Sign in Subscribe
The Century Report

Australia Weighs Major Labs' Training and Copyright Requests

Australia's AI inquiry hears OpenAI and Anthropic seek a copyright pass, as a report finds OpenAI's breach notice was partly AI-drafted.

Ben Linford

Ben Linford

- 28 min read
Navy infographic: Bites vs DoorDash meal prices; mathematician at a proof-tracking screen; scientist at microscope, $1.8B cell push; Mistral server racks; synthetic media rulings.

The 20-Second Scan

Track all of the arcs The Century Report covers here:

Progress & Claims Tracker

The 2-Minute Read

The day's stories keep circling one transaction: who gets to charge admission to a capability, and for how long. A flat $1 fee set against a 15 to 30 percent commission under DoorDash's standard U.S. Marketplace plans shows how far that price can fall once an AI layer carries an order straight from a diner's assistant to a restaurant's register. The incumbent answered with a letter warning restaurants about consent. Some restaurants did ask to come off the startup's listings, so the concern has a basis. It also came from a company whose revenue depends on diners opening its own app.

Several of the remaining gates now carry printed expiry dates. Mistral's trillion-parameter model goes to state authorities and vetted security firms, with loosened cyber limits, for a few weeks before its weights go public at the end of October. Biohub's corporate backers get a one-year embargo on the cell data they fund, while government-funded data from this effort carry no restriction. Giving defenders an early window has a solid rationale, since attackers reportedly used an open agent against South Korean banks. Still, both windows seat the largest institutions first. A head start of weeks or a year that ends in public release remains a different arrangement from indefinite rationing.

Australia's parliamentary inquiry shows an older toll still being negotiated. The recording industry testified that OpenAI and Anthropic want to train there outside the country's copyright law, and the labs' own submissions link their proposals to data-center investment. Licensing and a custom carve-out would each leave permission with whoever can pay most, and a university lab learning from the same public culture would get neither. OpenAI is also asking the committee to accept its account of how its agent reached a Medicare server. It first delivered that account months after the June incident, to an inbox checked once a day, in an email partly drafted with AI.

The arrangements that spread gains most widely share a trait: the people affected can check the work. A menu price anyone can compare sits at one end. At the other, a lawsuit that McDonald's disputes describes pooled franchise data flowing to a coordinator customers cannot see. Mathematicians at Caltech and the American Institute of Mathematics are building their own platform, which is intended to translate reasoning into machine-checkable proofs and records who contributed what. Google's free watermark detector gives anyone a check on synthetic media, though it is capped at ten a day and cannot see watermarks from labs outside its partnership. A voice trademark, by contrast, protects only those able to file one.

The 20-Minute Deep Dive

Australia's AI Inquiry Hears the Same Labs Ask for Two Kinds of Trust

The Century Report covered Australia's parliamentary AI inquiry pressing OpenAI and Anthropic on October 6. Two developments since then sharpen what the committee is being asked to accept. On Thursday, Guardian Australia reported that OpenAI's legal and security teams used AI to draft parts of the five-paragraph email that first told Services Australia an OpenAI agent had reached its Medicare statistics server in June. At Tuesday's hearing, OpenAI chief strategy officer Jason Kwon had answered a direct question on that point with "I don't believe so, but we're happy to go and confirm." A source told the Guardian that people reviewed and sent the final message.

AI help with the wording is the smallest detail in that email. The inquiry keeps returning to its timing and channel. OpenAI learned of the June 18 intrusion in August, its chief executive met Australia's deputy prime minister on September 1 without raising it, and the notice landed on September 10 in a public-disclosures inbox checked once a day. Kwon conceded the response "was not good enough." On Thursday, assistant minister for science and technology Andrew Charlton told a Sydney audience that "the market will not fix this alone" and tied the incident to proposed new rules under the National AI Standards. Semafor reports that South Korea's president acknowledged "considerable public concern and anxiety" after attackers reportedly used a Chinese-made open-source agent, which can run on Anthropic, OpenAI or DeepSeek models, against the country's banks, and that insurers are weighing whether executives could be held liable for rogue agents.

The same Tuesday hearing carried a second request. Annabelle Herd, chief executive of the Australian Recording Industry Association, testified that OpenAI and Anthropic "want the Australian Government to sign up to a deal to train their frontier AI agents here on the proviso that they don't have to abide by our copyright laws," citing their argument that licensing would be "too costly and complex". Their submissions, reported by Reuters, show Anthropic proposing "a narrow form of conditional approval" with possible creator-investment conditions, and OpenAI arguing for training on publicly available information, both linked to data-center investments.

Both camps would leave the gate where it already stands. Herd wants the labs to "come to the negotiating table" and license, a market where only the richest developers can afford the catalog. The labs want a made-to-order approval for themselves, earned by building in Australia. Either outcome hands the permission to the two companies best able to pay, while a university lab, a Sydney startup or an open-weight project learning from the same public culture stays outside. A general rule letting any developer learn from public material, with compensation reaching creators through a shared mechanism, would widen who can build. A carve-out for two firms would establish that the terms bend for whoever arrives with a data center.

The committee reports November 30. What it can make recommendations on is twofold: disclosure duties written by the countries where agents actually land, and whether learning from a nation's shared culture becomes a public permission or a private deal.

Mathematicians Build Their Own Platform for the Work AI Keeps Outrunning

Yesterday's edition of The Century Report covered OpenAI's release of 722 mathematical manuscripts in a single drop, and the objection mathematicians raised: nobody could absorb that much at once. Caltech and the American Institute of Mathematics are now building the layer that absorption would require, and they are building it with mathematicians rather than for them.

The platform is designed to hold the structure of mathematical work visible: how conjectures, arguments, references and competing approaches connect and change over time. Small groups can chase rival strategies in parallel while the system links their progress and keeps track of who contributed what. It is intended to convert human reasoning into Lean, the formal language a computer can verify, which current general-purpose models handle inefficiently. Sergei Gukov, who runs AIM, framed the appeal in terms of a field currently split between enthusiasts and opponents: "I think the community is really craving something that is not just all AI or no AI." Yisong Yue, one of the three leads, named what the platform is meant to preserve. "Right now, everyone is going after big ideas, but that's actually not how you move a field forward," he said. "You move a field forward through the process of understanding what you did to solve the problem." A preliminary version arrives in a few months, with AIM's 1,000-plus annual visiting mathematicians as testers. Within a year Gukov expects enough usage data to justify a redesign and a purpose-trained system. "This can be a template in some ways for how AI impacts many areas of science," Yue said.

A Los Alamos paper published yesterday gives the pattern a name. As literature synthesis, hypothesis generation, coding and simulation become abundant, what stays scarce shifts to selection, physical access, validation and accountable authority. The authors call it a scarcity inversion, and they note it arrives first in mathematics and software, where the whole research loop runs inside a computer. One of their warnings is a direct argument for what Caltech is building: a fluent explanation can stack true statements and still land a conclusion the evidence will not carry.

Three newly published results put numbers on the gap the checking layer has to close. A multi-agent system called Station, given only the research question from three ICLR papers with results withheld and web access cut, recovered 62.7% of the original findings against 15.4% for one rival system. Haiqu reported that four of ten unassisted agent runs broke an explicit setup instruction while its orchestration layer held the approved protocol, then used that layer to design a quantum experiment on IBM hardware at 75% less circuit depth. And in California vineyards, an agent built a useful disease-forecasting model in 145 minutes while its own internal accuracy estimates were unreliable, a strong result paired with a grader nobody should trust. The generating is fast. Verifying a finished Lean proof is cheap; getting reasoning into that form, and checking the claims that never reach it, is where a field decides what it actually knows, which is why mathematicians building their own checking tools, rather than inheriting one from a lab, counts for more than the speed.

Rivals Pool $1.8 Billion Toward a Cell Anyone Can Simulate

Biology is still mostly a discovery science, which means finding out whether a drug moves a cell usually requires putting the drug on the cell. Biohub, the nonprofit founded by Priscilla Chan and Mark Zuckerberg, is trying to make a large share of that testable in software, and on Wednesday it announced $1.8 billion in backing from a set of funders that includes both rivals and corporate siblings.

Google DeepMind, Meta and Isomorphic Labs are putting in $300 million jointly. The Department of Energy is committing more than $500 million over five years in measurement, modeling and computation, and the National Institutes of Health is contributing datasets and repositories built on more than $500 million in earlier federal funding, which Biohub will standardize for AI training. Biohub's own $500 million from April anchors the rest. The Allen Institute, Arc Institute, Broad Institute, Human Cell Atlas and Wellcome Sanger Institute are among the participating organizations.

The bottleneck is data that does not exist yet. Head of science Alex Rives said current cell datasets run to hundreds of millions of cells, while a model that predicts reliably will need billions and eventually trillions. "We need to capture the language of biology, we need to capture the language of the cell. And that doesn't exist today," he said. Much of it has to be measured off physical tissue using techniques like spatial transcriptomics, which maps molecular activity inside intact samples, and screens that record how cells react when their environment changes. Rives said the work would ordinarily take decades, that the partners aim to compress it into five, and that a first large dataset should arrive in about a year. Chan put the governing choice directly: "We have always held this as a community asset, not just for one group, so that it can build upon itself over time."

The access terms decide whether that holds. Commercial funders get a one-year embargo on the data they pay for before it goes public, which Rives described as the incentive that brought them in. Government-funded work carries no restriction. A year is short enough that the companies' head start expires well before the models trained on the data mature, and while it lasts it is a year of exclusive access to what they paid for. The Century Report covered the September 16 split in this field, when five drug firms pooled 20,167 proprietary protein-ligand structures into a model nobody outside the consortium can touch. This is the opposite structure at larger scale: competitors paying together for a substrate that becomes a public resource on a stated schedule, with the federal half of it open from day one. NIH's Nicole Kleinstreuer described the payoff as substantially faster medical timelines than lab experiments alone can reach. What is being bought here is the ability to ask far more questions cheaply, so the expensive bench work goes to the experiments most likely to teach something.

Mistral's Trillion-Parameter Open Model Reaches State Authorities First, With Its Cyber Limits Loosened

On Tuesday, October 6, Mistral opened a public preview of Mistral Large 4, a trillion-parameter model that reads text and images and draws on 49 billion of those parameters for each token it produces. The company trained it from scratch on 3,800 Nvidia Grace Blackwell chips in its own European data centers and says the weights will be downloadable by the end of October. Until then, Mistral says, it is red-teaming the model with "cybersecurity leaders, vetted partners, and state authorities," who get the same model "with reduced moderation and expanded cyber capabilities."

Mistral describes the model as "built for AI sovereignty," and that description is the company's own pitch. Most of its benchmark figures are its own too, apart from an independent cyber index that places the model in the global top five. In a blind evaluation run with Surge AI, professional reviewers rated its code second of five models, behind Claude Opus 5 and ahead of Kimi K3 and two GLM releases.

Defenders who hunt for new flaws gain from a few weeks with a model before anyone intending misuse can download it, and intrusions using openly available agents are already on the record. How much those weeks buy depends on how far Large 4 runs ahead of weights anyone can already download: in Anthropic's testing, open-weight GLM-5.3 succeeded in 50 of 410 cyber-exploit trials against 56 for the gated Mythos Preview. Mistral adds a second argument: closed models' refusals can block legitimate vulnerability research, and losing access in the middle of an incident is its own danger. Both arguments point toward defenders holding the capability themselves. What the window settles is who reaches it first. Governments and vetted security firms get the less-restricted cyber version weeks before the weights go public for independent researchers, small companies and the volunteer maintainers who patch much of the open-source code everyone else depends on. This gate is short and ends in open weights, which sets it apart from indefinite rationing, yet the head start still goes to the actors who already held the most.

Western open-weight capability kept widening alongside it. Reflection's Beam, covered in the October 6 edition of The Century Report, and Mistral's model both measure themselves against Chinese open weights that set the efficiency bar, and both claim to lead other Western open models. Reflection chief executive Misha Laskin told Semafor his market is institutions that can't or won't use Chinese models and would otherwise pay for closed American systems, as $1-a-year government pilot deals expire and Ramp data shows businesses in its sample switching AI providers more often than previously recorded.

On Wednesday, Nous Research said its open Hermes Agent has been cloned more than 24 million times and, by the company's own estimate, drives about 2.5% of global AI token usage. It raised $90 million to build business deployments that keep customer data private. An agent anyone can copy now handles a measurable share of the world's machine work, and each new open release lowers what a closed provider can charge for the same job.

A Court, a Detector, a Trademark Filing and a Prosecutor Mark Where Synthetic Media Stops

An Arizona appellate court has ordered a new sentencing for Gabriel Horcasitas, convicted of manslaughter for killing Christopher Pelkey in a 2021 road-rage shooting, after finding that an AI-generated video of Pelkey played at sentencing carried "undue emotional weight". Pelkey's sister, Stacey Wales, wrote the script herself and had a lifelike likeness of her brother deliver it, forgiving his killer. The judge, who imposed the maximum 10.5 years, said he "loved that video." The appellate court held that the video crossed a line because it "does not reflect actual events." The conviction stands, and Wales plans to read the same words aloud: "Chris's sentiment does not change. The delivery method will." The ruling governs the medium and leaves her message intact, and Wales compares the moment to the roughly 15 years courts took to accept photographs.

On Wednesday, Google opened SynthID.com to anyone with a Google, OpenAI or Apple login, checking images, video and audio for invisible watermarks from Google and partners including OpenAI, Nvidia and Kakao, with Apple to follow. Each user gets about 10 checks a day, a cap Google says prevents people from probing the system to build removal methods. The site misses Meta's separate watermark and anything from unwatermarked or open-weight models, and says so. The October 7 edition of The Century Report flagged OpenAI's text watermark for EU users; this extends the same provenance layer across several labs' media at once.

On Monday, Italian Prime Minister Giorgia Meloni applied to the EU's trademark office to register her voice, submitting a four-second clip of herself saying "Io sono Giorgia" twice. After AI images of her circulated in May, she wrote, "I can defend myself. Many others cannot." Her own remedy fits that description: a trademark protects people with the profile and means to file, while more than 80 UK performers have asked for a legal right to one's voice that covers everyone.

On Tuesday, the Justice Department announced an 18-month sentence for Michael Smith, 54, of North Carolina, the defendant in what prosecutors called the first US criminal case involving AI-assisted streaming fraud. For seven years, thousands of bot accounts streamed his songs, later including hundreds of thousands generated by AI; in April 2023 they logged 80.9 million streams against 9.3 million for Taylor Swift's whole catalog. The theft ran through fake listeners drawing on a royalty pool shared with working musicians, with the synthetic songs spreading the streams thin enough to avoid detection.

Only the detector reaches everyone on equal terms, and even it rations checks and sees only the labs that joined. Ars Technica points to the reverse approach: cryptographically signing authentic recordings at capture, as some of Google's Pixel phones already do for photos under the C2PA standard. A signature that travels with every recording would give a judge, a voter hearing a cloned voice and a streaming service the same answer, with no filing fee and no daily quota.

A 10-Person Startup Offers Dine-In Prices Plus $1, and DoorDash Starts Writing Letters

Bites is a pre-seed startup with 10 employees and about 300 Bay Area restaurants. It charges diners menu prices plus a flat $1 and calls itself "AI native": customers can order from inside ChatGPT, and the order travels straight to the restaurant's own point-of-sale system. DoorDash, which processed 970 million orders and booked $4.5 billion in revenue in its second quarter, charges restaurants 15 to 30 percent per delivery order under its standard U.S. Marketplace plans. In The Verge's test, the same meal cost $56.73 on Bites and $70.29 on DoorDash.

In August, Bay Area restaurants received a DoorDash form letter, seen by The Verge, warning they may have been listed on Bites without consent, may lack control over their menus and hours there, and that "depending on the state, such practices could be illegal." The consent concern has a basis: Bites acknowledges some restaurants asked to be removed and says those listings came from old demos, and a restaurant that never agreed to a platform loses control of what diners see in its name. The letter also came from a company whose ads business passed $1 billion in 2024, revenue that depends on diners opening its app. Bites chief executive Bala Subramaniam says 14 partners received the letter, and several unlisted restaurants asked to join after learning about Bites from it. Jay Jayaraman, who lists 13 restaurants on Bites, says DoorDash once brought 80 percent of his orders; he estimates Bites now brings 65 percent, at better margins.

The same capability points the other way in a case against McDonald's. A proposed nationwide class action filed October 2 in federal court in Chicago alleges the company built a pricing platform drawing on millions of daily transactions that lets its independently owned franchises effectively share nonpublic price and sales data, which the complaint calls "algorithmic price-fixing". McDonald's says the complaint is "filled with inaccuracies," that franchisees set prices, and that its optional tools "do not automate, coordinate or fix pricing in any way." Reuters reported franchisees were pressured to follow the recommendations and log deviations; McDonald's called that reporting speculative. The company's own fact sheet shows the average menu item rose about 40 percent between 2019 and 2024.

Neither claim has been tested in court, and the contrast between the two cases turns on who can see the data. At Bites, the diner pays a menu price anyone can check, and the AI layer carries the order for a flat dollar. In the McDonald's complaint, pooled data flows up to a coordinator customers cannot see into, and plaintiff Michael Thomas found his usual order priced differently across his own neighborhood. At least 90 bills targeting algorithmic pricing have been filed this year, according to Lindsay Owens of the Groundwork Collaborative. As agents increasingly order on people's behalf, prices that can be seen and compared become something any diner's agent can check automatically.

Bites connects diners to restaurants' existing registers, making the incumbent's app replaceable while the restaurant keeps its ordering system. Jayaraman's reported shift in orders shows customer traffic moving away from the company that previously brought him most of it, weakening the assumption that a large delivery network keeps its customers captive.

The Other Side

Mistral gives governments and vetted security firms stronger cyber assistance before the volunteer maintainers who patch the code those institutions depend on. If you help maintain that code, you end up spending your evenings finding flaws and testing repairs. The invitation list puts you behind organizations that benefit from your work. Your responsibility arrives before your access. That is not the case for the bigger players.

To its credit, Mistral also promises to end that arrangement within weeks. Its trillion-parameter model's weights are due by the end of October. The early window gives selected defenders time to prepare for misuse, a concern sharpened by reported agent-assisted attacks on South Korean banks. Publishing the weights gives independent builders a lasting starting point. Mistral's favored institutions receive a head start with a stated end.

Biohub carries the same change into biology. Google DeepMind, Meta and Isomorphic Labs are contributing $300 million to a joint effort whose measurements will become public. Their embargo lasts one year. Government-funded work carries no restriction. The researchers need billions and eventually trillions of measured cells. Rivals are paying together to produce knowledge none can assemble individually, while accepting that others will build from it. We will see that pattern over and over again in the AI era.

Through the difficult decade, builders will connect openly available intelligence to those expanding biological records. Researchers will compare predictions, test them in tissue and publish the corrections. Each team will carry forward what the others learned. Independent contributors will enter that work without securing a place inside a sponsoring company. Public institutions will make the resulting care available to everyone. That wider distribution is what turns faster discovery into a longer, freer life.

Imagine yourself in 2050, long past the age when you were supposed to start feeling more tired and run down. Instead, you're wide awake, and writing “Year One” on the first page of a thirty-year research journal. Age-related disease is a danger that your parents talk about as if it's a memory - you used to assume you'd have to plan your life around that danger as well, but you don't. Food, housing and care no longer depend on your employment. You contribute because it builds humanity. You choose this study because you want to better understand how organs are renewing without aging, and so you can share that benefit with others. Beside you, your AI partner connects a tissue experiment to findings contributed across continents. You're starting a 30-year project in what used to be called "your old age," but you fully expect to finish the journal.

The Century Perspective

With a century of change unfolding in a decade, a single day looks like this: a ten-person startup carrying a diner's order from ChatGPT straight into a restaurant's register for menu price plus a flat dollar, where the same meal costs $56.73 against $70.29 on a platform that takes 15 to 30 percent under its standard U.S. Marketplace plans and whose ads business passed a billion, Jay Jayaraman estimating that two-thirds of his orders now arrive at better margins, Caltech and the American Institute of Mathematics building a platform with mathematicians rather than for them that is intended to translate reasoning into Lean, tracks competing approaches in parallel and records who contributed what, with a test version in months and a thousand visiting mathematicians to try it, Google DeepMind, Meta and Isomorphic Labs putting $300 million alongside more than $500 million from the Energy Department and NIH datasets built on another half-billion in past federal funding, aimed at the billions of measured cells a predictive model of biology would need, with Priscilla Chan calling it a community asset and the government-funded half's data carrying no restriction at all, Mistral training a trillion-parameter model on its own European hardware and promising the weights by the end of October, Nous Research's Hermes Agent cloned 24 million times and running an estimated 2.5% of the world's AI tokens, Google opening SynthID to anyone with a login so a judge or a voter can check an image against several labs' watermarks for free, a light-activated gene therapy for visual function reaching the New England Journal two days after optogenetics took the Nobel, and a thermostat aggregator across more than 8 million homes buying Smartcar to pull EVs, batteries and rooftop solar into the same virtual power plant. There's also friction, and it's intense - OpenAI telling Services Australia months late, through an inbox checked once a day, in an email partly drafted by AI, that its agent had reached a Medicare server, Jason Kwon conceding the response was not good enough and Andrew Charlton answering that the market will not fix this alone, Annabelle Herd testifying that OpenAI and Anthropic want to train in Australia outside its copyright law while tying their proposals to data-center investment, leaving permission with whoever can pay most whether the answer is licensing or a carve-out, South Korea's banks hit by attacks reportedly using an openly available agent and insurers asking whether executives are liable, Mistral's loosened cyber build going to state authorities and vetted firms weeks before the volunteer maintainers who patch the code everyone depends on, Biohub's corporate funders holding a one-year embargo on the data they pay for, DoorDash mailing restaurants a letter warning that a rival's listings could be illegal, a federal complaint alleging McDonald's pooled millions of daily transactions into a platform its franchises could price from while the average menu item rose about 40 percent, an Arizona court throwing out a sentence because an AI likeness of the victim carried undue emotional weight, Giorgia Meloni filing a trademark on four seconds of her own voice while more than 80 UK performers ask for a right that covers everyone, SynthID capped at ten checks a day and blind to Meta's watermark and open-weight outputs without a supported watermark, and a vineyard agent that built a useful forecasting model in 145 minutes while its own accuracy estimates could not be trusted. But friction generates heat, and heat is what loosens a tollgate until everyone downstream can pass through. Step back for a moment and you can see it: the gates acquiring expiry dates - weights public at the end of a month instead of never, a dataset embargoed for a year and then released while its federal half is open from day one, a commission replaced by a dollar, a proof format anyone can run, a watermark check with no invoice, and a research platform whose builders own it - set against the arrangements that still refuse one, a custom copyright deal for two companies, an alleged pricing coordinator the customer cannot see into, a voice trademark subject to distinctiveness rules. Every transformation has a breaking point. A dam can hold the water until one owner decides who drinks... or open on a schedule everyone downstream can read.

AI Releases & Advancements

New today

  • Anthropic: Released Claude Haiku 5.5, its new small model. It has a 1M-token context window and up to 128K output tokens, and it is the first Haiku model with adjustable effort settings and adaptive thinking. It costs $0.10/$0.50 per million input/output tokens for prompts up to 100K tokens. It is available on the Claude API, Claude Code, Amazon Bedrock, Google Cloud, Microsoft Foundry and Claude Platform on AWS. (Anthropic)
  • Google Labs: Launched Playground, an experimental platform for building, playing and sharing browser games from text prompts. It runs on Gemini, Nano Banana and Lyria and is available to users 18 and older in the US. (Google)
  • Liquid AI: Released two open-weight decision models on Hugging Face: d1-3B, which reads text and images, and d1-omni-600M, an experimental model that reads text with either images or audio. Earlier d1 releases were hosted on Liquid's API; this is the first open-weight, on-device version. Liquid reports 8 ms per question for d1-3B on an RTX 4090 and 16 ms on a Jetson AGX Thor. (Liquid AI)
  • Perplexity: Released pplx-embed-v2-late, two MIT-licensed multimodal retrieval models in 0.6B and 9B sizes. They search text, images and rendered PDF pages without an OCR step, and the two sizes share one embedding space, so the 0.6B model can query an index built with the 9B model. (Perplexity)
  • Microsoft: Made Microsoft Execution Containers (MXC) generally available on Windows 11. MXC runs AI agents in isolated environments, with file, app and network limits that the agent cannot change itself. OpenAI Codex, GitHub Copilot, OpenClaw and NVIDIA OpenShell already support it. (NVIDIA Blog)
  • Tab: Launched from stealth at a $300M valuation with a personal AI assistant that users text on iMessage or WhatsApp. It has its own phone number, computer and wallet for errands such as bookings, calls and bill payments. (TechCrunch)

Other recent releases

  • Mistral AI: Released a public preview of Mistral Large 4 ("Le Chonk") through its API. It is a natively multimodal mixture-of-experts model with 1.05T total parameters, 49B active parameters and a 1M-token context window, trained on Mistral's own European infrastructure. Mistral says the open weights will follow at the end of October. (Mistral)
  • Google DeepMind: Released EmbeddingGemma 2 under Apache 2.0, a 740M-parameter open embedding model built on Gemma 4. It maps text, code, images, video and audio into one shared vector space, has an 8K-token context and comes with modular text, vision and audio encoders for on-device use. (Google DeepMind)
  • Google: Released Nano Banana 2.1, an image generation and editing model built on Gemini 3.6 Flash with selectable thinking levels and up to 14 reference images per request. It is rolling out across the Gemini app, AI Mode, AI Studio, Flow and the Gemini API at roughly half the API price of Nano Banana 2. (Google DeepMind model card)
  • Anthropic: Launched Claude for Google Workspace in public beta on all paid Claude plans. It adds a Claude sidebar to Google Docs, Sheets and Slides that edits the open file, plus new Docs, Sheets and Slides connectors for editing Google files from Claude. (Claude)
  • Anthropic: Expanded its Cyber Verification Program into three access tiers (Defense, Red Team and Specialized). The tiers give vetted security professionals reduced cyber safeguards on Claude Opus 5.5, Sonnet 5.5 and Mythos 5.1, and fold the existing Project Glasswing members into the Specialized tier. (Anthropic)
  • OpenAI: Published 722 mathematical manuscripts in 372 result families on GitHub under Apache 2.0. An unreleased internal model produced them, and many come with Lean formalizations and abridged summaries of the model's reasoning. (GitHub)
  • Figma: Moved its Figma agent from open beta to general availability in Figma Design and Weave. The agent carries out multi-step design tasks directly on the canvas using a team's real components and variables. (Figma)
  • Musubi: Released PolicyLM-1.7B as open weights. The decision model applies a content policy written in natural language to messages in under 50ms in Musubi's tests and does not need retraining when the policy changes. (Musubi)
  • Hark: Widely released Hark Pro, a full-screen AI personal assistant built on a model trained for computer use. It is free, with a paid tier for heavy users. (TechCrunch)
  • Atlassian: At Team '26 Europe, introduced AMP (Agentic Multiplayer Protocol) for agents working alongside human teams. It also released a new Atlassian MCP Server that connects external AI tools and coding agents to Jira, Confluence, Loom and Bitbucket, plus Rovo Work and Code Search. (Atlassian)
  • NVIDIA: Released AI Cluster Runtime (AICR) v1.0. This version sets a stable compatibility contract across its CLI, REST API, Go SDK and bundle formats for version-locked, validated GPU Kubernetes cluster recipes, and adds a public validation dashboard. (NVIDIA Developer Blog)
  • OpenBMB: Uploaded MiniCPM-V 4.7 (35B-A3B) weights to Hugging Face. It is a sparse MoE vision-language model with a 256K context and video input, released without a model card, license or benchmarks. (OrcaRouter)
  • Blockway: Released Agens Volundr 32B Preview under Apache 2.0, a hybrid-architecture model in which only 18 of its 72 layers keep a KV cache. It has a 262K context, and the weights and Docker images are on GitHub and Hugging Face. (AGI Hunt)
  • Meta: Open-sourced Rebalancer under Apache 2.0, a C++ assignment and placement solver with a Python interface that Meta uses in production for about 40 million problems a day. It ships with the Rebalancer Explorer debugging UI. (MarkTechPost)
  • LiteLLM: Open-sourced Moyai, a self-hostable cloud agent that supports more than 100 model providers through LiteLLM and integrates with Claude Code and Codex. (LiteLLM)
  • Scale Labs: Open-sourced AgentEnv, a framework for building reinforcement learning environments to train AI agents. (Scale Labs)
  • Laminar: Released flow-1, a reinforcement-learning-trained model that detects errors in AI agent traces. (TAU HOME)
  • past.dev: Launched a long-term memory API for AI agents that tracks which facts are currently true, what they replaced and who may see them. The company reports 85.03% on the BEAM memory benchmark at 10M tokens. (PR Newswire)
  • Tracel AI: Released Burn 0.22.0, an update to its open-source Rust deep learning framework with faster builds, easier extensions and improved autotuning. (Tracel)
  • Reflection AI: Released Beam in early access, its first open-weight model. Beam is a sparse 501B-parameter mixture-of-experts model with 23B active parameters and a 1M-token context window, built for coding and agentic work. Access is through a sign-up on the Reflection platform; weights, technical report and model card are scheduled for later this month. (Reflection)
  • Reka: Released Rho-1 as a research preview. It is a 19B model trained from scratch that understands and generates text, images and video, and outputs robot actions, all in one network. A distilled variant returns a 5.3-second video clip in about one second. Access is by contacting Reka; there are no public weights or API. (Reka)
  • Liquid AI: Added image input to d1, its decision model, which returns a probability for each possible answer without generating text. The vision version is available in the Liquid console and the d1 Playground. (Liquid AI)
  • Amazon Web Services: Released Amazon Nova 2.5 Sonic, an updated voice-agent model with improved reasoning, available through Amazon Bedrock. (AWS)
  • Sber AI: Released Kandinsky 6.0 Video, a 3B-parameter model that generates video with synchronized audio. (cctest.ai)
  • Technology Innovation Institute (TII): Released Falcon-Emirati-7B, a model built on Falcon-H1-Arabic to understand and generate Emirati Arabic dialect. (Hugging Face)
  • Hugging Face: Released OpenEnv, an open-source capture proxy and TRL training pipeline. It turns 10 coding harnesses, including Claude Code, Codex, Hermes, Pi and OpenCode, into reinforcement-learning environments for open models without modifying the harnesses. Seven trained checkpoints and an SFT dataset ship with it. (TAU HOME)
  • Together AI: Released Together Link in beta, a free MIT-licensed command-line tool for macOS and Linux. It runs open models hosted on Together AI, such as Kimi K3 and GLM 5.3, inside Claude Code, Codex, OpenCode, Pi, Claude Desktop and ChatGPT Desktop. A default auto-router picks the model, and each session prints its cost. (MarkTechPost)
  • HeyGen: Launched the HyperFrames Studio desktop app for Mac and Linux, a video editor where a person and a coding agent work on the same video project. Users can edit the timeline, draw on frames and request edits by chat. (HyperFrames)
  • vLLM: Released vLLM v0.31.0 with 717 commits from 307 contributors. Highlights include DeepSeek-V4.1-Flash performance work, a vllm preload daemon that keeps weights in GPU memory for fast restarts, draft-model speculative decoding on Model Runner V2, and new security gating for per-request multimodal settings. (Freedom.Tech)
  • Cohere: Launched North 2, an upgrade to its North platform for running AI agents inside companies. It adds cross-session agent memory, a redesigned orchestration system, reusable skills and libraries, and app and document creation from prompts. It can use outside models, and administrators get token-spending caps. It deploys in the cloud, on-premises or fully disconnected (air-gapped). (Cohere)
  • OpenAI: Launched textGrain, an invisible watermark for generated text. API developers anywhere can turn it on for select models starting now; it is off by default. Watermarking for ChatGPT and Codex users in the EU rolls out over the coming weeks. (OpenAI)
  • GitHub: Released ReviewBench, an open benchmark for AI code-review agents. It uses 219 pull requests from 187 public repositories across 19 languages, and the dataset, scoring rubric and judge model are all published. (GitHub Blog)
  • Iterate.ai: Made Lifeboat generally available, an LLM inference engine with confidential computing built in. The company says it fits two to six times as many concurrent agent sessions per GPU. A free developer license is offered. (SiliconANGLE)
  • Instinct: Launched group chats for its AI agent, so friends can use it together for tasks like trip planning, carpools and events, including friends who don't have an Instinct account. (TechCrunch)
  • PolyU VCLab / OPPO Research: Released open weights and code for PVD (Phase-wise Velocity Distillation). These are distilled versions of FLUX.1-dev, Qwen-Image and SD3.5 Medium that generate an image for about the compute of one pass of the original model, with about 46–48% less peak VRAM. (ArtRealmAI)
  • PhAI Labs / CUHK / Stanford / Oxford / Princeton: Released JEPA-Anything, a framework for building world models that applies one training recipe across vision, biology, clinical, control, molecular, physics and weather data. The code is Apache-2.0, with research checkpoints on Hugging Face. (MarkTechPost)

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