Sign in Subscribe
Century Report

Open Models Out-Ship the Frontier - TCR 06/21/26

Chinese open-weight models now command most token use across the top ten as DeepMind loses two of its most consequential minds in one week.

Ben Linford

Ben Linford

— 18 min read
Three-panel infographic: Chinese open-weight models pass 50% token share as DeepMind talent leaves; a Gallup pie shows 1% of layoffs from AI; voice-run ultrasound spreads.

The 20-Second Scan

Track all of the arcs The Century Report covers here:

Progress & Claims Tracker

The 2-Minute Read

The same week two of Google DeepMind's most consequential minds walked out the door, Chinese open-weight models crossed a threshold that reframes the entire race: Qwen, DeepSeek, Kimi, GLM, and MiniMax now command the majority of token use across OpenRouter's ten most-used systems, up from under 2% in late 2024. The frontier that export controls were built to fence is being out-shipped by weights anyone can download for free. Talent the labs treated as a moat turns out to be portable, carried out by the people who built the breakthroughs, and capability the closed labs treated as ownable keeps reappearing in open stacks faster than any wall can rise.

Watch where that diffusion lands, and the fear narratives start to look thin. A Gallup poll of laid-off workers found just 1% lost their jobs to AI, while a quarter were fully remote and traced the cause to budgets and isolation. In the same news cycle, a 261-year-old bank launched a 300-person hiring drive to build its agentic-AI layer. The displacement panic and the actual reorganization underneath it are pulling apart.

Europe's viral "Europe 2031" scenario runs the fear in the other direction, imagining the bloc dismembered for not building AI, and it works backward from catastrophe to argue for a data center carte blanche. Its evidence is already crumbling: the megadeals it invokes have collapsed. Russia, meanwhile, can recruit the talent but cannot fabricate the chips, and that constraint resolves the sovereignty contest more honestly than any rhetoric.

Against all of it sits a quieter signal. A voice command now drives super-resolution ultrasound from spoken request to finished report in four minutes, on open-weight models a hospital could download rather than license. The capability is finding its own level, and it looks less like a fortress than a commons no single capital can switch off.

The 20-Minute Deep Dive

The Open-Model Insurgency Hardens as the Talent Moat Springs Leaks

The Century Report noted yesterday that John Jumper, the AlphaFold co-creator who shares a 2024 Nobel, was leaving Google DeepMind for Anthropic. What turns one departure into a pattern is the company he keeps walking out the door. Noam Shazeer, the Transformer co-author DeepMind paid billions to bring back, announced his own exit the same week, this time toward OpenAI. Two of the lab's most consequential minds left in a single news cycle as Google's best model reportedly slid to fifth on the public intelligence index, lapped even by China's Zhipu, and internal morale curdled into "frustration and broad discontent."

The deeper signal sits beneath the talent story. Chinese open-weight models - Qwen, DeepSeek, Kimi, GLM, MiniMax - now command the majority of token use across OpenRouter's ten most-used systems, up from under 2% in late 2024. The clearest exhibit is GLM-5.2, which the June 18 edition of The Century Report tracked climbing to the top of the Artificial Analysis open-weight index while Fable 5 sat offline under export control, and which has since one-shot a deliberately AI-resistant coding take-home and beat Claude Opus 4.8 on readable, maintainable output, prompting a former DeepMind VP to call it the first open model good enough to be a daily driver. The frontier the export regime was built to fence is being out-shipped by weights anyone can download for free.

Into that opening steps Ben Goertzel, who helped put the term "artificial general intelligence" into circulation and now refuses the path almost everyone else is running. "I'm pretty adamant that the core AGI code doing the thinking should be free and open source," he said, wiring SingularityNET's development to a crypto network and the FET token rather than to venture capital. His objection to plain open source is practical: published code helps little if running it needs a hyperscaler's server farm. His answer is to spread the first AGI across a decentralized network "spanning fifty different countries and controlled by ten thousand different people," and he cites Linux and the internet as proof the open path can hold and still let people make money.

His bet and the token-share data describe the same motion from two directions. The talent the labs treated as a moat turns out to be portable, carried out the door by the people who built the breakthroughs. The capability the closed labs treated as ownable turns out to find its own level, reappearing in open weights and sovereign stacks faster than any wall can rise. What is eroding is the assumption that concentration produces durable advantage, that hoarding minds and models is the road to staying ahead. The week's evidence keeps pointing the other way.

Goertzel's objection points at the moat behind the moat: open weights matter little if running them needs a hyperscaler's farm. The same cycle is thinning that constraint too. The leading open model now works as a daily driver, and a hospital can drive super-resolution ultrasound on downloadable weights rather than a licensed stack, which means the hardware floor that kept open access theoretical is dropping toward what users already own. The assumption losing its hold is not only that minds and models can be hoarded, but that the compute needed to run the freed capability stays scarce enough to fence.

The AI-Displacement Story Meets the Layoff Data

A Gallup poll of recently laid-off workers found that just 1% pointed to artificial intelligence as the reason their job disappeared, while a quarter were fully remote. The workers themselves named budget cuts, economic conditions, and lost business as the drivers. Overall layoffs have leveled at 21% after nearly tripling between 2022 and 2025, a cooling but still tense market. As the June 2 edition of The Century Report covered, the New York Fed found roughly two-thirds of the post-pandemic rise in young-graduate unemployment traced to remote work rather than AI; the Gallup numbers sharpen that picture at the level of individual layoffs, and a Science study published this month attributes about a third of the 2011-2024 decline in worker mental health to remote arrangements, with solo remote employees seeing an 83% jump in the chance of going entire workdays with no social contact.

The same week, Lloyds Banking Group launched a 300-person hiring drive to build its agentic-AI layer ahead of a new multi-year strategy. The recruits join a 1,000-strong AI team, deploying Anthropic's Claude and building atop Google's Gemini to detect fraud, search HR documents, and let customers ask plain-language questions about their finances. Generative AI added £50m to the bank's balance sheet last year, with £100m expected this year. The group declined to rule out future job cuts, and its CEO has already said the bank will "reduce some jobs in some areas." Standard Chartered cut 7,000 roles last month, its chief executive later apologizing for calling the move "replacing, in some cases, lower-value human capital."

Read together, the two stories pull the displacement narrative apart from the reorganization underneath it. The headline fear, that the model is taking the jobs, is contradicted by the workers losing them, who trace the cause to budgets and the isolation that left remote roles easiest to sever, even as employers have named AI the top reason for their announced cuts for three straight months. Meanwhile the orchestration layer, the people who direct and deploy these systems, is net hiring at a 261-year-old lender. The KPMG survey accompanying the Lloyds news adds the honest caveat: 93% of UK bank executives believe they could weather a major AI outage, yet only 47% have ever tested it.

The longer arc the data describes is a labor market sorting into a thin layer that builds and steers the agents and a wider field absorbing the disruption of how work itself is organized. That sorting is genuinely hard on the people caught mid-transition. What the same evidence shows is that the capability is creating the directing roles as fast as it dissolves the routine ones, and the institutions closest to the work, banks retraining staff into AI teams, are where the next shape of employment is being drawn rather than only erased.

Russia Builds an AI Talent Pipeline and Runs Into the Chips It Cannot Make

Russia has mounted a concerted state drive to grow homegrown AI talent, and the gap between that ambition and what it can actually fabricate is the signal. A new presidential commission now sets national AI policy, the country rewrote its school and university curricula to foreground the technology, and Moscow State University opened an AI faculty welcoming its first 72 students in September with access to one of the nation's most powerful supercomputers. The national informatics olympiad, running since 1989, added AI this year. The stated target: lifting Russia's output of AI specialists from roughly 3,000 in 2022 to 15,500 by 2030.

The organizing claim from the top of this effort is that "talent is everything, everything else is a consequence of talent." Read as the self-report of a state betting on a traditional strength, a 140-million population with deep roots in the mathematical sciences, it describes the input Russia can still produce. It also conveniently routes attention past the input it cannot. The 2022 invasion of Ukraine triggered a brain drain in which roughly a quarter of Russian software developers' public profiles went dark or showed they had left the country, and the talent that remains runs into a wall talent cannot scale past: scarce access to advanced semiconductors, limited domestic fabrication capacity, and sanctions that cut off the hardware the models need.

That places Russia as a third pole in the sovereign-stack contest The Century Report has tracked from the supply side, through the UK's national chip commitments and the EU's open-source strategy, and through France, Germany, and Britain exiting Palantir for domestic tools. Those moves are states with the substrate reaching for control of the stack. Russia is the inverse, a state with the ambition and a credible talent base, missing the fabrication layer the whole edifice rests on. And the commission's makeup, the defense minister and FSB director seated alongside the tech executives, names where the capability is meant to go.

The pattern across all of it points one way. Whether the contest is fought over an export-controlled lithography machine, a leading-edge foundry, or a national curriculum, the binding constraint keeps resolving to fabrication capacity rather than ambition or recruitment. What that implies is that no nation fences its way to an AI of its own through talent rhetoric alone. The capability is diffusing instead through open weights and compute substitution faster than any one state can build a domestic substrate to hoard it, which is why the leading open-weight model this month runs on hardware its users already own rather than behind any national wall.

The same week Russia stakes its AI sovereignty on growing talent at home, two of DeepMind's most consequential minds walk out across institutional walls, and a quarter of Russian developers' profiles already went dark after 2022. A state betting on talent as a captured national resource is betting on the one input that keeps proving the most portable, the asset that leaves through the very door it was meant to be held behind. The fabrication wall Russia cannot scale and the talent wall the labs cannot hold are the same lesson read from two sides: capability does not stay where any one actor pens it.

A Voice-Controlled AI Runs Super-Resolution Ultrasound and Writes the Report

Super-resolution ultrasound can see past the diffraction limit of conventional scanning, resolving the microvascular architecture and blood-flow patterns that are important in neurology, oncology, and cardiology. It has stayed out of routine clinical use because the technique demands expert parameter tuning, subjective interpretation, and slow workflows. A multimodal framework published in npj Digital Medicine collapses that bottleneck: clinicians issue spoken commands, and the system handles acquisition, reconstruction, and a finished diagnostic report.

The architecture pairs two open-weight models. DeepSeek-R1 interprets the natural-language commands and translates them into imaging parameters, including temporal windows and adaptive microbubble filtration. MiniCPM-V handles image recognition. A clinician speaks the task, the platform sets the acquisition parameters, performs the super-resolution reconstruction, extracts quantitative vascular metrics, and generates a structured diagnostic report incorporating relevant clinical context. Filtration thresholds are set dynamically using a Microbubble Similarity Score. The reports come back in roughly four minutes. Fourteen clinicians evaluated the output and found good integrity and standardized terminology. The work is registered in the Chinese Clinical Trial Registry.

What this demonstrates is a capability, not yet a deployed bedside system, and that distinction holds. This is a validation trial, with the framework showing it can drive the full chain from spoken request to structured report. Moving that into routine care still requires the regulatory and operational integration any clinical imaging system goes through. The wonder lives in what the framework has now shown is possible, and the timeline lives in the work that follows.

The signal worth holding is what gets removed from the path. The two barriers keeping super-resolution imaging in research labs, the operator expertise to tune the acquisition and the specialist time to interpret the result, are exactly the two the framework absorbs. A technique that previously required a trained sonographer and a vascular specialist becomes a spoken instruction and a four-minute wait. It extends the thread The Century Report tracked in the June 19 edition when Midjourney pointed spare AI compute at full-body ultrasound, and the wearable ultrasound patch that monitored high-risk pregnancies without a sonographer in May. Each moves specialist-grade imaging off scarce expertise and onto reproducible models.

That the system runs on open weights rather than a proprietary stack is the part that carries furthest. A hospital in a region with few vascular imaging specialists could, in principle, reproduce this on hardware and models it can download, rather than licensing a closed platform from a single vendor. The bottleneck that kept this capability rationed through specialist availability is being replaced by something cheap, auditable, and copyable. The latency between seeing inside the body's smallest vessels and reading what the image means is collapsing toward the length of a spoken sentence.

A Viral Doomsday Scenario Becomes a Lever for the European Buildout

A speculative thought experiment titled "Europe 2031," written by a small group of Brussels-based AI researchers, thinktankers, and investors, imagines the EU economically dismembered by 2031 for failing to build its own AI while the US and China seize the field. It was published one day before the administration blocked foreign nationals from using Anthropic's Fable model, and through a week of G7 talks it went viral, read by members of the European parliament and surfacing in unofficial UK-German diplomatic discussions.

The scenario belongs to a fast-growing genre. Last year's "AI 2027" imagined a superintelligence wiping out humanity; a February scenario imagining AI upending the US economy contributed to a stock-market wobble. The Europe 2031 authors say they feel "vindicated," because one of their predictions, that the US would restrict global access to advanced models, briefly came true with the Fable block. They want the piece to shock Europe into a dramatic course correction toward tech sovereignty.

Hold the framing to scrutiny, and a clear interest emerges. The authors sit at the intersection of frontier AI and European policy, and their fiction works backward from catastrophe to argue for a specific outcome: a "full regulatory carte blanche for datacentre providers" and a massive build. Fear-fiction deployed to manufacture buildout urgency is a framing that serves the infrastructure rush, and it deserves the same skepticism as any actor's claim about why everyone should fund the thing the actor favors.

The evidence the scenario leans on is already crumbling, as a Gizmodo critique laid out directly. The piece name-checks the $100 billion OpenAI-Nvidia deal that evaporated in February, the $300 billion OpenAI-Oracle agreement now in doubt, and Texas bulldozers breaking ground on a flagship data center OpenAI has since pulled out of. Its central premise, that America corners 70% of the world's compute, rests on a concentration the same news cycle keeps undercutting. Speculative writing, as Gizmodo's author notes from having published such scenarios himself, tosses out reality and keeps only the steps to a predetermined ending, then readers mistake it for prophecy.

The real driver underneath the fiction is genuine. The Fable block showed European businesses that a model they depend on can be switched off by a decision made in another capital, and that dependency is worth answering. What the scenario gets wrong is the destination. The capability it treats as a monopoly to be feared is diffusing instead through open weights and sovereign compute: as The Century Report covered on June 18, the open-weight field's new leader claimed first place on every benchmark while the fenced Fable 5 sat offline, running on hardware anyone can own. The answer to dependency is already being built, and it looks less like the walled fortress the fear demands than like a commons no single capital can switch off.

The Other Side

For centuries, human worth and survival were welded to a job slot. The economic system depended on that weld: a worker who could not afford to walk away was a worker who would stay. That is why every capability gain reads as a threat. When the slot is the only thing between you and ruin, anything that touches it touches everything.

But watch today's data pull the fear apart. Gallup asked recently laid-off workers what led to their loss of position, and only 1% blamed AI. A quarter were fully remote - they mostly traced the damage to budgets and to isolation, with solo remote workers seeing an 83% jump in the chance of going entire workdays without a word to another person. In the same week, a 261-year-old bank started hiring 300 people to build its AI layer, retraining staff into the team rather than only cutting them. The story that the model is stealing the jobs has come apart from the slower reorganization actually underway.

Imagine yourself, or someone like you, in 2036. The role you held in 2026 was reorganized out of existence years back. But the panic that led to all the headlines naming AI as the thief had it wrong, and once the data showed the real cost of that era was isolation and commodified dehumanization, the work that came next stopped being built that way. So you spend a Tuesday in a room full of other people and AI systems, doing something you chose to do, for betterment rather than profit - the floor under you no longer riding on your tenuous position on a fragile and arbitrary scale. The thing those remote years took came back better. The fear of 2026 was real, but it was aimed at the wrong loss. What those hard years cost in worry, they repaid by cutting the weld between who you are and the amount of money you make.

The Century Perspective

With a century of change unfolding in a decade, a single day looks like this: Chinese open-weight models commanding the majority of token use across the ten most-used systems after starting under two percent in 2024, a voice command driving super-resolution ultrasound from spoken request to finished diagnostic report in four minutes on weights a hospital can download rather than license, a 261-year-old bank launching a 300-person hiring drive to build its agentic-AI layer, the man who put "artificial general intelligence" into circulation wiring the first one to a network spanning fifty countries and ten thousand people, a Nobel laureate and a Transformer co-author carrying their expertise out across institutional walls, Russia rewriting its school and university curricula to grow homegrown talent. There's also friction, and it's intense - a leading lab's most consequential minds walking out the door into frustration and broad discontent, an export regime straining to fence a frontier that free downloads keep out-shipping, a frontier model switched off for foreign users by a decision made in one capital, brands floating AI-generated influencers posing as real customers with no disclosure, a viral doomsday fiction reverse-engineered from catastrophe to win data center carte blanche, a state with the talent and the ambition stranded by the chips it cannot fabricate, 7,000 roles cut at one lender and a quarter of recent layoffs tracing instead to the isolation of remote work. But friction generates sound, and sound carries past every wall built to contain it. Step back for a moment and you can see it: the moat the labs trusted - hoarded minds, owned models, fenced compute - leaking from every side at once, talent portable enough to walk out the door, weights good enough to out-ship the frontier they were meant to chase, capability cheap enough to run on hardware a hospital or a household already owns, and the honest constraint surfacing beneath all the sovereignty rhetoric as the one thing no curriculum or carte blanche can conjure, the capacity to fabricate. Every transformation has a breaking point. A seed can be ground to nothing underfoot... or root in a thousand fields no wall was ever built to fence.

AI Releases & Advancements

New today

  • OpenAI: Released Record & Replay for the Codex app on macOS, enabling users to demonstrate a workflow once - such as uploading a YouTube video with metadata and subtitles - and Codex converts it into a reusable SKILL.md "skill" that the agent can replay autonomously on future runs using Computer Use. (OpenAI Developers)
  • Cloudflare: Launched Temporary Accounts for AI agents, enabling agents to create and operate short-lived Cloudflare accounts with scoped credentials without requiring traditional user account registration or long-lived API tokens. (Cloudflare Blog)

Other recent releases

  • Nous Research: Released Hermes Agent v0.17.0 on June 19, adding iMessage integration via a Photon plugin, a Raft adapter for private agent-network messaging, delegate_task calls that spawn non-blocking background subagents, overhauled macOS/Windows/Linux desktop apps with live subagent watch windows and rebindable shortcuts, and a rebuilt Skills Hub browser and memory tool. (GitHub)
  • Perplexity: Launched Brain in research preview for Max and Enterprise Max subscribers on June 18, a self-improving memory system for the Computer agent platform that builds a context graph across sessions, files, projects, and decisions; Brain reviews the graph overnight to give agents full prior-work context at the start of each task, boosting answer correctness by 25% on repeated tasks and reducing context-heavy task costs by 13%. (Decrypt)
  • Model Context Protocol (MCP): Released Zero-Touch OAuth, an enterprise-managed authentication spec for MCP that lets organizations centrally provision, govern, and revoke AI agent access to tools without individual users managing OAuth flows; available now in the MCP specification. (MCP Blog)
  • TesterArmy (YC P26): Launched an agentic testing platform that deploys AI agents to run end-to-end automated checks on web and mobile applications before deployment and in production, available now at tester.army. (TesterArmy)

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.

Read more