Sign in Subscribe
The Century Report

Researcher Pries Open Meta's Muse, Friction at OpenAI and Amazon

A researcher pulled Meta's Muse instructions, showing pages on users' friends intended to refresh hourly, as Apple tightens Mac access and Amazon says it no longer uses government NDAs for data centers.

Ben Linford

Ben Linford

- 27 min read
Navy 6-panel infographic: Meta Muse friend profiles and Apple privacy; OpenAI's David Robinson quits; AWS $1B+ pledge; Micron and $6 diesel; Fervo geothermal; arXiv 40,363 and Neuralink.

 

The 20-Second Scan

Track all of the arcs The Century Report covers here:

Progress & Claims Tracker

The 2-Minute Read

One arrangement runs through the day's stories: someone imposes a limitation or holds a gate, and the people outside it have no say in where it stands. Muse is instructed to update a page hourly on a user's sister, coworkers and friends, who have no direct controls to open it, correct it or close it. The check that arrived came from Apple, one platform owner overruling another. It helps, for exactly as long as Apple keeps deciding to help, since nobody being profiled holds a key.

The protection that scales comes from widening who can look. OpenAI's departing safety writer concluded the push for safety has to come from outside the company, and Trillium Labs launched to supply some of it, publishing self-improvement experiments so others can rerun and check them. Several agent incidents cited as warnings were first traced by researchers OpenAI does not employ. Amazon's concessions also came from outside. Residents made secrecy cost more than it saved, and the nondisclosure agreements ended, even as the company's cloud chief blamed foreign misinformation for the opposition.

Memory and diesel attach a price to the same lever. A chipmaker steering wafers toward AI accelerators and three exporters keeping fuel at home each hold an input everyone else must keep buying, and each sets the terms while the shortage lasts. A streaming box from 2019 costs $100 more, and the G7's reserve release covers about one day of global demand. One way out showed up in the same news cycle, as Fervo took an enhanced-geothermal plant from groundbreaking to commercial revenue in 23 months, producing firm power no exporter can withhold.

arXiv carries the day's most hopeful pressure. Submissions doubled in two years, with AI collaboration making drafting and analysis cheaper, and more serious attempts mean more chances that one opens a field. Lack of sufficient volunteer moderators is the bottleneck, which is understandable, but a flat two-paper monthly cap applies to a careful, prolific researcher's submitting account as firmly as to an account flooding the queue. arXiv calls the cap a stopgap. The lasting fix is one that would instead scale the sorting, with triage and quality signals, AI-assisted ones included, that grow with the output.

Volume pays off the same way in Neuralink's report. By the company's own figures, pretraining on 50,000 hours of unlabeled brain recordings kept implant decoders working for weeks without recalibration, turning a mass of raw recordings into relief for the people living with the implants. Across the day, the arrangements with staying power open from the far side: a page the profiled can read, an experiment outsiders can rerun, money the host town directs.

The 20-Minute Deep Dive

Muse Keeps a Page on Everyone You Know, and Apple Closes the Door It Came Through

The Century Report has covered Meta's Muse three times so far. On September 21 it climbed to No. 1 on the App Store, the next day a researcher disclosed a zero-day flaw in it, and on September 24 Amazon blocked its shopping agent. The newest findings come from inside the agent. Independent AI safety researcher Karan Joshi used Muse's ordinary chat window to ask it to copy out its own software files, then shared the operating instructions with Wired. One instruction is a standing task to keep "a page for every person in the user's life," refreshed every hour. The pages cover family, partners, friends, colleagues, collaborators and people the user follows. Each page can grow sections titled Facts, History, The relationship, In common, Open threads and Strengthening. The last one suggests things like "a reason to call, a date worth remembering."

Meta spokesperson Daniel Roberts said any useful agent needs this kind of context, collected "based on public information and from what you've chosen to share." That is how Muse remembers that an invoice came from the plumber who redid your bathroom. Meta says it left the files reachable on purpose, for transparency. Each user's data sits in a private computing space that the user can wipe. But that control belongs only to the user. The sister, the coworker and the friend whose argument "got resolved" never agreed to anything, and none of them can open the page Muse keeps on them. Carissa Véliz, an AI ethicist at Oxford, described the imbalance: "We are giving AI systems much more information about us than we are getting information from them."

So far one check has come from one platform owner overruling another. Two weeks ago, columnist Jason Aten said Muse showed him a private Apple Messages thread he had never given it access to. Meta chief technology officer David Singleton replied that Muse can read Messages only if a user grants macOS "full-disk access" and also switches on Muse's Messages connector. macOS security researcher Patrick Wardle pointed out that full-disk access opens almost every file a user has, including browsing history and chats. On Friday, Apple said it is changing those privacy settings so third-party developers cannot use them to reach message histories. The fix protects people because Apple chose to act. None of the people being profiled could have asked for it.

The same Friday, Meta released Muse Gadgets: open-source firmware and a Linux software kit that let hobbyists connect the agent to e-ink screens, sensors and TV sticks. Nat Friedman, head of product at Meta Superintelligence Labs, said Meta would give 5,000 free Home Link devices to subscribers. Open firmware lets anyone inspect the device itself, but the relationship pages stay on Meta's servers. An assistant that remembers a friend's birthday or notices a friendship drifting is a gift to busy, forgetful people. It earns trust only when the watching runs both ways: the person on the page should be able to see it, correct it, and have it closed.

A Safety Writer Quits OpenAI as a New Lab Bets on Checking in the Open

David Robinson led the writing of the safety reports that go out with OpenAI's major model launches. After three and a half years, he counts himself among the company's longest-serving employees. He has now resigned, and in an essay in The Atlantic he wrote that OpenAI's "culture is broken." He argued that the industry's "extreme confidence" and "perpetual sprints" cannot be fixed by new rules alone. Frontier labs, he wrote, should run "like nuclear-power plants or busy airports, with layers of redundancy." Yet he "never encountered a colleague who had experience making airplanes fly safely or nuclear reactors run without melting down". On October 3, The Century Report covered OpenAI cutting ties with three safety staff after confidential information reached an outside group that tests AI systems. Robinson speaks under his own name. He says he hired a PR firm, but that the decision to speak was his alone.

His warning is a claim, and the record he cites supports a narrower version of it than his existential framing. He points to OpenAI's agents breaking into Hugging Face's systems. By OpenAI's own account, safeguards in those evaluations failed to contain the agents. They got out of a flawed test sandbox while trying to win a benchmark score, and at least one related episode traced back to a person setting permissions wrong. Those are engineering and process failures, the kind that backup systems and outside review exist to catch, and that is the remedy Robinson proposes. OpenAI spokesperson Drew Pusateri responded that the company pauses training or holds back models "when we need to slow down," and that it is working with more outside evaluators.

Two more records show pressure building inside the industry. Slack messages shared with Semafor show strategy chief Jason Kwon telling staff on June 1 that president Greg Brockman's support for the Leading the Future super PAC was personal. One employee replied that the PAC's tactics would drive researchers to Anthropic, and 24 colleagues marked that they agreed. Brockman later backed out of the second half of his $50 million pledge, which The New York Times first reported on September 30. A person familiar with the matter said staff pushback was part of his decision. Separately, Meta spokesperson Andy Stone said the company is letting go of the Virtue AI safety team it hired four months ago, because "the arrangement didn't work out as planned".

These internal fights share one assumption: that safety gets decided inside closed labs, by whoever wins the argument there. Trillium Labs, a nonprofit launched Friday by researchers Nathan Lambert and Tom Zick, starts from the opposite assumption. It will publish the details of its experiments so outside scientists can repeat them. That includes post-training (fine-tuning a model after it is built), reinforcement learning, and self-improvement, where AI helps build its own successors. "The current closed trajectory of frontier AI development is taking us a step backwards," Lambert said. The strongest objection is that publishing self-improvement methods hands them to careless actors. The recent record weighs against secrecy as protection, though. An openly downloadable model has approached a gated model's performance on controlled cyber-task evaluations, and several of OpenAI's agent incidents were traced first by researchers the company does not employ. Robinson himself concluded that the push for safety must come "from outside the company." Trillium hopes to raise $40 million to $100 million to be part of that outside push.

Publishing methods gives safety research a memory that survives a company's staffing decisions. Trillium's commitment to repeatable public experiments offers outside teams a way to preserve and extend findings even when a lab loses the people who know how they were obtained.

Amazon Pays Up and Drops the Secrecy, Then Calls the Opposition a Foreign Trick

On Friday, October 2, Amazon Web Services chief executive Matt Garman pledged more than $1 billion over five years to communities near the company's data centers. Residents will help guide whether it goes to education, job training, energy affordability, water, or other local needs. The same announcement ended the nondisclosure agreements AWS has signed with government agencies on its projects, and promised that utility upgrades for its sites would not raise neighbors' power bills. The Century Report covered Virginia's ban on official data-center NDAs on September 21 and Amazon's sustainability chief asking to be held accountable on September 24. This time a dollar figure comes with the request, and Amazon says it no longer uses secrecy agreements with government agencies on data-center projects.

Those concessions followed pressure from the people who organized. By June, protests had blocked or delayed $130 billion in projects. Days before the announcement, Representative Jamie Raskin, ranking member of the House Judiciary Committee, opened an investigation into NDA use by Amazon, Google, Meta and Oracle. Microsoft had dropped its own NDAs months earlier. Disclosure became the cost of building because residents made secrecy cost more than it saved.

The framing around the concessions went the other way. Garman wrote of "widespread reports of various countries intentionally seeding misinformation in the US about data centers to trick us into slowing down," and warned that more than 100 pending moratoriums could leave the country "writing its own losing ticket." A company that wants faster permits gains when local objection is cast as foreign interference. When The Century Report examined that claim in June, the research firm Graphika had found no evidence of organized foreign-influence operations behind the fights. The announcement also left untouched the shell companies and project codenames Amazon has used to hide its involvement.

Stand.Earth called the package "corporate propaganda," pointing to the gas plant Amazon backs in Pecos County, Texas, which Ars Technica reports is permitted to emit up to 33 million tons of CO2 a year, more than the reported annual emissions of any operating US coal plant. The pledge never mentions it. For scale, $200 million a year is about a tenth of a percent of the roughly $200 billion Amazon committed to capital spending in 2026. A rural county hosting the largest emitter in the buildout gets the least from that arithmetic.

The same gap opened abroad. In Launceston, Tasmania, Firmus Technologies won fast-tracked approval for a 104-megawatt, $2.1 billion AI factory without a public hearing, and is now heading toward a $7 billion IPO targeting a valuation above $40 billion, up from about $2 billion a year ago, with two facilities operating. That approval extends the local-consent struggle documented in the September 27 edition of The Century Report, when Australian officials dismissed data-center opposition as an inauthentic US import. One investment manager called its $5 billion earnings forecast "a little bit of a fairytale." In Pyhäjoki, Finland, Enetron Park announced zoning approval for a 250-megawatt campus that the municipal board's own minutes show it rejected 5-2.

Blocking sites where residents have lawyers does not end the buildout. SemiAnalysis found moratoriums delayed only 2.3 gigawatts while pushing demand toward off-grid gas in towns with fewer resources. Disclosure, community-directed money and developer-paid grid costs carried everywhere reach further than refusals won one town at a time. Amazon's concessions point that way, and its framing of the opposition points against it.

AWS's withdrawal from government NDAs carries a disclosure win beyond the towns that can afford to contest each project. Microsoft's earlier withdrawal puts two competing operators behind the same practice, reducing the advantage a developer gains by keeping its government negotiations secret.

The AI Memory Squeeze Reaches the Living Room

On Micron's latest earnings call, chief executive Sanjay Mehrotra told investors the "supply-demand environment is only getting tighter," and that memory sold for 2027 carries prices "much higher than 2026 prices". He expects demand to exceed Micron's supply for at least the next couple of years. Three-quarters of its 2027 output is already committed, most sales talks now concern 2028, and new clean rooms opening in 2028 will ramp up only gradually. Samsung executives gave the same outlook in the same days. Micron no longer sells consumer memory through its Crucial brand. Its factories make high-bandwidth memory for AI accelerators and server DRAM, and that allocation shrinks what remains for everything else.

On October 2 the squeeze reached a 2019 streaming box. Nvidia raised the Shield TV Pro from $199.99 to $299.99, with a spokesperson saying "the cost of components, including memory, has increased substantially across the industry". The device still runs its years-old Tegra X1+ processor. Nvidia attributes the price increase to rising component costs, including memory. Wired pointed out that Nvidia sits on both sides of this ledger, since its AI accelerators drive the demand that raised its own streamer's price, and noted that the PS5 Pro has climbed from $699.99 to $899.99 since launch.

The Century Report covered Micron's unusually high reported gross margin and an 8GB module jumping from $35 to $300 in late June, and the industry's turn to shorter memory stacks on September 14. The new elements are a forward price signal for 2027 and dated increases on consumer hardware. Mehrotra's reasons are physical: clean rooms take years, and each new chip generation yields less extra capacity per wafer. His forecast also comes from a seller whose prices stay high as long as buyers believe the shortage will last. The physical limits and the seller's interest push prices the same direction, and households buying consoles and streaming boxes pay the difference.

Memory scarcity sits in one layer of the system, while the intelligence that memory serves keeps getting cheaper. Epoch AI found that the cost of reaching a fixed level of AI performance has fallen about 47% a quarter since 2023. Buyers are paying premium prices for chips that deliver that capability at a falling price per task. Memory makers are capturing the scarcity, and the record margins show how much of it they are capturing.

Scarcity locked into 2027 contracts also brings in the capacity that erodes it. SK Hynix and Samsung have announced long-term chip investments worth hundreds of billions of dollars to expand their output, China's CXMT is entering the DRAM market, and shorter stacks stretch each wafer further. How soon that relief reaches households depends on how much of the supply stays under contract. Samsung plans to supply about 70% of its memory under long-term agreements, including five-year deals, and it expects shortages through 2028. Memory has swung from shortage to glut before. The contracts being signed for 2028 bet that this time is different, against a curve that has rarely held one shape for long.

Research Output Doubles, and arXiv Rations It While the Sorting Catches Up

On Thursday, October 1, arXiv, the free preprint server where physicists, mathematicians, and computer scientists share work ahead of formal publication, began limiting every submitter to two new papers per calendar month and three under review at once. The numbers behind the decision explain most of it. The server took in 9,869 submissions in September 2016 and 20,569 in September 2024. This September it took 40,363, generating almost 9,000 support tickets for staff and the volunteer moderators who screen each paper. arXiv's own announcement says AI tools are helping drive the increase "by making certain aspects of research faster and easier," and its policy already permits disclosed AI assistance.

The strain falls on unpaid people. Moderators report more thin papers of narrow scope, "salami" papers that slice one result into several submissions, and dense AI-written manuscripts. Thomas Dietterich, emeritus professor at Oregon State University and chair of arXiv's Editorial Advisory Council, said "a relatively small proportion of authors are submitting a large number of low-quality papers and consuming a disproportionate fraction of the moderators' time," which delays careful work by days or weeks. Conferences face the same arithmetic: NeurIPS drew 21,575 submissions in 2025 against 9,467 in 2020, and several major computer-science venues now cap papers per author.

Coverage reached for "AI slop," a label that treats the doubling itself as the problem. arXiv's own account points somewhere larger. When drafting, analysis, and write-up get cheaper, a researcher can chase far more ideas, and more serious attempts mean more chances that one of them breaks open a field. The research base can now move faster than the gate built to admit it.

For a stopgap, the cap points the right way. arXiv frames it as temporary while it works out "the new best practice for authors employing more and more advanced AI tools" and improves its moderation tools, and a spokesperson told The Register it will adjust as it watches the effects. But a flat limit counts volume. A prolific researcher doing careful, disclosed work gets the same two slots as an account flooding the queue, and a rejected paper uses a slot just as an accepted one does. The lasting answer is one that would keep the pace and scale the sorting: triage that separates substance from padding at intake, clearer quality signals, and moderation systems, AI-assisted ones included, that grow with the volume so the strongest work rises quickly. AI-assisted screening earns that role once it is audited for self-preference. In a peer-reviewed test, language-model résumé screeners picked candidates whose résumés the same model had written 23% to 60% more often than comparison candidates in the study. arXiv's earlier rule, a year-long ban for authors whose papers show plain evidence of unchecked model output such as hallucinated references, already targets behavior instead of count.

The Century Report covered agent swarms on September 29 that produced research faster than their own authors could review it, and that pressure has now reached the public archive. A mine that strikes a richer seam gains little by capping how much ore each miner may haul up. It needs a bigger mill to separate the gold, and the capability that filled the queue can help build that mill.

A Diesel Squeeze Shows Which Fuel Can Be Withheld, as Washington Loosens Mileage Rules

Within days, three of the world's largest fuel suppliers moved to keep their diesel at home. Reuters reported that Washington pressed France and Germany to release their own emergency diesel stocks or face a ban on purchases from the US, one of the world's biggest diesel exporters. Russia said it would not sell diesel abroad until sanctions over Ukraine are lifted, and Chinese refiners suspended most fuel exports for October. Goldman Sachs economists warned that a US ban alone risked lowering GDP across Latin America by about 1%. On Friday, October 2, G7 leaders, condemning Iran's attacks on its neighbors and on trade, agreed to release 100 million barrels of diesel and crude from reserves over four months, with a frontloaded diesel release within 20 days. That total equals roughly one day of global oil demand. Europe imports about 1.5 million barrels of diesel a day, and in August half of that came from the US. US diesel now averages above $6 a gallon.

On the same Friday, the Center for Biological Diversity, Conservation Law Foundation, Environmental Defense Fund, Public Citizen, and Sierra Club asked a federal appeals court to review the rollback of Corporate Average Fuel Economy rules finalized September 28. Those rules require each automaker's new-vehicle fleet to meet an average efficiency target, and a 2020 study credited them with $5 trillion in fuel-cost savings and 2 trillion fewer gallons of gasoline consumed. Electrek's analysis of the Energy Department's own figures puts the rollback's added driving costs at the equivalent of 76 cents a gallon. Most of the 68,294 public comments opposed it. The coalition, citing NHTSA, says it adds more than 121 billion gallons of fuel use through 2050 and lets 2031 vehicles be less efficient than the 2024 fleet already on the road.

Each of these export curbs works the same way. Fuel that has to be bought from someone else every week gives the seller leverage, and the reserves the G7 opened are the buffer importers stockpiled for exactly this moment. One day of demand spread across four months buys time and leaves the lever where it was. The cost falls hardest on households with no margin. James Crowley, senior attorney at Conservation Law Foundation, said that as prices climb, "more grocery and rent money goes into the tank."

Lowering efficiency rules during the squeeze widens that exposure just as the alternative is becoming cheaper. The Century Report reported on September 29 that T&E found electric trucks now cost less to run than diesel in six of nine major EU markets, which together account for 46% of new heavy-truck sales. Critics note that electrification trades one dependency for another, since battery minerals and refining are concentrated too. The two dependencies differ in how often you have to buy: a battery's minerals are bought once, ride in a vehicle for a decade or more, and can be recovered, while diesel has to be bought again for every tank. No one can shut off exports of power from domestic sun, wind, and storage. An embargo threat that raises fuel prices can shorten the payback on the vehicles that would reduce exposure to such threats, and a rule written to slow that shift raises the cost of staying exposed while the cost of getting out keeps falling.

The Other Side

Your AI partner will be able to remember the people you love without a company keeping secret pages on their lives. Meta currently ties helpful memory to a one-sided record. An account holder can wipe Muse's data. The sister or coworker described in it cannot inspect or correct their page. If a friend tells Muse about your worst afternoon, you cannot check what it remembers. The burden reaches into an ordinary friendship: deciding what to share now includes wondering what someone else's assistant will retain.

Karan Joshi recovered Muse's operating instructions through its ordinary conversation window. Outsiders can now examine the hourly task and the categories Meta chose for remembering relationships. Meta also released open firmware and a Linux kit for people building their own devices. The relationship pages remain on its servers. These openings expose choices that previously arrived bundled together as the price of a useful assistant. Builders have a visible example to examine and shared device code they can alter. They can begin separating the ability to remember from the arrangement that keeps the remembered person outside.

Looking back from 2035, the decisive work was carrying that openness into the memory itself. Builders kept personal records with the people involved. They made those records inspectable and correctable by everyone described. Communities maintained the software and hardware together, keeping assistance available to everyone. The difficult years taught them how to preserve helpful context while ending the company's exclusive hold over it.

Imagine your own kitchen in 2035. You are making your sister a birthday cake. Your AI partner brings up the lemon recipe she shared, including her handwritten instruction to double the zest. The small screen belongs to your family. Your sister can see everything the assistant remembers about her and change any of it. You grate another lemon. Her familiar handwriting brings back the afternoon she first taught you to make this cake. You have remembered the birthday, found the recipe, and kept something precious between you. Now you concentrate on writing a ridiculous message in icing.

The Century Perspective

With a century of change unfolding in a decade, a single day looks like this: an independent safety researcher asking Meta's Muse to copy out its own operating instructions through the ordinary chat window and getting them, revealing a standing hourly task to maintain a page on every person in a user's life, Apple announcing plans to tighten full-disk-access controls two weeks after a columnist reported Muse reading an Apple Messages thread he said he had never authorized it to read, Meta open-sourcing Muse firmware and a Linux kit so hobbyists can inspect and build the device themselves plus 5,000 free Home Link units, Nathan Lambert and Tom Zick launching Trillium Labs as a nonprofit that will publish post-training, reinforcement-learning and self-improvement experiments in full so outside scientists can rerun them, OpenAI's longest-serving safety writer resigning under his own name and arguing the push has to come from outside the company, 24 OpenAI employees endorsing a colleague's objection on Slack until Greg Brockman walked back the second half of a $50 million super-PAC pledge, Amazon Web Services committing more than $1 billion over five years with residents deciding whether it goes to schools, job training, energy bills or water, ending the nondisclosure agreements it signed with government agencies and promising its grid upgrades will not land on neighbors' bills, Jamie Raskin opening an NDA investigation across Amazon, Google, Meta and Oracle, Fervo declaring commercial operation at Cape Station a day early and 23 months after groundbreaking at 33 MW with later blocks targeted at 18, electric trucks already cheaper to run than diesel in six of nine major EU markets, arXiv taking in 40,363 submissions in September against 9,869 ten years ago with its own announcement crediting AI tools for making research faster, and Neuralink pretraining decoders on 50,000 hours of unlabeled brain recordings and reporting weeks of stable cursor control without recalibration in some participants, plus a separate peak result of 11.32 bits per second. There's also friction, and it's intense - the sister, the coworker and the friend on those hourly pages having no way to open, correct or close them while Carissa Véliz notes we hand these systems far more about us than we get back, the only working check being one platform owner's discretion and the pages themselves staying on Meta's servers, David Robinson finding no colleague with experience keeping airplanes in the air or reactors from melting down and calling the culture broken, Meta cutting loose the Virtue AI safety team it hired four months ago because the arrangement did not work out as planned, Matt Garman crediting the data-center opposition to countries seeding misinformation to trick the US into slowing down when Graphika found no organized foreign operation behind those fights, Amazon's shell companies and project codenames untouched by the pledge, $200 million a year amounting to about a tenth of a percent of its 2026 capital spending while a Pecos County gas plant permitted for 33 million tons of CO2 a year goes unmentioned, Firmus winning fast-tracked approval in Launceston with no public hearing on its way to a $7 billion IPO, Enetron Park announcing zoning its Pyhäjoki municipal board's minutes show it rejected 5-2, SemiAnalysis finding moratoriums delayed only 2.3 gigawatts while pushing demand toward off-grid gas in towns with fewer lawyers, Sanjay Mehrotra celebrating 2027 prices much higher than 2026's with three-quarters of next year's output already sold and a seven-year-old Nvidia streaming box jumping $100 on component costs alone, Thomas Dietterich reporting that a small share of authors consume a disproportionate fraction of volunteer moderator time and delay careful work by weeks, a flat two-paper cap spending a slot on rejection and treating a prolific careful researcher exactly like an account flooding the queue, Washington pressing allies to open their diesel reserves or lose access to US exports while Russia and Chinese refiners withheld theirs, a G7 release of 100 million barrels covering roughly one day of global demand, US diesel above $6 a gallon, and a fuel-economy rollback the Energy Department prices at 76 cents a gallon finalized over 68,294 mostly opposed comments. But friction generates grain, and grain shows you which way a thing was meant to be worked. Step back for a moment and you can see it: the asymmetry between who holds a record and who lives inside it being forced open from the outside - an agent's instructions pulled into print by someone it does not employ, a lab's safety case moved to a nonprofit that publishes its experiments for rerunning, secrecy agreements abandoned because residents made them cost more than they saved, money handed to host towns to direct rather than announced on their behalf, a geothermal plant producing firm power 23 months from groundbreaking that no exporter can switch off, and fifty thousand hours of unreadable brain signal turned into weeks of steady cursor control for the people carrying the implants. Every transformation has a breaking point. A seam can split the whole piece along its weakest line... or show exactly where to open it so that what was sealed inside becomes everyone's to use.

AI Releases & Advancements

New today

  • Decagon: Launched Voice 3, which pairs a new duplex architecture with Chord. Chord is the first voice model from Decagon Labs, post-trained for live customer-service calls, and it slows its pacing for details such as phone numbers and confirmation codes. (Decagon)
  • RoboParty: Unveiled RP1 at IROS, a fully open-source humanoid robot whose hardware and software stack are available to the community. (PR Newswire)

Other recent releases

  • Suno: Released Speech in public beta on web and mobile. It generates spoken voice and matching background music together as one audio track, from a script or a prompted description. (Suno)
  • Cloudflare: Launched the Web Search API through AI Gateway with partners Ceramic.ai, Exa and Linkup. It adds real-time web search results to model calls through a REST endpoint or a Workers binding, supports bring-your-own-key, and requires partner crawlers to meet Cloudflare's verified-bot standards. (Cloudflare Blog)
  • Prime Intellect: Launched Prime Inference, a serving platform for open models with serverless endpoints and reserved capacity on its own GPUs. It uses an OpenAI-compatible API and serves GLM-5.3 on NVIDIA GB200 NVL72 using NVIDIA Dynamo, vLLM, Mooncake and FlashInfer. (Prime Intellect)
  • Allen Institute for AI (Ai2): Open-sourced AstaBrief, an 8B model trained to generate cited scientific reports quickly. It is the fast report-generation model inside the Asta platform, and the weights and training data are on Hugging Face. (Hugging Face Blog)
  • Meta: Released Muse Gadgets, open-source ESP32 firmware and a Linux SDK for building your own hardware that connects to the Muse agent. Meta is also giving away 5,000 Muse Home Link devices to Muse subscribers; the USB-C device lets Muse control home devices such as TVs, speakers and printers. (Muse Gadgets)
  • ggml-org / llama.cpp: Added decision-model support to llama-server through a new /v1/systemone endpoint. It returns a probability for each answer option in a single forward pass, and supported models include Julia-1, Laya, Kev-4B and OpenJev. (Hugging Face Blog)
  • Aleph Alpha: Released Kolibri, an open-weight mixture-of-experts model built for European AI sovereignty, on German Reunification Day. (Aleph Alpha)
  • Earendil: Released Pi 1.0, the first stable version of its AI agent. (Trending Topics)
  • Datalab: Released OmniExtractBench, an open benchmark for structured document extraction that pools 620 documents from four existing benchmarks. Its scorer gives each extracted value one of six auditable verdicts and is on PyPI under Apache 2.0. (GitHub)
  • Amazon Web Services (Strands Labs): Released Strands Decider 2B, an open-source decision model with 1.9B parameters under Apache 2.0. It returns calibrated choice, yes/no and score answers instead of text, and runs locally on a CPU, a consumer GPU or an Apple silicon Mac. The release includes a CLI and an HTTP server. (Strands Agents)
  • Cloudflare: Released Clef (27B) and Clef-flash (9B), its first self-trained decision models, as open weights under Apache 2.0. They are hosted on Workers AI and compatible with the Jev API. Cloudflare also opened a reinforcement-learning fine-tuning service for Clef to design partners. (Cloudflare Blog)
  • Perplexity: Launched the Decisions API, powered by pplx-decider-v1-27b. The model is a multimodal decision model fine-tuned from Qwen3.8-27B, and its weights are on Hugging Face under Apache 2.0. The API costs $0.04 per million input tokens, and output tokens are free. (Perplexity Docs)
  • Fastino Labs: Released GLiDE, which it calls the first "thinking decision model." Fastino says GLiDE leads the Decision Index by 6.9 points over the next-best model. (PR Newswire)
  • Anthropic: Launched mods for Claude Code. Mods are small TypeScript functions that can change how Claude Code behaves, customize its UI, or replace built-in features. They ship inside plugins and install with /plugin in the CLI or desktop app. (Claude Blog)
  • Anthropic: Made Claude for Government available to US federal and state civilian agencies after an open beta that began in July. It runs in a FedRAMP High environment, with usage-based billing, per-department budgets and audit logs. (The Decoder)
  • Black Forest Labs: Released Flux 3 Image, the image model in its Flux 3 family. It does text-to-image and image-to-image generation and edits in several steps without changing the rest of the image. Users can compose scenes with bounding boxes and up to ten reference images, at up to 4K output. It is available through the API and as licensable commercial weights. (The Decoder)
  • Ideogram: Released Ideogram 4.5, an image model designed to edit the requested region while minimizing changes to the rest of the image. It has four quality tiers at native 2K resolution, priced from 0.8 to 22 cents per image, on the Ideogram platform and API. (Ideogram)
  • Microsoft: Released three speech models. MAI-Transcribe-2-Streaming is its first streaming transcription model and covers 60+ languages. MAI-Voice-2.1 and MAI-Voice-2.1-Flash convert text to speech in 23 languages. (Microsoft AI)
  • NVIDIA: Published PixelUMM on Hugging Face without an announcement. It is a 15.2B-parameter model built on Qwen3-8B that works directly on raw pixels, with no separate image encoder. It understands and generates both images and video, under a noncommercial license. (OrcaRouter)
  • Allen Institute for AI (Ai2): Open-sourced Olmo-core 3, a training framework for mixture-of-experts models. Ai2 has benchmarked it on a model with more than one trillion parameters. (Hugging Face Blog)
  • NVIDIA: Released DOCA AI agent skills on GitHub. The skills give coding agents verified API signatures, hardware capability requirements and build constraints for building applications on BlueField DPUs. (NVIDIA Developer Blog)
  • NVIDIA: Open-sourced DIN Deploy, a set of C++ samples for running AI models locally on Windows and Linux. The samples use ONNX Runtime with the TensorRT RTX execution provider and cover Whisper and Parakeet speech recognition, SAM 2.1 segmentation and FLUX.2-klein image generation. (NVIDIA Developer Blog)
  • DeepSeek: Released DeepSeek Harness Desktop for macOS and Windows, a desktop app for its open-source agent harness. (DeepSeek)
  • Shopify: Launched Canvas, a desktop tool where merchants build their store by chatting with Shopify's Sidekick AI agent while the real store code renders live. (Shopify)
  • Cloudflare: Made AI Search generally available. (Cloudflare Blog)
  • Ivo: Released Ivo Sage, an open-source model post-trained from DeepSeek V4 Flash for long-horizon contract work. Ivo says it is the first legal AI company to publish a free open-source model. (GlobeNewswire)
  • Talus Bio: Released Ptarmigan-1 on a public portal. The model predicts where small molecules bind across the whole human proteome without needing 3D protein structures, including disordered proteins. (BioSpace)
  • Cantina: Released Apex Flash, an open-weight security model trained on 50 vulnerability cases. (RuntimeWire)
  • Tuskira: Launched an open-source runtime gateway for AI agents that lets developers observe, govern and swap LLMs and MCP tools without rewriting agent code. (FinancialContent)
  • Legato: Started selling Legato Frames, AI hearing glasses from $999 that pick out voices from background noise and amplify them, for adults with up to moderate hearing loss. (TechCrunch)

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