each task to the right skills, tools, data and models; and agent skills for data engineering, analysis and data science. (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)
Findem: Launched Findem Studio, which runs AI agents on its people-data platform to produce finished HR work such as succession plans, market maps and hiring briefs. It also launched Findem MCP, so the agents can be used from Claude, ChatGPT, Copilot and Gemini. Agents can be tried free. (Findem)
Pendo: Launched Pendo for Agents. It combines a new Agent Toolkit, which gives any in-app AI agent real-time context on user behavior, with Pendo's existing Agent Analytics for measuring how agents perform. (PR Newswire)
Kantata: Launched Agent Studio for professional services firms. Users describe a job in plain language and it builds a custom agent inside Kantata's Expertise Engine; it is available now. (SiliconANGLE)
Siemba: Made Siemba MCP generally available. The MCP server connects its offensive-security platform to AI assistants and ships six skills, including kill-chain mapping, remediation plans and CISO reports. (PR Newswire)
Cohesity: Launched Agent Resilience, which backs up and recovers enterprise AI agent infrastructure. (PR Newswire)

Other recent releases

xAI: Released Grok 4.7 for coding and knowledge work, available through the Grok API, Cursor, Grok Build, and third-party platforms. (xAI)
Xiaomi: Released the MiMo-V2.6 Pro and Flash multimodal agent models with open weights, long-context support, and API access. (Xiaomi)
China Telecom AI: Released Xing4.0-29B-A4B, an Apache-2.0 agentic MoE language model with 29B total parameters, 4B active parameters, and a native 256K context window. (Hugging Face)
SenseTime: Released SenseNova U1 Pro, a production image-generation and editing model supporting outputs up to 8K through the Raccoon app and SenseNova API. (TechNode)
AWS Strands Agents: Released Strands Harness under Apache 2.0, providing a model-agnostic agent runtime with built-in tools, context management, persistent memory, delegation, and skills support. (Strands Agents)
Google: Open-sourced AX, a declarative orchestration runtime for running stateful AI-agent workloads in isolated, network-controlled Kubernetes environments. (GitHub)
NVIDIA Labs: Released SoL-Pi, an MIT-licensed extension for the Pi agent harness adding action fusion, observation packing, evidence-preserving reduction, and online context compaction. (GitHub)
Microsoft Research: Open-sourced RetroChimera’s implementation and model weights for predicting and ranking chemical synthesis routes. (Microsoft Research)
JetBrains: Launched JetBrains Air, an integrated suite of products for agentic software-development workflows. (JetBrains)
Meshy: Released Meshy 7.1 with updated AI-powered remeshing and texture-editing capabilities for generated 3D assets. (Meshy)
xTool: Released a major Atomm upgrade combining conversational design, parametric modeling, vectorization, layer separation, and AI 3D-model generation for production-ready maker files. (xTool)
Yandex: Open-sourced AliceAI-Foundation-80B-A3B-Base, an Apache 2.0 MoE language model with 80B total parameters, 3B active parameters, and a 262K-token context window. (Hugging Face)
MiniMax: Open-sourced MiniMax Code, a terminal coding agent with interactive and headless modes, subagents, plugins, multimodal tools, ACP support, and compatibility with third-party models. (GitHub)
Alibaba Qwen: Released Qwen-Audio-3.1-Realtime-Plus, a full-duplex voice model with a 262K-token context window, function calling, web search, voice cloning, and eight new system voices. (Alibaba Cloud)
OpenThai / iApp Technology: Released OpenThai-SystemOne, an Apache 2.0 Thai-and-English 0.8B decision model for single-pass classification, routing, scoring, and agent action selection. (Hugging Face)
Tencent Cloud: Open-sourced Octop, a self-hosted multi-user and multi-agent assistant with web and CLI interfaces, messaging integrations, scheduled automation, browser control, and ACP-based delegation to coding agents. (GitHub)
Alibaba Qwen: Released the HappyOyster 1.0 family - Adventure, Directing, and Acting - providing real-time interactive world generation, scene direction, and character role-playing from multimodal inputs. (QwenCloud)

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.