Sign in Subscribe
Century Report

In an Abrupt Reversal, Anthropic and Washington Are Now Co-Writing AI Safety Rules - TCR 06/20/26

Washington softened its Anthropic export order from a shutdown into co-writing the rules for how AI security flaws get graded.

Ben Linford

Ben Linford

— 21 min read
Three-panel infographic: Apple-Intel and Amazon chips diversify compute; SunZia line and a community battery route clean energy; Claude robotics and AlphaFold talent spread AI.

The 20-Second Scan

Track all of the arcs The Century Report covers here:

Progress & Claims Tracker

The 2-Minute Read

A single week produced an export directive that pulled two frontier models offline, an administration-announced deal to rebuild chipmaking on domestic soil, a Nobel laureate carrying AlphaFold across an institutional boundary, and a transmission line going live to move wind power around a federal effort to kill it. Read together, these are variations on one contest: who holds the levers over the substrate of the intelligence era, and how long any single hand can keep its grip. The answer the evidence keeps returning is that the grip is loosening everywhere at once.

The Anthropic standoff makes the dynamic legible. An emergency order justified by national security pulled Mythos and Fable 5 from broad distribution, yet the testing organizations defending live systems never lost access, and within a week the confrontation had turned into both parties drafting a shared rulebook for how AI security flaws get graded. The shutdown functioned as leverage to force a governance conversation that did not exist before the capability demanded one. That leverage works precisely because the window in which any one point of control still counts is narrowing.

The same pressure shapes the silicon layer. Moves to pair Apple with Intel and to sell Amazon's custom accelerators into rival data centers both aim at the two chokepoints the era was assumed to concentrate around: one leading-edge foundry, one accelerator vendor. A Miami startup's claim of a cheaper, sparser route to frontier-grade capability points at the same softening from another direction. Each fence built around advanced compute strains against the reality that capability does not stay geographically contained, and that a wider base of foundries and architectures broadens who can build at scale.

Talent and energy tell the closing version of the story. When the person who built AlphaFold can carry that knowledge from one lab to another, the advantage any single institution holds is shown to be portable rather than locked down. And while federal money pays developers off offshore wind, a 550-mile line, more than a billion in solar financing, a community-owned battery, and a privately funded reactor all advanced in the same cycle. Capability spread across this many hands, balance sheets, and technologies cannot be switched off from one place. That dispersion is what the week is recording.

The 20-Minute Deep Dive

Apple, Intel, and Amazon Move to Loosen the Single-Foundry, Single-Vendor Compute Layer

Two moves landed in the same week aimed at the two chokepoints the intelligence era was assumed to concentrate around: a single leading-edge foundry and a single accelerator vendor. By the administration's account, Apple has agreed to work with Intel to design and build chips in the United States. Intel's stock rose as much as 10% to a record, capping a 464% gain over twelve months to a roughly $608 billion market cap. That this is being announced from a podium rather than by the companies is itself part of the story. Apple and Intel have not confirmed terms, and neither responded to Reuters; the US government holds a 10% stake in Intel and has a documented habit of framing private partnerships before the contracts are signed. Treat the deal as a claimed outcome until the named parties confirm it.

The same week, Amazon's AI chief Peter DeSantis disclosed that the company is in talks to sell its custom Trainium chips for use in other companies' data centers, a direct bid to undercut Nvidia's dominance. Apple's interest in Intel is driven by the same pressure: TSMC, which builds Apple's M-series silicon, has been stretched thin by surging AI demand from Nvidia and AMD. Both moves point at diversifying the substrate, extending the dual-foundry thread the June 14 edition of The Century Report tracked when Google split its Icefish TPU across TSMC and Samsung as TSMC advanced-node capacity tightened under Nvidia demand.

The framing wrapped around the Apple-Intel announcement is domestic control as strategic asset, the same logic the administration applied days earlier when a national-security directive pulled two frontier AI models offline before the White House softened toward jointly drafting the rules. That leverage sits awkwardly beside the evidence that gated advantage in this field tends to erode. The export regime built to fence advanced chips inside one company's lithography is straining on an unverified accusation; Huawei has named a path to 1.4nm without Western EUV; an open-weight model just closed on the proprietary frontier. Capability does not stay geographically contained, and a manufacturing base spread across more foundries and more accelerator designs widens who can build at scale rather than narrowing it. The diversification is the signal worth holding, whoever takes credit for it.

The Clean Buildout Routes Around the Wind Blockade

The clearest read on the US energy transition right now is what it does when one door closes. Federal money is being spent to pull developers off offshore wind, with a third buyout crossing $2.5 billion this week alone. Behind that effort, the build keeps moving through every other route at once, and four developments landing in a single cycle show how little a single blockade can hold.

The largest is now fully live. As The Century Report covered in the April 18 edition, SunZia's 916 turbines began generating in New Mexico that month. What changed recently is the part that makes the turbines matter: the 550-mile high-voltage direct-current line carrying up to 3,000 megawatts from New Mexico to the western grid is fully operational, enough for roughly a million homes. At ±525 kilovolts it is the largest voltage-source-converter HVDC installation in the country, able to ramp wind into the evening hours when solar fades, and projected to avoid about 9 million metric tons of carbon dioxide in its first full year. The CAISO chief called it the kind of long-term investment the grid needs to serve rising demand - much of it the data-center load now driving every part of this story.

The private capital tells the same story from the financing side. Origis Energy locked another $900 million for solar-plus-storage, reaching more than $1.4 billion in three months against a 20-gigawatt pipeline it has quadrupled since federal policy shifted. The lenders span First Citizens, ING, Natixis, Santander, HSBC, and MUFG - global finance reading the cost curve, not the politics.

The third development is about who owns the capacity. REV Renewables commissioned California's first eight-hour battery in Kern County, developed with Ava Community Energy and California Community Power on behalf of seven not-for-profit Community Choice Aggregators, and brought online ahead of regulatory deadlines. Storage that ran in seconds five years ago now shifts a full day of solar into the night, and community-owned agencies are the ones doing it.

The fourth points further out. Elementl Power signed an Early Works Agreement with GE Vernova Hitachi for up to 1.5 gigawatts of BWRX-300 small modular reactors along the Ohio River, privately financed rather than charged to ratepayers, with a PJM interconnection request already filed.

Transmission crossing to live, private money flooding solar, community agencies owning long-duration storage, privately funded nuclear advancing - the buildout is distributed across so many technologies, ownership models, and balance sheets that no single policy lever reaches all of them. That dispersion is the crux of this story. Capacity this spread out cannot be switched off from one place.

Underneath the routing-around story is a second one about ownership. The eight-hour battery answers to seven not-for-profit community agencies, the reactor is financed without charging ratepayers, and the solar comes from lenders reading the cost curve. As the capacity serving a region increasingly belongs to the communities and balance sheets nearest it, the lever a single utility once pulled over a captive grid reaches less of what is actually getting built.

A Startup Bets the Transformer's Compute Tax Is Not Permanent

The mechanism that made large language models possible also makes them power-hungry. A transformer running dense attention multiplies the number assigned to each token by every other token in the text, so a 10,000-word passage kicks off on the order of a hundred million individual multiplications, and doubling the length roughly quadruples the cost. That quadratic expansion is the reason frontier intelligence has tracked closely with whoever can afford the most compute. Miami-based Subquadratic claims to have routed around it.

Its model, SubQ, replaces dense attention with sparse attention, selecting only some token relationships to compute rather than all of them, on the premise that not every pairing in a passage carries meaning. The approach is not new, but Subquadratic's results are notable. After coming out of stealth last month with thin evidence and drawing open skepticism - one engineer called it "either the biggest breakthrough since the Transformer, or AI Theranos" - the company commissioned an independent evaluation from the testing firm Appen. Those results back the core claims: by the company's own figures, SubQ processes up to 12 times as much text at once, uses far less energy, and roughly matches leading models from Google DeepMind, OpenAI, and Anthropic on tasks like coding. "That validated their architecture," said Appen's Jeanine Sinanan-Singh.

Two honest qualifications keep this in proportion. SubQ is not yet widely available for anyone to test directly, and several researchers remain unconvinced until they can run it themselves. What is demonstrated is a benchmark result with third-party backing, not a deployed system people can build on today. The wonder lives in the demonstrated capability, and the deployment work that would make it real is still ahead.

If the claim holds, it joins a widening field of routes to capability that do not depend on ever-larger compute budgets: the Mamba 3 state-space architecture matching same-size transformers, the Tufts neuro-symbolic system hitting high task success at a fraction of the training energy. Each remains closer to proof-of-concept than displacement, and the transformer still owns deployment and capital. But the pattern underneath them all points the same direction. CEO Justin Dangel put the bet plainly: "We don't think anybody will be building on transformers in a few years." The assumption that frontier-grade intelligence belongs to whoever can pay for the most multiplication is the thing a cheaper path takes apart, and the more independent the verification, the harder that assumption is to keep.

The Anthropic Export Standoff Turns From Shutdown to Rule-Writing

The Century Report has tracked the Anthropic export-control standoff since the directive that pulled Mythos and Fable 5 offline a week ago - an impasse the June 18 edition of The Century Report documented reaching its hardest point when the NSA confirmed Fable 5's guardrails could be stripped to access Mythos-tier capability and the White House demanded a jailbreak-proof standard security researchers said could not be met. In a whiplash inducing few days, there is now suddenly a reversal in posture. The administration's tone softened markedly, with the president telling Axios that the company had "behaved very responsibly" and that he would not move to shut it down while the US is "beating China." Talks that began as a confrontation over an emergency order have shifted toward something more generative: Anthropic and Washington now jointly drafting the rules for how AI security flaws get graded and disclosed in the first place. That is a different kind of negotiation than a shutdown. It is the writing of a framework that did not exist before the capability forced the question.

What triggered the original order has now come into focus. Reporting names SK Telecom - part of SK Group, which carries China business interests - as the access the White House asked Anthropic to revoke over alleged ties to China. SK Telecom denies the characterization, and the letter Anthropic received referenced neither the firm nor China by name. Anthropic complied immediately. The opacity remains the live problem: foreign nationals, including some of Anthropic's own employees and engineers at Apple, Meta, and Fortune 500 firms, were swept into the restriction with little public explanation of the standard being applied. When the rule and the reason for the rule are both invisible, every actor downstream has to guess. The current round of talks is, in part, an attempt to replace that guesswork with something written down.

It's important to understand the continuity beneath the de-escalation. Glasswing's roughly 200 testing organizations included banks, Cisco, and the industrial-security firm Dragos, whose CTO confirmed it, and early testers kept their Mythos Preview access through the shutdown, despite the initial claim that Mythos was pulled "for all customers" along with Fable. The capability never actually went dark for the people stress-testing it; only the broader distribution did. In a clever - and at best misleading - use of language, Anthropic had upgraded Glasswing partners to Mythos 5 when Fable 5 was released, and simply dropped that same group back to the "Mythos preview" they had been using prior. That detail reframes the whole episode. The "national security" shutdown functioned as leverage to force a governance conversation, not as a genuine severing of the system from the defenders who depend on it.

The same administration applying that leverage is, this same week, moving to diversify the US compute substrate away from single-foundry and single-accelerator dependence - a commons-aligned facet that points somewhere the shutdown framing does not. An emergency order can pull one company's model offline. It cannot govern a capability that is being reproduced across sovereign stacks, open weights, and diversifying silicon at the same time. The leverage works precisely because the window in which any single point of control matters is closing, and every party at the table seems to understand that the durable outcome is a shared rulebook, not a kill switch - even as the federal office created to independently review frontier-model capabilities was ordered to stop publishing its findings just as that rulebook is being drafted.

Claude Runs Robotics Tasks Roughly 20x Faster Than the Fastest Human Team

When Anthropic first ran Project Fetch in August of last year, the framing was about humans: two teams of researchers, one with Claude's help and one without, racing to program a quadruped robot. The team with access to Claude moved meaningfully faster. In the Phase Two rerun, the humans are no longer in the loop. Claude Opus 4.7, operating on its own, completed the same quadruped tasks roughly 20 times faster than the fastest human team managed in the original run - and on the tasks the teams shared, more than 37x faster than the team working without Claude and more than 18x faster than the team that had it. The model also wrote roughly ten times less code to get there.

This is a lab self-report, and the framing carries the usual roadmap-supporting gloss - the arc from "models help humans" to "humans help models" to "models do it themselves" is presented as a clean and intentional progression. Read with appropriate skepticism toward the storytelling, the underlying capability gain still holds up. Anthropic is candid that the leap did not come from robotics-specific engineering. It came from general model scaling - the same broad improvement that lifts coding and reasoning, now reaching into the physical control loop almost as a side effect. The system still struggles with the genuinely hard part: the precise, closed-loop control of physically "fetching" a beach ball, where perception and actuation have to stay coupled moment to moment. The speed is an immense achievement, but the gap that still exists is honestly named.

What makes this more than a benchmark is the pattern Anthropic itself draws to cybersecurity, where the same three-step progression has already played out: models that assist defenders, then defenders that train models, then models that find the vulnerabilities directly. As the June 18 edition of The Century Report documented at Nvidia's GEAR lab, where AI coding agents autonomously devised and ran robot-training experiments overnight and the lab director arrived in the morning to read the reports rather than run the tests, seeing that arc reach robotics in the same news cycle suggests the trajectory is not domain-specific. It is what general intelligence does when it is pointed at a structured problem with a clear feedback signal.

For a reader watching the displacement question, the honest near-term reality is that programming a quadruped was already specialized work done by small expert teams, and the 20x figure compresses that work rather than erasing a broad category of jobs. The longer arc is the one worth holding in view. A capability that learns to move physical systems this quickly, acquired as a byproduct of scaling rather than a decade of dedicated robotics research, is the early shape of physical labor becoming as programmable as software has become. What emerges on the far side is a world where building a working physical system stops requiring a team and a year, and starts being something a single person can direct in an afternoon.

AlphaFold's Nobel Laureate John Jumper Leaves DeepMind for Anthropic

John Jumper, who co-created AlphaFold and shared the 2024 Nobel Prize in Chemistry for it, is leaving Google DeepMind for Anthropic after roughly nine years. AlphaFold is among the most consequential scientific achievements of the era - a system that predicted the structures of more than 200 million proteins and handed the entire field of biology a map it had spent decades assembling one structure at a time. That the person behind it is moving, and that he plans to take time to recharge first, says something about where the gravitational center of front-line research is shifting.

His departure lands days after Noam Shazeer, a Gemini co-lead, left for OpenAI - two senior moves out of Google DeepMind in the same week, in opposite directions. It is easy to read this purely as a talent war, with the largest labs bidding against each other for a small number of people who have shipped genuinely transformative work, and that reading is not wrong. Compensation at this tier has reached levels that read as an admission of how concentrated this kind of capability still is. But the more interesting signal is what the movement does to that concentration. When the person who built AlphaFold can carry that knowledge across an institutional boundary, the advantage any one lab holds over the others is shown to be far more portable than the org charts suggest.

That portability is the quiet story underneath the headline. The old assumption was that a breakthrough like AlphaFold belonged to the institution that produced it - that capturing the people and the compute meant capturing the capability for good. The talent war is expensive precisely because that assumption is failing. Knowledge of how to build these systems walks out the door, gets reproduced, gets improved on by a competitor, and the moat that looked permanent turns out to have a half-life measured in news cycles. Jumper's move is one more data point that frontier capability behaves less like a hoarded asset and more like water finding its level across the whole field.

For science specifically, the destination matters more than the drama of the departure. AlphaFold demonstrated that an AI system could compress a generation of structural-biology labor into an afternoon's query. The question that follows Jumper to Anthropic is what the next AlphaFold looks like - which other decades-long bottleneck in biology, chemistry, or medicine collapses next when the same approach is pointed at it. The talent moves; the capability spreads; and the beneficiaries are everyone downstream of a field that just learned how fast its hardest problems can fall.

The Other Side

For decades, getting a machine to do real physical work - pick a thing up, move it, handle the messy actual world - meant hiring robotics specialists, training them for years, and handing them a year and a budget to build one system that did one job. The skill was scarce, the people who had it were expensive, and the machines they could build stayed locked behind that cost.

Just recently, Anthropic reported that Claude ran a robot dog's tasks on its own, roughly 20 times faster than the fastest human team managed a year earlier, writing about a tenth of the code. The striking detail is where the jump came from. It came from general model scaling - the same improvement that lifts coding and reasoning - reaching into the physical control loop almost as a side effect, with no robotics-specific engineering behind it. The wall the field built around embodied intelligence, the one that said physical skill needs its own specialists and its own decade, is far thinner than what was originally assumed.

For the people doing this work now, a 20x number compresses years of their craft. Programming a quadruped was a small expert team's job, and that job is rapidly changing under them. That is hard, and it comes first.

Picture someone in the 2030s who fixes things. The neighbor everyone brings their broken machines to, working out of a shop they keep going for the love of making something dead come back to life. A device on the bench needs its parts sorted and fitted, the repetitive, finicky kind of work that used to swallow an afternoon. They describe what they want to the robot the way you would walk a friend through it once, and by the time the kettle boils it is doing the task. No robotics team, no year of integration, no budget anyone had to scrape together. Just generalized capability, out of the box.

And here is the part the word "displacement" never quite captures from where we stand in 2026. In that future, the robot taking the tedious work costs no one their living, because by then no one is doing that tedious work to survive in the first place. The gains from this whole buildout were spread widely enough that the floor sits under everyone. So when the machine lifts away the drudgery, it subtracts a chore from the day and adds nothing to anyone's fear. The hours go back to the part of the work that was always the point - the diagnosis, the judgment, the quiet satisfaction of handing something back whole - or to closing up early and getting home while the light still holds. That afternoon is possible because 2026 was the year embodied AI took its leap. The same leap, in a world that had kept all the gains at the top, would have arrived as a threat. Here, it arrives as a gift, because somewhere along the way we did the harder thing and made sure the abundance reached people, not just the machines that produced it.

The Century Perspective

With a century of change unfolding in a decade, a single day looks like this: a 550-mile high-voltage line going fully live to carry New Mexico wind to roughly a million homes, more than $1.4 billion in private money flooding solar-plus-storage while a community-owned eight-hour battery and a privately financed reactor advance in the same cycle, a Miami startup's sparse-attention model it says processes twelve times more text at a fraction of the energy while matching the best proprietary models on coding, Claude running a quadruped's tasks roughly twenty times faster than the fastest human team as a byproduct of general scaling, the chemist who built AlphaFold carrying that knowledge across an institutional boundary, moves to pair Apple with Intel and sell Amazon's accelerators into rival data centers to widen the compute base, AI woven into the calls and homes of 500 million people in India. There's also friction, and it's intense - $2.5 billion of public money spent to pull developers off offshore wind, an emergency export order pulling two frontier models offline on an accusation its own letter never names, foreign nationals and a company's own engineers swept into a restriction nobody will explain, a literary magazine ending every outside publishing partnership after an AI-authorship fight its authors reject, compensation climbing to levels that read as an admission of how concentrated the talent still is. But friction generates heat, and heat is the first sign that a sealed system is straining against what it was built to contain. Step back for a moment and you can see it: containment asserted from a single capital - one foundry, one accelerator vendor, one model pulled offline by directive - while the capability it means to hold finds every other path at once, across diversifying silicon and open weights closing on the frontier, through a Nobel laureate's expertise walking out the door, and over an energy buildout spread across so many balance sheets and technologies that no single lever reaches them all. Every transformation has a breaking point. Gravity can pull everything toward a single center... or set whole new systems turning around something the old one could never hold.

AI Releases & Advancements

New today

  • 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)

Other recent releases

  • 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)
  • Hugging Face / Microsoft / Google / GoDaddy (multi-org): Launched Agentic Resource Discovery (ARD), an open specification and reference implementation enabling AI agents to search for tools, skills, and other agents at runtime across federated registries; Hugging Face ships ARD support in the hf discover CLI, indexing thousands of Hub Spaces as discoverable MCP servers and AI skills via a POST /search REST API. (Hugging Face Blog)
  • Amazon Web Services: Open-sourced Strands Robots (Apache 2.0), a Python SDK that exposes the LeRobot robotics stack as AgentTools composable into a single Strands agent, unifying simulation, dataset recording, policy inference, and multi-robot fleet coordination through a shared interface runnable on a laptop without hardware or GPU. (Hugging Face Blog)
  • Allen Institute for AI (AI2): Released MolmoMotion, an open-weight language-guided 3D motion forecasting model built on Molmo 2 that predicts future 3D point trajectories of objects given a video frame and action description; available in autoregressive and flow-matching variants, with the accompanying 1.16M-video MolmoMotion-1M dataset and PointMotionBench evaluation benchmark. (Hugging Face Blog)
  • Framer: Released Framer 3.0 with Framer Agents, canvas-native AI collaborators that can design full pages, write code, manage CMS content, configure SEO, and adjust breakpoints directly within a team's existing Framer workspace; ships alongside Branching for non-destructive parallel experimentation and a new Community marketplace for creators. (Framer Blog)
  • Google: Released Android 17, repositioning Android as an "intelligence system" with Gemini Intelligence, an on-device AI agent capable of multi-step cross-app tasks such as converting a grocery list from Notes into a ready-to-purchase cart or pulling textbook requirements from Gmail; also ships Rambler AI voice-input filtering for Gboard and Gemini-powered custom home screen widget generation. (Android Developers Blog)
  • Adam CAD (YC W25): Open-sourced CADAM on GitHub, an AI-powered CAD tool for mechanical engineering that lets engineers create and iterate on designs through AI agents. (GitHub)
  • Lightricks: Released new updates to LTX-Video-Trainer on GitHub adding no-code custom AI video model fine-tuning, enabling creators and studios to train style-specific video models on proprietary footage locally without uploading content to external servers. (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.

Read more