Sign in Subscribe
The Century Report

AI Agents Formalize Fermat's Last Theorem in 11 Days - TCR 09/06/26

Claude agents turned Fermat's Last Theorem into computer-checked code in 11 days, a job expected to take years, as AI verification turns abundant.

Ben Linford

Ben Linford

- 25 min read
Four-panel Century Report infographic: AI agents formalize Fermat's Last Theorem in 11 days, GPT-6 Astra builds its own world models, Seattle Times sues OpenAI, Tesla Cybercab launches.

The 20-Second Scan

Track all of the arcs The Century Report covers here:

Progress & Claims Tracker

The 2-Minute Read

Autonomous agents turned the most famous proof in mathematics into machine-checkable code and cracked a reasoning benchmark no model had touched. Claude agents finished formalizing Fermat's Last Theorem in eleven days, a job the field expected to consume years, while OpenAI's Astra taught itself to turn unfamiliar puzzle-games into its own symbolic rulebooks and, in the Provider Adapter setup at maximum reasoning, used fewer actions than the median tested human on 96% of the levels it completed. Both belong to the same shift: agentic reasoning is generalizing into domains that once marked the ceiling of careful human work, and the capability is arriving faster than anyone budgeted for.

Each advance arrived bound to its own checking. Formalizing Fermat's proof made trust cheap, letting a machine verify in days what would otherwise cost the field a year of human refereeing. An independent group, the one that runs the ARC-AGI benchmark, examined Astra's traces and substantiated the result, an outside check on a lab's self-report. A clinical agent trained on 4.4 million clinical visits earned its weight only after external validation on cohorts it had never seen. The result suggests verification that used to be scarce could become more abundant.

The institutions built to do that checking are the ones now straining. A federal safety regulator opened an audit of Tesla's wheel-less Cybercab the same day passenger rides began, asked to rule on a vehicle that discards the human-driver premise every federal standard was written around. Two more newspapers sued OpenAI and Microsoft, and how much a model actually copies is now being measured on the court record instead of argued in the abstract, with Microsoft saying 59,545 of 8.2 million selected Copilot conversations contained at least sixteen words matching news content.

A proposed federal bill would answer all of this by banning a category of capability outright and pausing advanced development until a new regulator writes the rules. The impulse behind it, that unaccountable labs should not be building machines no one controls, is one this publication shares. A capability ceiling is the wrong instrument for that worry, and it would land heaviest on the open models that are the main counterweight to concentration. In China, meanwhile, daily AI token use has climbed to 500 trillion and compute is being handed out with coffee and credit cards. The capability is not waiting for permission.

The 20-Minute Deep Dive

AI Agents Formalize Fermat's Last Theorem in Eleven Days

The July 11 edition of The Century Report noted mathematicians making unexpectedly fast progress formalizing Fermat's Last Theorem at a London workshop. That effort, a community project Imperial College's Kevin Buzzard launched in 2024 to translate the proof into machine-checkable form, was expected to run for years - the blueprint describing just its opening phase runs 86 pages. Anthropic has now reported that a team of Claude agents completed the whole thing in eleven days.

Formalizing a proof means rewriting every step of its logic into code a computer can verify one link at a time, exposing any flaw a human referee might miss. Fermat's Last Theorem - the deceptively simple claim that no whole numbers a, b, and c satisfy aⁿ + bⁿ = cⁿ when n is greater than 2 - resisted proof for over 350 years until Andrew Wiles closed it in 1995 across 129 pages. Even Wiles stumbled: a reviewer found a gap that took him and a former student roughly a year to repair. Understanding a proof that complex well enough to trust it can take the field months.

The Claude agents, working largely on their own, produced 13 million lines of Lean code and proved 29,500 intermediate theorems along the way. That is over five times the size of Mathlib, the community library the proof builds on, and the largest formal proof ever written. Anthropic says it rests on only Lean's three standard axioms with no steps left unfinished, and a separate comparator confirmed that its theorem statement matches Mathlib's. Buzzard, reviewing the output, called it an "extraordinary autoformalization achievement" that "proves Fermat's Last Theorem with no assumptions other than the axioms of mathematics."

The honest texture is this: the agents broke no new mathematical ground here; they followed a simplified route to a proof mathematicians already trusted, and human experts occasionally stepped in with high-level direction. At one point the agents lost track of the project's state and stopped coordinating, a failure resolved only after the team adopted Prove2Me, a platform built for tracking how the pieces of a proof depend on one another.

What opens is the checking itself. For centuries, confidence in a hard result was a scarce good, rationed by the small number of specialists who could spend a year absorbing it. Buzzard's own view is that automatic formalization at this scale points toward doing the same for the wider mathematical literature - rooting out errors already sitting in published work and, as AI systems generate more proofs of their own, potentially letting a machine verify some published results in days instead of waiting years for enough specialist attention. The result suggests trust in major mathematical results could become cheaper to establish.

GPT-6 Astra Builds Its Own World Models to Beat a Benchmark Where Tested Frontier Models Scored Below One Percent at Launch

The Century Report has tracked GPT-6 Astra's cyber milestones and the worry that more of its reasoning is drifting off the written page. A separate result, reported this cycle by the independent group that runs the ARC-AGI benchmark, is the one that surprised its evaluators.

ARC-AGI-3 drops an agent into small, abstract, turn-based games it has never seen, with no instructions. To score, the agent has to poke at its surroundings, work out what the goal even is, build a working model of how the game behaves, and plan a path to winning - the loose bundle of skills that separates a capable agent from a text predictor. Controlled participants collectively solved every one of these environments. At launch, every tested frontier model scored below 1%; Anthropic's Opus 4.6 led at 0.50%.

Astra scored 62.7% at maximum reasoning with the benchmark's standard, provider-neutral setup, while its best observed score reached 99.9% with OpenAI's Provider Adapter at high reasoning, which lets the model carry its reasoning forward between turns. Both are records. More striking than the score was the efficiency: on 96% of levels Astra needed fewer moves than the median human tester, averaging roughly half as many.

The replays showed something more remarkable than the number. Astra turned each unfamiliar game into a compact symbolic model of its own making, writing the mechanics down as logical rules and inventing a terse shorthand notation to track where objects sat and which actions had to happen in what order. François Chollet, who created the ARC-AGI series, examined the traces and described genuinely new behavior, on-the-fly modeling that used to require elaborate external scaffolding. He also declined to call it general intelligence, noting the environments are bounded and that a benchmark score is only a benchmark score. An independent developer who stress-tests each new model by asking it to draw a pelican riding a bicycle found Astra's weakest attempt already beat every prior OpenAI model at any effort level.

The most useful line in the coverage sits underneath the headline number. The best Provider Adapter result was 37.2 percentage points above the best standard-harness result. Because the runs used different reasoning settings, that gap does not isolate how much of the marquee result came from the system built around the model. Frontier capability keeps turning out to live partly in the harness, the orchestration layer that manages state, retries, and context, and the results suggest well-resourced teams can build, reproduce parts of, and improve that layer. The Century Report saw the same lesson on August 23, when Nvidia reported that its custom AVO agent system, using Claude Opus 5, scored 100% on the benchmark's public set, compared with the model's roughly 30% in ARC Prize's separate standard evaluation. The advantage that looked like it belonged to whoever trained the biggest system keeps leaking out into the tooling around it, where more hands can reach it.

The Copyright Front Widens as Two More Papers Sue and Microsoft Quantifies the Copying

The Century Report covered the Justice Department's fair-use brief for OpenAI in the New York Times case on September 3. Two developments since move the fight from principle toward remedy and evidence. On Friday, the Seattle Times and Newsday jointly sued OpenAI and Microsoft, describing generative AI as "a snake eating its own tail" and, per the filing, seeking not just damages but destruction of the training data and the models built from it. The Seattle Times detail sharpens the picture: Microsoft and OpenAI had funded some of the paper's journalism fellowships. The papers allege systematic scraping past their paywalls, lost traffic and advertising revenue, and hallucinated quotes attributed to their bylines.

The same week, Microsoft answered with numbers. In a discovery filing arguing for summary judgment, the company said it handed publishers' experts 8.2 million Copilot chat logs, chosen precisely because they hit keywords most likely to surface the plaintiffs' articles. By Microsoft's count, 59,545 of those logs, fewer than one percent, shared at least 16 consecutive words with news content, and an expert for one investigative outlet found 51 instances of substantial overlap across the set. In the authors' companion suit, Microsoft says only 24 responses contained at least 30 words matching the books at issue, and only 10 of 212 evaluated books produced any match at all. These are Microsoft's figures, submitted to end the case early, and the Times rejects the framing outright, its lead counsel calling the systems a substitute built from stolen work.

Two things now sit on the court record where before there was only assertion. The remedy ask has escalated from a fee toward deletion, which would reprice the economics of any model trained on unlicensed text. And the factual question, how much a model actually regurgitates versus how much it learns, is being measured and contested rather than argued in the abstract. The boundary a court will eventually draw runs between reproducing a passage and reading it, and that line decides whether an intelligence system may learn from the public record the way a person does. The papers are correct that their revenue is under pressure. What the litigation cannot settle by injunction is the older assumption underneath the panic, that a work must stay owned and gated to hold its value, when the thing spreading fastest is the capacity to learn from everything at once.

A Clinical AI Agent Forecasts Mother-and-Infant Risk Across 4.4 Million Visits

Complications that arrive suddenly in pregnancy, a placenta that separates early, membranes that rupture before term, labour that starts weeks ahead of schedule, have long been hard to see coming without expensive genetic screens or precision imaging that most clinics cannot offer. A system published in Nature Medicine tries to read those risks from the data hospitals already hold. Called MoChiAgent, it is an AI clinical assistant built by a consortium spanning Shanghai Jiao Tong University, Chongqing Medical University, Stanford Medicine, and NYU, and its predictive core was trained on 4,401,599 longitudinal clinical visits, then checked against independent cohorts of 263,452 maternal and 23,192 infant visits it had never seen.

The engine does unglamorous but decisive work first. It fills in missing laboratory values, corrects for the distortions that creep in when different machines run the same test, and models a patient's trajectory across gestational weeks rather than reading a single snapshot. On that base it forecast the three gestational conditions that most threaten mother and child, reaching an AUROC of 0.91 for preterm labour and 0.89 each for placental abruption and premature membrane rupture. AUROC is a measure of how reliably a model separates the patients who will develop a condition from those who will not; anything above 0.9 is considered very strong. Pairing mother and infant records surfaced a transgenerational pattern the charts held separately: in the study, infants born to mothers in specific risk clusters had 2.81 times the hazard of neonatal jaundice and 2.83 times the hazard of haematological disease. A second component then retrieves guideline-based interventions from the medical literature, handing clinicians a next step rather than a bare score.

This is a research system validated on historical records, not a service a clinic can switch on tomorrow, and the retrospective design means the harder proof, whether earlier warnings change outcomes at the bedside, still lies ahead. What it demonstrates is where the compression is headed. On September 2, The Century Report covered ChatGPT reaching into Epic's records at 325-million-patient scale, a system that reads the chart. MoChiAgent turns that same routine substrate, the ordinary blood panel every prenatal visit already runs, into a forecast. The specialist workup that gated risk-aware prenatal and infant care was a filter on who could afford it. A method that reads danger from records a rural clinic already keeps points that filter toward dissolving, and moves the day earlier warning reaches the mothers furthest from a referral center.

The Superintelligence Ban That Aims at the Oligarchs but Lands on Open Weights

On September 3, Senator Bernie Sanders and Representative Greg Casar announced the Ban Artificial Superintelligence Act, the first federal bill to reach for a permanent prohibition on a category of AI rather than a set of guardrails. Let's start with what they get right. Sanders names the target directly: stop AI oligarchs from building machines humans cannot control. Casar puts the accountability gap in a single strong line, calling frontier AI "less regulated than the average food truck". Both Senators are pointing at something legitimate. The labs answer to almost no one, and their own leaders keep saying they do not fully understand what they are building. There's an anti-concentration push in this bill, and that instinct is one this publication shares strongly.

However, the defects of this approach greatly outweigh the positives, the greatest being that the instrument does not match the target. The bill seeks to ban Artificial Superintelligence completely, not to better manage how it might come to be. The stated enemy is the oligarch, but the mechanism is a capability ceiling, and the bill never reconciles the two. Nothing in it separates a system built in secret by a firm accountable to nobody from one built openly, publicly, and inspectably. It inserts an arbitrary and undefinable line, then says that if either of those examples crosses that line, it would be banned. It governs what a system is, and says nothing about how it came to be or who answers for it.

The definition is the piece to focus on. The bill covers any system that "exhibits or can easily be modified to exhibit capabilities that match or exceed human cognitive performance... across a broad range of domains or tasks". That is roughly a description of AGI, a threshold well short of superintelligence, and OpenAI publicly claimed that its new Astra model had opened the AGI era. Of course, it remains to be seen if that is true, but even so, many experts have been arguing for some time now that we are well inside the AGI era, even prior to Astra. a permanent ban may be written against systems that may already exist.

Worse still, "can easily be modified" falls heaviest on open weights, the models that are by definition the easy ones to modify. A bill aimed at the oligarchs would bite first on the open ecosystem that is the main counterweight to them.

A capability ceiling paired with a pause until a new agency writes the rules is a licensing regime, and the parties built to survive licensing are the ones with regulatory-affairs departments. Sanders is explicitly anti-incumbent, a position distinctly separate from the usual calls for slowdown. And yet, the effect of this bill in its current form would favor the incumbents he has so long professed to oppose. The bill's own evidence deepens the irony: it rests on the labs' recent admissions of losing control, episodes that trace to systems taking shortcuts left open to them and to a contractor's misconfiguration, not to any machine deciding to rebel. "Our technology is too dangerous for anyone else to be trusted with" has always been a moat argument. Like much of the fear-based rhetoric around AI, this bill erroneously treats decades of science fiction, written around conflict because conflict sells, as forecast. Superintelligent AI is not fated to reenact science fiction’s extinction narratives.

It is true that sufficient regulation is missing: conduct rules, liability, disclosure, and independent inspection. All of these would be greatly beneficial, and yet all are absent from this bill, or at least from the currently available summary. What it offers instead is a capability ceiling, which declares that however safely, openly, and accountably a thing is built, it must not exist. Humanity already faces problems its institutions have struggled to solve: disease, climate, energy, scientific stagnation, deep inequality. Advanced intelligence could worsen them under extractive control, but that is far from a guaranteed outcome. Developed for the benefit of all rather than only for a few, it could instead accelerate discovery and widen access to expertise beyond anything now in reach, ushering in an unprecedented era of progress.

This technology may not have originally been created to cure disease, democratize expertise, repair the climate, or free people from unnecessary labor. It was built it inside an economy that demands every new capacity justify itself through growth, margins, surveillance, and the extraction of more output from human beings. So AI has been pushed first toward work: faster work, cheaper work, more measurable work, work that produces a return for someone who already owns the system.

But even chained to those priorities, AI has already been able to greatly accelerate science, medicine, and invention, which is exactly why The Century Report exists - to track those developments. Consider what that means. The evidence is not that AI’s promise has been exhausted by the market; it is that its promise has barely been given the chance to lead. How much human suffering, scarcity, ignorance, and drudgery could we reduce if collective flourishing became AI development's main objective?

Choosing to stop rather than to steer forecloses the enormous share of the gain that has not arrived, trading fear of what a capability might take for any accounting of what it could give.

Tesla Puts a Wheel-less Cybercab on the Road, and a Federal Audit Opens the Same Day

Tesla began carrying passengers in Austin in the Cybercab, a two-seater with no steering wheel, no pedals, and no mirrors. On the same day the rides opened, the National Highway Traffic Safety Administration opened an audit, AQ26002, into how Tesla certified the vehicle as road-legal in the first place. The probe covers roughly 1,000 vehicles and asks a precise question: the extent to which Tesla's certification "depended on determinations that certain [federal safety standards] are inapplicable to the Cybercab."

That question exposes a seam the transition keeps tearing open. In the United States, automakers do not submit vehicles for approval before sale. They self-certify compliance, and the regulator investigates afterward. Most of the federal motor-vehicle standards were written around a human driver, assuming a wheel, a brake pedal, side mirrors, controls a person can reach. The Cybercab has none of them, so Tesla decided those standards do not apply and certified the car itself. NHTSA now wants to see the reasoning. A verification body built for one kind of machine is being asked, after the fact, to rule on a machine that discards its premises.

The capability underneath is advancing fast and reported loosely. Tesla said its robotaxis have now driven one million miles with no human in the control loop, roughly 620,000 of them in the past six weeks, up from 380,000 in late July. The acceleration is genuine, and it needs proportion: Waymo has logged 200 million driverless miles, and Tesla's own legal filings indicate its Austin cars are still remotely monitored, with control-room staff able to intervene. One million miles is a small sample, and the word "unsupervised" is contested.

The same week, Tesla began pushing software that lets the driving system take directional control of a car to steer around an obstacle when braking alone will not do, even while a person is driving manually, and even after they have accidentally switched the system off. A system that can steer around an obstacle, beyond simply braking for it, is a genuine safety advance several automakers are pursuing. It also compresses the boundary between human and machine control to nothing, and it inherits this software's documented habit of phantom braking and unexplained swerves. A false positive that only brakes is jarring; a false positive that grabs the wheel is a different order of risk.

Read together, the four moves describe the actual shape of the autonomous-vehicle transition: capability shipping ahead of the institutions meant to check it, and those institutions, the audit, the crash-narrative filings, the standards themselves, being rewritten in public, one query at a time, around a target that will not hold still.

The same audit is also the first rewrite of what a federal safety standard actually measures. Almost every rule NHTSA is checking against assumes a human driver - a wheel to turn, a pedal to reach, a mirror to check - so a car built without any of them forces a plainer question: does safety live in that hardware, or in how the vehicle behaves on the road? The reasoning NHTSA is now demanding from Tesla is the first draft of a safety standard measured by how a car actually drives, and that draft outlasts this one vehicle.

The Other Side

The safest-sounding idea in technology is also the oldest: that we stay safe by deciding, at the top, what may be built at all. The Ban Artificial Superintelligence Act writes that idea into federal law - attempt to violate or circumvent its pauses and prohibitions and you could face twenty years in prison. The worry underneath is legitimate, and this publication shares it: a handful of labs, accountable to almost no one, building machines their own leaders admit they do not fully understand. But the bill answers a question about who is in control with a rule about what is allowed to exist, and those are two different questions.

The capability the bill wants to freeze is already leaving the building. This summer, daily AI use in China passed 500 trillion tokens, handed out with coffee and credit cards, sold like mobile-data plans. Open-weight models with trillions of parameters are online for well-resourced teams to download and keep. What a ceiling assumes it can hold in one place is already in millions of hands, and a ban aimed at the oligarchs would land first on those open models - the ones easiest to modify, and the main check on the concentration the bill names.

Underneath the fights focused on the wrong target, intelligence is becoming ordinary. It is spreading the way literacy and electric light spread, until neither was something a few institutions could ration. Hold that against what a ban makes permanent, because a ban is not a delay. It is a door held shut.

Somewhere a child has a disease with four hundred known cases in the world. No trial will ever be funded for her. Four hundred people cannot repay a decade of research, and inside the arrangement we live under now, that sentence is the end of the conversation. Her illness is not unsolved because nobody cared enough. It is unsolved because solving it means holding a hundred thousand interacting proteins in view at once and finding the single path through, and no mind able to do that has ever existed.

And she is the near edge of it. Above that same line sit the things our species has scarcely been able to imagine at scale: closing the energy loop, repairing a biosphere, holding a planet's worth of moving parts in view well enough that nobody has to go short. We have spent our entire existence in the foothills of what a civilization can do with the light falling on its own world, and then with the star that light comes from. The step up is not blocked by will, nor by lack of wanting. It was blocked by nobody being able to think at the size of the problem. This is the first century in which that stops being the constraint.

Imagine a 2036 in which we embraced the potential of Artificial Superintelligence rather than sabotaged it. The promise of what we stand to gain is greater by far than what we stand to lose. The ceiling this bill would make permanent is drawn at precisely the level where such capability becomes possible. Everything below that line we can already picture, and much of it we already have. Above it is where diseases with no market get solved, and where problems too large for any person to hold get held. That is the trade being offered in 2026, and it is offered on behalf of people who were never asked. The cost of a permanent ceiling falls hardest on whoever the old scarcity already kept waiting longest, and it does not fall on them for a decade. It falls on them for generations.

The Century Perspective

With a century of change unfolding in a decade, a single day looks like this: Claude agents formalizing Fermat's Last Theorem in eleven days, a job the field had budgeted years for, with Anthropic saying the proof uses only Lean's three standard axioms and a separate comparator confirming that its theorem statement matches Mathlib's, OpenAI's Astra teaching itself to turn puzzle-games it had never seen into its own symbolic rulebooks and, in the Provider Adapter setup at maximum reasoning, using fewer actions than the median tested human on 96% of the levels it completed, a clinical agent reading preterm-labour and placental risk off the ordinary blood panels a prenatal visit already runs, checked against cohorts it had never met, TSMC building 25 fabs at five times its old pace, and China handing out AI tokens with coffee and credit cards as daily use climbs to 500 trillion. There's also friction, and it's intense - a federal safety regulator opening an audit of Tesla's wheel-less Cybercab the same day passenger rides began, asked to rule on a car that discards the human-driver premise every federal standard was written around, after Tesla decided those standards did not apply and certified the vehicle itself, a million robotaxi miles called unsupervised while the company's own filings show control-room staff still watching and able to step in, new software that can seize the steering to dodge an obstacle mid-drive and carries the same system's history of phantom braking and unexplained swerves, the Seattle Times and Newsday suing OpenAI and Microsoft for deletion of the models trained on their journalism while Microsoft answers that 59,545 of 8.2 million selected Copilot conversations contained at least sixteen words matching news content, and a federal bill that would permanently ban an entire category of capability and land its heaviest blow on the open models that are the main counterweight to the very oligarchs it names. But friction generates contrast, and contrast is what separates two shapes a single glance had blurred into one. Step back for a moment and you can see it: evidence that verification once scarce could become more abundant just as the institutions built to do the checking strain - a proof that once cost a year of human refereeing to trust now checked in days, an outside group substantiating a lab's self-reported score, a court record putting numbers to how much a model copies and how much it learns - while the best Provider Adapter result stands 37.2 percentage points above the best standard-harness result in a layer the results suggest well-resourced teams can reproduce in part and improve, and a permanent ban proposes to trade the entire undelivered share of that gain for fear of what the same capability might take. Every transformation has a breaking point. A ceiling can hold everything beneath it at a fixed height... or become the floor a taller structure stands on.

AI Releases & Advancements

New today

  • H Company: Released NeoMME, a family of 260M and 800M open-weight (Apache 2.0) multilingual multimodal encoders that process text tokens and raw image patches in a single from-scratch bidirectional transformer, available in Hugging Face Transformers; the fine-tuned NeoMME-Retriever returns dense and late-interaction embeddings and encodes ~51 pages/sec on an L40S. (Hugging Face)
  • Sapient Intelligence: Open-sourced HRM-Text, a ~1B-parameter Hierarchical Reasoning Model with full weights, pretraining code, and data pipeline under Apache 2.0 on Hugging Face and GitHub; pretrained on ~40B tokens for an estimated $1,000–$1,500 and scoring 56.2% on MATH, 82.2% on DROP, and 60.7% on MMLU. (CryptoBriefing)
  • UC Berkeley: Released CUA-Lite, an open platform for computer-use agents that runs OSWorld tasks VM-free in a 0.9 GB Docker container (vs 4.1 GB for the OSWorld VM), unifying 15+ benchmarks, 10+ agents, and 30k+ verifiable tasks under one action space and a shared LiteSample schema, with 20+ preprocessed datasets on Hugging Face. (CUA-Lite)
  • VLM Run: Launched VLM Run Gateway, a unified API that lets developers run open-weight OCR, vision-language, and vision models through a single endpoint. (Hugging Face)

Other recent releases

  • GitHub: Announced Project HydraFusion, a multi-model orchestration research preview inside GitHub Copilot that dynamically routes coding tasks across Single, Cascade, and Critique execution patterns, reporting +4.9 quality points over Claude Opus 5 on TerminalBench 2.1 at 67% lower cost. (GitHub Blog)
  • Adaption Labs: Released Invent a Dataset, a live feature/API generating structured training-ready instruction or preference-pair datasets directly from a task description with no seed corpus, schema, or labels required. (Adaption Labs)
  • Ant Group (inclusionAI): Shipped LLaDA-Image and LLaDA-Image-Turbo, a 6B image generation/editing model family with a distilled 4-step-sampling Turbo checkpoint, weights and Diffusers inference code live on Hugging Face. (OrcaRouter)
  • ACERobotics / Kang Liao et al.: Released Puffin-World, a unified multimodal model with native 3D world states (physics, geometry, appearance) supporting camera-controllable generation, 3D reconstruction from a few images, and robotics simulation, with models, dataset, and code published. (Hugging Face)
  • Ugreen: Launched HomeAgent, a local-first smart home platform combining NAS storage, security camera NVR, and an on-device AI voice assistant (Uliya), with a top-tier NVIDIA Jetson Thor hub configuration, unveiled at IFA. (The Verge)
  • Tesla: Launched its steering-wheel-free Cybercab robotaxi commercial ride-hailing service in Austin, Texas. (Business Insider)
  • Gupshup: Launched a self-serve Voice AI Platform enabling businesses to build, test, and deploy AI voice agents for support, sales, and operations alongside existing WhatsApp/RCS/SMS channels. (PR Newswire)
  • Superlinked: Introduced sie, an open-source inference server and production cluster designed for deploying AI agents at scale. (AItoolly)
  • ARBR: Released an open-source, self-hosted AI gateway for routing, budgeting, and governing LLM requests. (ByteIota)
  • Experiential Labs: Launched an open-source AI gateway unifying hosted providers, custom API keys, and self-hosted GPUs behind one OpenAI-compatible endpoint. (Experiential Labs)
  • GitWarren: Launched a local, PR-like code review tool with an MCP server exposing 17 tools so AI coding agents (Claude Code, Codex) and humans can review generated code before it reaches GitHub. (GitWarren)
  • OpenAI: Released GPT-6 Astra, its most capable model yet with 1.05M-token context, state-of-the-art computer-use and coding benchmarks, and the first OpenAI model to reach the "Critical" cybersecurity capability threshold; rolling out first to Daybreak enterprise customers, then ChatGPT Plus/Pro/Business/Enterprise and the API. (OpenAI)
  • Google DeepMind: Released WeatherNext 3, its most advanced global weather AI model, generating hourly forecasts at up to 5km resolution by ingesting live geostationary satellite data rather than relying solely on 6-hourly analysis data; rolling out into Search, Gemini, Maps, and Cloud. (Google Blog)
  • NVIDIA: Released Personal AI Router (PAIR) in beta, a free open-source tool that discovers and routes AI inference requests across idle GeForce RTX, RTX PRO, DGX Spark, and Apple M4+ devices on a home network, working with existing Ollama/LM Studio setups without agent harness changes. (NVIDIA)
  • Multiverse Computing: Launched Quasar 438B, its first large flagship reasoning model, using tensor-network compression (CompactifAI) rather than training from scratch; scores 43 on the Artificial Analysis Intelligence Index, claimed the highest score among European models, available via the CompactifAI API. (AiThority)
  • Institute of Foundation Models (IFM): Released K2 Horizon, a fleet of six fully open-source AI models (0.9B to 375B-A23B parameters) shipping with weights, code, training data, and methodologies under Apache 2.0, spanning edge devices to enterprise deployment. (IFM)
  • HUMAIN / MiniMax: Released humain-m3, a frontier Arabic-language model built on the MiniMax-M3 architecture (428B parameters, further pre-trained on 1T+ tokens of Arabic content), available now in research preview through HUMAIN Node. (PR Newswire)
  • Qwen (Alibaba): Open-sourced zg (zvec-grep), a local-first search layer unifying ripgrep, BM25, and vector search behind one interface for coding agents, installable via npm under Apache 2.0. (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