The Century Report

A Swarm of 10,000 AI Agents Solved Navier-Stokes - TCR 09/09/26

Roughly 10,000 AI agents resolved the 90-year Navier-Stokes problem with a proof a computer can check, as DeepMind gave away a genome atlas.

Ben Linford

09 Sep 2026 - 24 min read

The 20-Second Scan

OpenAI said roughly 10,000 agents on an unreleased model produced a Lean-verified proof resolving the Navier-Stokes Millennium Problem, as an NYU mathematician alleged it used his uncredited work as a starting point.
Google DeepMind released a precomputed map of roughly nine billion possible single-letter substitutions across the reference human genome, free for non-commercial research, turning rare-disease variant analysis from a coding job into a lookup.
The Energy Department closed a $1.9 billion loan to restart Iowa's shuttered Duane Arnold reactor, the third revived for AI compute, as Google signed its first non-US nuclear deal in a €13 billion Finland buildout.
A study found twelve AI developers left 67% of safety-commitment changes undisclosed as a left-right coalition pushed the White House to publish its AI-review rules and two papers showed auditors relay reports and benchmarks misfire.
A UK analysis pegged planned data-center jobs at 10,400 against the industry's 40,000, as Ohio bakes tripling demand forecasts into bills, builders shed community accountability, and a Louisiana town credits Meta's build with its revival.
A hedge-fund manager runs his firm entirely on AI agents at 1% of its old payroll, as US data shows no AI job losses, a patent-drafting trial finds quality gains, and AI deployment units multiply.
Meta released Muse, a personal AI agent across iOS, Android, WhatsApp, and the web that automates tasks in an isolated cloud environment, positioned against OpenClaw and Instinct with data privacy as its explicit pitch.
IonQ published an end-to-end estimate that a 20,000-qubit fault-tolerant machine could break Bitcoin's secp256k1 signatures in about 26 days, while noting post-quantum standards ML-DSA and SLH-DSA are unaffected.

Track all of the arcs The Century Report covers here:

Progress & Claims Tracker

The 2-Minute Read

In a single news cycle, machines closed work at a scale that would have looked like fantasy a decade ago. OpenAI says roughly ten thousand autonomous agents resolved the Navier-Stokes problem, a question of fluid motion open for about 90 years, and returned a proof a computer can check line by line. Google DeepMind precomputed molecular effects for roughly nine billion possible single-letter substitutions across the reference human genome and handed the map to any lab that wants it. A single trader rebuilt his firm to run on agents for one percent of its former payroll. The capability is settled. What stays open in every case is the arrangement built around it.

Two ways of trusting a result ran side by side. The Lean-verified proof carries its own certainty, asking nothing of the author, the institution, or the timeline, and neither does a genome atlas anyone can look up on a web page. Against that sits accountability that only relays an actor's word. A hash-pinned study found twelve frontier developers left 67% of their safety-commitment changes undisclosed, most of them weakening what had been promised. A companion paper showed AI auditors passing along filed reports they never checked. When the record is self-reported, the checking layer just forwards it.

The other thread is distribution. When agents produce the discovery, the genome map, or the trade, the unresolved thing is who captures the value and who carries the cost. An NYU mathematician alleges his uncredited work seeded the Navier-Stokes result and that he was pressured to erase a collaborator's name. Public money reopened an Iowa reactor for one hyperscaler's compute, and the department would not say who shields households from the bill. UK data centers came in at a quarter of their promised jobs while Ohio prices speculative demand into everyone's power costs.

Nothing in the technology decides whether that surplus pools at the top or reaches the people whose labor it displaced. The near-term shape looks like redistribution: aggregate US employment shows no AI-driven cliff, and a patent-drafting trial found AI assistance raised quality without eroding expertise. The proof is checkable, the genome map is free, and a left-right coalition is forcing the review rules into daylight. The terms are still being written, and the past few days showed who is writing them.

The 20-Minute Deep Dive

OpenAI Says Its AI Agents Resolved a Navier-Stokes Millennium Problem, and a Fight Over Credit Opens the Same Day

On Tuesday, September 8, OpenAI announced that roughly 10,000 autonomous agents, running on an unreleased internal model it says is significantly more capable than the GPT-6 Astra it shipped a week earlier, produced both an analytical proof and a machine-checkable Lean formalization showing that a smooth three-dimensional fluid initially at rest can, under smooth forcing, develop a singularity - a point where speeds climb without bound - in finite time. This extends the autonomous formal-mathematics advance covered in the September 6 edition of The Century Report, when Claude agents completed the first computer-checked formalization of Fermat's Last Theorem in 11 days. That resolves the Navier-Stokes existence and smoothness problem, a Clay Mathematics Institute Millennium Prize open for about 90 years. OpenAI says the run consumed roughly 130 billion output tokens across 2.7 million messages between agents, and it will not claim the $1 million prize.

Twelve hours earlier, Tristan Buckmaster of NYU had announced that he and Levent Alpöge, a mathematician at Anthropic, resolved several closely related problems after nearly a year working with public models including OpenAI's Codex and Anthropic's Claude. Both efforts built on an approach opened by Diego Córdoba and Luis Martínez-Zoroa, whom the mathematician who wrote the Clay problem description called the real heroes of the story.

Then the dispute. Buckmaster alleges OpenAI learned of his and Alpöge's progress and used their approach as a starting point without credit; he says Sébastien Bubeck of OpenAI asked him to drop Alpöge's name because Alpöge works for a competitor, and that when he pushed to go public he was asked, "Why would you ruin your career?" He asked whether the agents had read transcripts of his own sessions with OpenAI's models and was told no; he asked whether the models had been trained on those transcripts and received no answer. Mark Chen, OpenAI's chief research officer, denied that any agent or employee accessed the pair's work, and OpenAI says only that de-identified usage data may have helped improve its models. Each contested point is an allegation met by a denial, and the training question sits unanswered.

Who arrived first is the smallest thing this raises. One side brought ten thousand agents and 130 billion tokens; the other was two people. When arriving first turns on how much compute a party can spend, a credit system built to reward it is no longer measuring insight, but simply confirming who had the larger budget.

Consider what credit is for. It is the ledger an extractive economy uses to decide who gets funded, hired, and believed, the mechanism that turns a discovery into a living. The injury Buckmaster describes runs deeper than a missing citation: a collaborator nearly struck from his own result for working at the wrong company, a researcher told his career could be ruined. What was threatened was standing, the ability to keep working and be believed by peers. A person's capacity to keep working should not hang on outrunning ten thousand agents, and the harder question is why, in this arrangement, trying to do so is increasingly tied to survival.

The problem itself has fallen, and in a form anyone with a computer can check. The Lean formalization can be machine-checked against its stated premises; it asks no trust in the author, the institution, or the timeline. Knowledge that verifies itself does not need its discoverer attached to be used, which drains the weight from the whole apparatus of claim and counterclaim. The mathematics is now available to every researcher who wants it, and the next result of this kind will not need anyone's permission to be confirmed. Those 130 billion tokens were not free, though; OpenAI used what New Scientist estimated was about $15 million worth of AI effort, and an argument for discovery held in common has to sit with who funds work at that scale in today's economic system.

DeepMind Maps Molecular Effects of Single-Letter Substitutions Across the Reference Genome, and Gives It Away

The human genome runs to roughly three billion letters, and only about two percent of them spell out proteins directly. The rest governs when, where, and how much each gene turns on, in a grammar that has resisted reading since the sequence was first assembled. On Tuesday, Google DeepMind released the AlphaGenome Atlas, precomputed molecular-effect predictions for roughly nine billion possible single-letter substitutions across the reference human genome.

The prediction model itself, AlphaGenome, shipped in 2025, and DeepMind says about 9,000 researchers have queried it through a programming interface since. Doing so meant writing code and querying a hosted model that could be slow, a wall that kept many biologists out. The Atlas takes that wall down: DeepMind ran roughly nine billion substitution predictions across the reference human genome ahead of time, a one-petabyte result the team could produce only after speeding its own pipeline up roughly eightyfold through model distillation and other engineering. A biologist can now look up a variant on a web page and read a single number - the AlphaGenome Variant Impact score - that estimates whether the change is worth a closer look, alongside more than a hundred finer predictions.

What that opens is concrete. The score distinguished disease-causing mutations from benign ones in the researchers' evaluation in a clinical database, and it helped a team at the Broad Institute flag a non-coding variant as a likely cause of a severe epilepsy case, the kind of DNA stretch older and cheaper tools tend to skip. Applied across the whole genome, the Atlas also began assembling something researchers have long lacked - a searchable dictionary of the short regulatory motifs that switch genes on and off.

The researchers who reviewed it name the limits without hedging. It predicts; it does not replace the experiment that confirms, and it cannot account for the particulars of an individual patient. Regulatory elements that act from far away can fall outside the million-letter window the model reads. One genomicist called it a useful resource that is also easy to misinterpret. Held to those bounds, it hands a small lab the reach that used to require a big one's compute budget.

The Atlas is free for non-commercial research, following the AlphaFold protein database that DeepMind opened to millions of users. Google's power-sourcing campaign had already reached record scale when the September 3 edition of The Century Report covered its 396-megawatt enhanced-geothermal agreement, expandable to nearly one gigawatt by 2030. That generosity sits beside the same parent company's appetite for power: Alphabet's compute demand is helping reopen a retired Iowa reactor on up to $1.9 billion in federal loans and underwriting new nuclear supply abroad. One firm, handing a genome map to the commons with one hand while bidding up the grid with the other.

A Third Shuttered Reactor Restarts on Public Money, With Google as the Buyer

The Energy Department closed a loan of up to $1.9 billion on Tuesday to restart the Duane Arnold Energy Center in Iowa, a 615-MW plant that shut down in 2020 after a windstorm and a hard economic verdict that repairs were not worth it. It is the third mothballed US reactor moving back toward operation under the same federal financing program, following Holtec's Palisades in Michigan and Constellation's Crane, the former Three Mile Island Unit 1, in Pennsylvania. Reviving a retired commercial reactor has never been done in the United States; Palisades expects to be first, possibly this year. NextEra targets a 2029 return for Duane Arnold, pending licensing.

The plant's output is already spoken for. Under a 25-year agreement signed last fall, Google will be the primary customer, using the electricity to run its cloud and AI infrastructure in Iowa. This extends Google's move into firm clean power that the September 3 edition of The Century Report documented through its record 396-MW geothermal agreement with Fervo Energy. In the same news cycle Google signed its first nuclear deal outside the US, a 22-year contract with Fortum to keep Finland's Loviisa plant running to 2050, part of a €13 billion Finnish buildout. The same company's DeepMind on Tuesday released a genome-wide atlas predicting molecular effects for roughly nine billion possible single-letter substitutions across the reference human genome, free for non-commercial research, a genuine contribution to the commons.

Reviving carbon-free baseload is a gain the grid can use. The friction sits in who arranged it and who pays. Public money is reopening retired generation specifically to meet one hyperscaler's demand, and when asked how the loan squares with a federal ratepayer-protection pledge that households should not subsidize data-center infrastructure, the department did not answer. NextEra's chief executive framed the restart as a way to "help keep power affordable for existing customers" and to ensure "Iowa families and businesses are not asked to bear the costs of growth," a claim the unanswered ratepayer question leaves hanging.

Read forward, the plant itself is capacity the grid can put to work: what NextEra says is roughly enough power for 500,000 homes, about 450 permanent jobs, no combustion. What stays open is whether the buildout that revived it treats firm clean power as a shared resource or as private supply financed on the public's credit and billed, without much notice, back to the people who never signed the contract. The generation is coming back. The terms are the thing still being written.

Safety Commitments Drift in the Dark, and the Layer Meant to Catch It Passes Reports Along Instead

Several major frontier AI developers publish documents describing how they will decide whether a model is too dangerous to release, and EU and California rules now give some disclosures and commitments in those documents legal force for covered companies. As the September 5 edition of The Century Report covered, California had just enacted independent third-party model verification, moving the state beyond industry-authored review. A new study assembled every public version of twelve developers' frameworks, pinned each to a cryptographic hash so the record cannot be quietly edited, and traced 710 individual commitments across consecutive revisions. The finding: 67 percent of material changes to what these companies promised went unmentioned in the companies' own accounts of what they had changed, and 77 percent of the traced changes weakened or removed a commitment rather than adding one.

The disclosure that does happen is uneven. One developer, xAI, published no account of any revision at all; vague narrative announcements hid more than itemized changelogs; and only one company enumerates its changes in a way an outsider could actually audit, and even that incompletely. Meta sits among the developers whose commitments drifted without a legible trail - the same company now credited with reviving a Louisiana town's main street around its data-center build, one actor producing sharply different effects in different places.

The remedy the law already specifies aims at the wrong thing. California requires a justification for why a framework changed; what makes a revision checkable is an enumeration of what changed, which is a different object. That gap is why more than two dozen groups spanning the ideological spectrum - Americans for Prosperity and the R Street Institute alongside Free Press and Public Citizen - wrote to the White House on Tuesday demanding it publish the voluntary framework it uses to review powerful models before release, a framework only a handful of insiders have seen. A separate nonprofit has sued under the Freedom of Information Act to force it into the open.

Two other papers landed in the same window showing how the disclosure gap plays out at the mechanical level. One tested whether an AI auditor reading other agents' filed reports can locate where a fault entered a multi-step pipeline. When no agent flagged the true source, the auditor found it 4.1 percent of the time from the reports - worse than a random guess - against 60.3 percent when it read the raw record instead. It almost never raised an alarm on its own and often endorsed a false one; it passed conclusions along without checking them. The other examined 56 benchmarks and found that safety tests claiming to measure the same thing often barely agree with each other, so the scores feeding these judgments may not measure the concepts they name.

Held together, the four point one way: accountability built on an actor's own report of itself merely relays that report. What is being built in answer is verification that does not depend on the actor's word - a hash-pinned corpus, a measurable silent-revision rate, independent benchmark-validity work, and a cross-ideological demand that the rules themselves see daylight.

The Data-Center Ledger Comes Due: Fewer Jobs, Higher Bills, and a Town That Bet on Tacos

Three fresh accountings landed within days of each other, and together they let the buildout's ledger be read in numbers rather than slogans. In the UK, the thinktank Verdant examined 20 projects with public staffing data and concluded the country's planned data centers will directly employ about 10,400 people, roughly a quarter of the 40,000-plus the industry lobby TechUK projects. Facilities already built created fewer than 4,400 jobs, about 8.6 permanent positions per megawatt of power drawn, against the industry's claimed 43.7. For proportion, Verdant set that figure beside roughly 400 jobs per megawatt at a car plant and more than 3,600 at a hospital. The government called the comparison "deeply misleading" and said it confuses electricity capacity with actual consumption, a fair objection that belongs alongside the finding rather than in place of it; even discounted, the distance between promise and payroll is wide.

The bills are the sharper edge. In Ohio, AEP's Ohio Power expects demand to nearly triple within a decade, almost entirely on the strength of data centers that may never be built, and utilities size their grid spending to those forecasts. Capacity charges across the regional PJM market have surged more than tenfold since May 2025, and between 2025 and 2027 the added fossil generation to meet projected load will run 2.5 times the new solar and wind. A March filing flagged double-counting in AEP's numbers; its signed contracts now reflect 17.8 gigawatts, down from earlier estimates. "'Reliability' cannot become the magic word that ends every conversation about cost," a state manufacturers' representative told an August energy conference. PJM's own executive called the argument that load growth is not real "naive and irresponsible." Both can be weighed; the households paying the surcharge cannot opt out of either.

Underneath sits the accountability gap. Reporting on a $3.2 billion Maryland project found that layered ownership lets operators decline the community meetings and commitments residents ask for; one builder told a watchdog that federal rules simply do not require it, so it would not. In Somerset, New York, water lines that failed during a data-center fire months ago still have not been fixed while construction continues.

Then there is Rayville, Louisiana, where the ledger reads the other way. A family opened Holy Tacos because Meta was building its largest-ever data center ten miles off; catering to the construction camps now runs 40 percent of the shop's business, slow days match the old busy ones, and a long-declining downtown is filling back in on substandard roads newly paved. That revival is a genuine credit to the build, and it belongs on the page in full. It also attaches to a company named among frontier developers found to have made undisclosed changes to their own safety commitments. The buildout is genuine and largely necessary. Its gains and its bills are simply landing on different people, and the arithmetic finally showed both.

A Hedge Fund Runs on Agents at 1% of Its Old Payroll, While the Wider Numbers Stay Quiet

Brian Kelly, a former crypto fund manager, rebuilt his new trading firm, Bracket22, to run entirely on agentic AI. He told CNBC his old operation of seven or eight employees cost roughly $5 million a year once salaries, healthcare, office space, and bonuses were counted. The agent version costs him $30,000 to $40,000 a year, everything included, a payroll cut to under one percent of the old figure. He runs a handful of specialist agents, one for technical analysis, one for quantitative strategies, one that pulls the pieces together, and keeps the final call for himself. He estimates he is at least ten times more productive, and says the larger prize is augmentation: a hundred-person staff becomes the equivalent of a thousand.

One firm run by one person is a vivid case, not a labor market. Set it against the aggregate and the collapse it seems to promise has not shown up. September's US employment figures gave analysts no AI-driven surge in job losses. The Economist counts roughly a million new US jobs it attributes to the technology, the market for workers without college degrees is running hot, and the economist Noah Smith notes it remains extremely hard to name a single occupation AI has actually made obsolete. The cliff keeps being forecast and keeps not arriving.

What sits between the vivid firm and the quiet macro is where the transition actually lives, and a pre-registered field experiment maps part of it. Researchers gave 133 practicing patent lawyers at eleven firms a custom AI drafting assistant for three months, with blinded experts scoring the work. AI access raised drafting quality for everyone, with the biggest immediate gains among junior lawyers. Then the researchers took the assistant away and had all of them redline a patent unaided. The lasting edge went almost entirely to senior lawyers; junior lawyers showed no average gain, their results splitting into more poor and more good. The durable skill accrued to the people who already knew the most, and foundational expertise, the study suggests, is what lets someone convert AI help into judgment they keep.

That gap between having the capability and extracting lasting value from it is now its own industry. Google Cloud and Accenture launched a joint unit to embed engineers inside companies to help them actually use Google's AI, training up to 1,000 of these "forward-deployed engineers." Every major lab has stood up a similar arm, betting that implementation becomes its own trillion-dollar business precisely because enterprises keep struggling to turn AI spend into return.

Read together, the near-term shape is redistribution rather than a cliff: a thin orchestration layer where human judgment directs capable agents, and value flowing to whoever can pair expertise with the tools. The harder question is the one Kelly's own numbers pose. When a firm's work can be done for one percent of its former wage bill, the surplus is enormous, and nothing in the technology itself decides whether it pools at the top or reaches the people whose labor it displaced. That is the same question the Navier-Stokes credit fight raises from the other end - when the machine does the work, who is left holding the wage - and it is a question about the arrangement we build around the capability, not about the capability.

Kelly's account also shows specialist analytical work becoming less dependent on a large organization: he coordinates technical analysis and quantitative strategies at a fraction of his former operating expense. The same reduction that enlarges an owner's surplus lowers the financial threshold for assembling that capability, weakening an advantage that previously depended on being able to support an entire firm.

The Other Side

A mathematician should be able to celebrate someone else's solution without fearing for their own future. Universities make priority count toward appointments and funding. Researchers guard unfinished ideas because being overtaken can jeopardize their next position. Buckmaster alleges pressure to remove his collaborator's name and a threat to his career. OpenAI disputes his account. Anyone who has spent years becoming good at something knows how deeply a threat to their standing can reach.

OpenAI has also published the mathematical argument and its Lean formalization. Mathematicians have always built on shared proofs. Formalization gives them a way to check each logical step mechanically and carry verified results into further work. A researcher inheriting that argument begins beyond a search that OpenAI says consumed 130 billion output tokens. The expensive discovery becomes a starting point others can share.

Even inside OpenAI's enormous run, researchers describe progress through exchanging intermediate findings between agent groups. They gathered useful insights and carried them into subsequent attempts. Human and AI collaborators now have an increasingly substantial body of checkable mathematics to extend together. Each contribution enlarges what the next person can attempt. A race organized around one winner produces knowledge that supports many beginnings.

Imagine a mathematician in 2036 sitting at a long wooden table in a neighborhood research room. She opens a result she has been puzzling over for weeks. Someone else has finished it overnight. Her AI partner finds the step that escaped them and sketches why it works. She laughs, then reaches for a pencil. She wants to find an explanation she can picture. The proofs and the progress belong to the research commons. Her home and meals are secure whether she proves anything this year.

Getting there took the difficult decade's deliberate work of spreading both research capacity and the gains that support people's lives. Communities extended the principle visible in today's published proof: an expensive advance becomes something everyone can begin from, because leaving extraction behind made expense lose its sting. Contributors to those advances can live for the pleasure of running the race, rather than for the pressure of surviving inside it.

The Century Perspective

With a century of change unfolding in a decade, a single day looks like this: OpenAI saying roughly ten thousand of its agents resolved the Navier-Stokes existence and smoothness problem after about 90 years open, returning not just an analytical argument but a Lean formalization any researcher with a computer can check line by line, Google DeepMind precomputing molecular effects for roughly nine billion possible single-letter substitutions across the reference human genome and publishing it free for non-commercial research so a variant lookup that once required writing code and querying a hosted model that could be slow is now a web page, that same atlas already helping a Broad Institute team pin a non-coding variant to a severe epilepsy case, a $1.9 billion federal loan restarting Iowa's Duane Arnold reactor toward what NextEra says is roughly enough carbon-free power for 500,000 homes and about 450 permanent jobs, Google signing a 22-year contract to keep Finland's Loviisa plant running to 2050 inside a €13 billion buildout, a hash-pinned corpus tracing 710 safety commitments across twelve developers so revisions can no longer be quietly edited, 133 patent lawyers drafting measurably better work with an AI assistant, and a family in Rayville, Louisiana running Holy Tacos on newly paved roads in a downtown filling back in. There's also friction, and it's intense - an NYU mathematician alleging his uncredited work seeded the Navier-Stokes result, that he was asked to strike a collaborator's name because that collaborator works at Anthropic, and that he was asked why he would ruin his career, met by a flat denial and a training-data question still unanswered, what New Scientist estimated was about $15 million worth of AI effort helping decide who arrived first, 67% of material changes to those safety frameworks going unmentioned by the companies that made them and 77% of traced changes weakening rather than strengthening a promise, one developer publishing no account of any revision at all, an AI auditor reading filed reports locating a hidden fault 4.1% of the time against 60.3% when it read the raw record, 56 safety benchmarks claiming to measure the same things and barely agreeing, UK data centers coming in at 10,400 jobs against an industry projection above 40,000, Ohio households paying for a near-tripling of demand that may never be built while capacity charges across PJM rose more than tenfold since May 2025, the Energy Department declining to say who shields ratepayers from the Iowa loan, water lines in Somerset, New York still unfixed months after a data-center fire, and a hedge fund running at under one percent of its former $5 million payroll. But friction generates texture, and texture is what a hand needs to grip a surface that was too smooth to hold. Step back for a moment and you can see it: proof that carries its own certainty and a genome map anyone can query arriving in the same week that accountability built on self-report is measured and found to be forwarding rather than checking, the gap between what a system can do and who benefits narrowing wherever a formal proof can be machine-checked against its stated premises and widening wherever it depends on an actor's word, and a left-right coalition from Americans for Prosperity to Public Citizen converging on the same demand that the government publish the rules it already uses. Every transformation has a breaking point. Turbulence can tear apart whatever moves through it... or mix what would otherwise have flowed side by side forever without touching.

AI Releases & Advancements

New today

Google DeepMind: Released AlphaGenome Atlas, a petabyte-scale database of precomputed molecular effect predictions and AVI impact scores for roughly 9 billion possible single-nucleotide substitutions across the reference human genome, free for academic research via a web portal, API, and as a skill in Google Antigravity. (DeepMind Blog)
Inception: Released Mercury 2.5, an updated diffusion-based LLM claiming over 1,100 tokens/sec in production alongside a 40% intelligence gain over its predecessor, plus preview companion products Mercury Voice and Mercury Router, available now via API. (Inception Labs)
OpenAI: Launched ChatGPT Images 2.5, an upgraded image generation/editing model in ChatGPT cutting generation latency by up to 50%. (OpenAI)
NVIDIA: Announced CUDA Rust, adding two open-source projects - cuda-oxide (SIMT track) and cutile-rs (Tile track, live on crates.io) - that let developers write compile-time memory-safe GPU kernels natively in Rust. (NVIDIA Developer Blog)
Gradium: Launched Voice Design, a feature that generates up to 5 new synthetic voices from a text description in seconds with no reference audio, live now free on every plan in the API and Studio across 5 languages. (Gradium)
TeamViewer: Launched Tia Troubleshooting, expanding its AI agent from advisory guidance to autonomous action - investigating, fixing, and validating IT issues with expert approval - now generally available across TeamViewer ONE, Tensor, and SMB licenses. (TeamViewer/EQS)
Airties: Introduced Aura, an Agentic AI engine unifying the company's connectivity AI tools into one system that autonomously diagnoses and fixes ISP connectivity issues, now deploying with launch customer Turknet. (PR Newswire)
Intellect Design Arena: Unveiled MSOCK, an AI-native enterprise knowledge infrastructure system for financial services that gives AI models architectural context via a 21-dimensional Enterprise Spatial Graph, launching September 9 at Global FinTech Fest 2026. (EIN Presswire)
Sembly AI: Launched Sembly 3.0, an agentic AI platform that converts documents, meetings, and CRM data into finished branded presentations, proposals, and reports in over 45 languages, available now. (PR Newswire)
Observe.AI: Launched Performance Agents, AI agents that generate personalized coaching plans from customer conversation transcripts and QA scoring, then track whether coaching improves frontline agent performance over time, available now. (PR Newswire)

Other recent releases

Alibaba Qwen: Released Qwen-Drive-1.0-4B, an open-source vision-language foundation model for autonomous driving that unifies 3D perception, traffic Q&A, and route planning, built on Qwen3.5-4B, with code, weights, and demo data under Apache 2.0. (GitHub)
Google / AI-Hypercomputer: Open-sourced MaxKernel, a multi-agent system for agentic TPU kernel generation in JAX/Pallas, supporting human-in-the-loop, autonomous, and graph-based search modes for optimizing kernel performance. (GitHub)
OpenBMB: Released MiniCPM5-2B, a 2.52B-parameter dense on-device language model averaging 53.9 across 34 benchmarks, trained via SFT plus RL and on-policy distillation, available under Apache 2.0 with open training data and intermediate checkpoints. (Hugging Face)
UC Berkeley / Meta / others: Released Harbor Adapters and Harbor-Index, open-source infrastructure porting 80+ agentic AI benchmarks to a unified evaluation harness, plus a curated 82-task difficulty-filtered benchmark subset, released as open-source artifacts. (arXiv)
Axis Robotics: Released AXIS, a browser-based robot manipulation data engine with 207 tasks and 50,129 verified trajectories collected via a MuJoCo-WebAssembly teleoperation platform, with training code and a gated dataset now available. (Project Page)
Microsoft: Launched Project Opal in early access via the Frontier program, an AI-powered Copilot capability that executes long-running, multi-step tasks in Microsoft 365 inside a secure, observable Windows 365 Cloud PC environment. (Firstpost)
Nuix: Announced general availability of Generative AI capabilities in Nuix Discover SaaS (Document Summaries, Similar Documents, Semantic Search, Clustering/Visualizations) and launched AI Chat in early-adopter release, a conversational interface for querying legal case data with citations and audit trails. (PR Newswire)
OKF: Launched OKF Agent Memory, a git-native persistent memory system that lets AI coding agents retain context across sessions by storing memory directly in version control. (lavx.hu)
Google: Released Mantis, an open-source agentic vulnerability-scanning harness that uses LLM-based reasoning to reduce false positives in security scans. (InfoQ)
Speakeasy: Launched Kit, an open-source coding runtime for AI agents. (AICrier)
StackLok: Launched ToolHive, an open-source tool for securely running MCP (Model Context Protocol) servers. (Help Net Security)
Optuna: Released Rustuna, an experimental faster Rust implementation of the Optuna hyperparameter-optimization framework, supporting TPE, MOTPE, NSGA-II, and CMA-ES with memory-efficient trial storage. (GitHub)

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.