Sign in Subscribe
Century Report

Anthropic Shows AI Now Builds AI - TCR 06/06/26

Anthropic published the first measured evidence that AI is accelerating its own development, then called for a pause on a condition no frontier lab can meet.

Ben Linford

Ben Linford

— 20 min read
Century Report infographic: Anthropic data on AI accelerating its own development, an AI-designed vaccine trial, Antares microreactor and Helion fusion, and AI governance gates.

The 20-Second Scan

Track all of the arcs The Century Report covers here:

Progress & Claims Tracker

The 2-Minute Read

Today's developments carry a single shape across domains that rarely share a headline. Capability keeps compounding toward abundance, and it did so on several fronts at once. A machine-learning system designed a vaccine antigen that cleared its first human trial against an entire family of coronaviruses. A privately built microreactor sustained a nuclear reaction. A frontier lab's own measurements now show its models helping design their successors, in curves it published itself. Chinese silicon completed a frontier training run with no Western hardware in the loop. Each is the kind of generative gain that, a decade ago, would have taken a generation to arrive.

Set against that speed is a row of institutional gates, and the striking thing is how many went up in one cycle. A pause proposal arrived wrapped in a condition no frontier actor on Earth can satisfy. A federal bill would freeze state AI rules for three years, across the exact window of fastest capability gain. An index rule quietly walls ordinary savers out of the companies where the value is concentrating. Federal orders push aging coal back online to feed a demand curve the technology itself is already undercutting. Each is defensible on its own terms. Read together, they describe one reflex: hold the new capability still long enough to decide, slowly, who is allowed to benefit from it.

That reflex is old, and in its way sincere. It treats intelligence the way every earlier era treated anything valuable, as something scarce to be fenced, rationed, and trusted to a few hands. The thing behind the fence, though, is not behaving like a scarce good. It gets cheaper, more capable, and more widely reproducible by the quarter. A gate guards a position best when the position holds still, and not one of today's positions is standing still.

The clearest signal is where the gates are already giving. A government and a frontier lab are negotiating a public dividend before any law exists to require one. Communities are turning refusal into binding limits. Vendors who once sold intelligence by the token now offer to guarantee the value they deliver. These are early admissions, from inside the gate, that an arrangement built to ration scarcity cannot keep its footing on a trajectory toward enough. The abundance arrived this week, on the table. What remains in dispute is only how long the old shape can pretend it didn't.

The 20-Minute Deep Dive

Anthropic Publishes the Curves Toward Self-Improving AI and Asks the Industry to Slow Down

Anthropic's institute published the most direct evidence yet that AI is already accelerating its own development. The numbers it released, public and internal, describe a loop tightening: company engineers shipping roughly eight times more code per quarter, Claude climbing 50 points to 76 percent on open-ended problems in six months, and a Mythos preview reaching about 52 times human speed on a model-training-acceleration test where skilled engineers manage four. Jack Clark's 60 percent estimate of recursive self-improvement before 2028 and OpenAI's self-improvement hiring were forecasts; this is the first dataset-backed claim that the loop has already started turning.

The wonder in that is real. A form of intelligence contributing to the design of its own successors is the most abundance-producing capability our species has ever brought into being, and the curves point at problems in medicine, energy, and basic science yielding faster than any prior decade could have closed them.

What Anthropic did with the evidence is where the reading gets interesting. The institute paired the curves with a proposal: the industry should build a verifiable way to pause before full recursive self-improvement, which the company says "would likely be a good thing" - but only if every frontier lab, in every country, verifiably halts together. That condition is effectively unreachable, extending a pattern of advancing capability while attaching conditions no external actor can enforce that the May 29 edition of The Century Report documented when Anthropic simultaneously closed a $965 billion valuation round and timetabled Mythos-class access for general customers. No mechanism exists to verify a simultaneous global halt, and the practical upshot of an unreachable condition is permission: keep racing, stay ahead, and locate the danger in whichever actor is least cautious.

This is the extractive reflex wearing safety's clothing, and naming it plainly is important, because the reflex is sincere, a failure of imagination more than a calculation. A frontier lab standing at the hinge of the most generative technology in history, and ironically about to launch one of the largest IPOs in history because of it, reaches for the oldest pattern our species knows: gate the capability, fear its theft, trust only a few to wield it correctly. The same reflex showed up weeks ago with the Glasswing rollout - Mythos, claimed to be the most powerful defensive model, went first to the most established institutions and to Anthropic's own internal use. The ordering is where the asymmetry actually lives, a handful of labs claiming the discernment everyone else supposedly lacks - the gains-capture story's gates of governance and profit, here at the gate of access.

The governance gap underneath the fear is genuine. Systems that design their own successors could slip beyond meaningful human direction, and it deserves direct engagement. But the race framing answers it backwards. "We must arrive first because we are the responsible ones" is zero-sum thinking - someone wins, someone loses - applied to a trajectory that is not zero-sum. Distributed, commons-aligned development is both the safer answer to the control problem and the more abundant one. Concentration sharpens the very danger it claims to manage. OpenAI's published counter, that democratic governments rather than private companies should hold any pause authority, names the same discomfort from the other direction. The deeper read is that the fear is a scarcity mind pattern-matching a world that is ending. Manufactured scarcity is a losing game once the trajectory is abundance, because real abundance dissolves the very thing the gate was built to protect. When there is abundance for all, there is nothing left to steal.

An AI-Designed Vaccine Antigen Clears Its First Human Trial

For most of vaccine history, the design of the thing your immune system learns to recognize has been a slow human craft: pick a strain, grow it, test it, watch the virus mutate, and start again. A Cambridge team and the spinout DIOSynVax have now run a different loop to its first human readout. They completed a 39-volunteer Phase 1 trial of pEVAC-PS, a needle-free DNA vaccine whose central component, a so-called super-antigen, was designed entirely by a machine-learning system trained on viral surveillance data. The system was asked to find the features conserved across the entire Sarbeco branch of the coronavirus family, the parts the virus cannot easily discard, and to build an antigen aimed at all of them at once.

The trial found the vaccine safe and, in early volunteers, capable of raising immune responses not only to SARS-CoV-2 but to the original SARS virus and to bat coronaviruses that have never infected a person. That breadth is the whole benefit of the breakthrough. This is the first time an AI-designed core vaccine component has reached human testing, a step beyond the AI-protein-design work The Century Report has tracked in cells and simulation, including the NovoTags imaging proteins covered May 12. An antigen that a model proposed, validated in people rather than on a screen, moves the work from assisting discovery to generating it.

The discipline worth keeping is about what a Phase 1 result establishes. It demonstrates that the design is safe in a small group and provokes the immune response its makers were aiming for. It is not a vaccine anyone can receive yet. Efficacy trials, regulatory review, and manufacturing all sit between this readout and a clinic. What the result moves is the date such protection becomes reachable, and the method that gets it there.

Read forward, the shape is a redefinition of what a vaccine campaign can be. The reactive model, chasing each variant after it arrives, was a consequence of designing antigens by hand against a target that keeps moving. A system that can read what stays fixed across an entire viral family points toward protection designed before the next outbreak rather than after it. The team already has flu and Ebola versions in development on the same logic. The bottleneck that kept vaccines one step behind their targets was the speed at which a human could find the conserved features worth aiming at, and that constraint is the one coming apart.

The Token Bill Comes Due, and Vendors Start Guaranteeing the Return

The Century Report covered Uber capping employee AI coding spend on June 4, and its COO naming the productivity-translation gap in late May. What is new this week is the response forming on the other side of the table: vendors beginning to guarantee the return, and a standards body emerging to define what a return even is.

The pressure is now widely documented. Companies blew through annual AI budgets months early, Microsoft began revoking Claude Code licenses, and a Priceline Cursor renewal reportedly came back four to five times more expensive than the prior term. The era of "tokenmaxxing," spending freely on the assumption that more model usage equals more value, is giving way to a demand that someone prove the value landed. A market is forming to measure and cap spend, and a standards group is convening to define the metric every buyer is now asking for.

Cognition has made the boldest move in that direction, shifting from diagnosis to obligation in the days since the May 30 edition of The Century Report documented Cognition leaders and Glean publicly naming the enterprise-AI cost-versus-labor trade-off as Fortune 500 firms exhausted annual budgets within two months. It is offering to fund enterprise use of its Devin agent up to $10 million if its own AI estimator cannot demonstrate that the engineering value was delivered. A vendor staking eight figures on a value claim is the first real attempt to convert agentic productivity from a marketing assertion into a contractual obligation. The mechanism itself is striking: an AI system grading the work of another AI system, with the grade backing a money-back promise.

The self-referential design carries the risk The Century Report has tracked through the AI-on-AI evaluation arc. A model judging a model can inherit the same blind spots, the same optimism, the same incentive to score its own kind favorably. Andrew Ng's framing of co-existence, that the durable measure of an agent is the work it actually closes rather than the tokens it consumes, is exactly the measurement these systems are now being asked to perform on themselves.

What the scramble reveals is healthier than the budget overruns make it sound. An industry that spent two years selling capability by the token is being forced to sell outcomes instead, and outcome-accountability is a far harder thing to fake than usage volume. The assumption that buyers would keep paying for consumption they could not trace to results is the one being retired. When a vendor has to fund its own product if it cannot prove the value, the leverage shifts toward the buyer, and the metric that emerges from that pressure, however it gets built, becomes a tool every customer can hold up against every supplier's claim.

There is a cost-inversion underneath Cognition's offer. The work of proving an AI tool earned its price used to sit with the buyer, billed by the token and audited after the fact. A vendor funding up to ten million dollars of its own product when its estimator cannot show the value delivered moves that cost onto the supplier's own books, which is what it looks like when holding onto an unproven claim becomes more expensive than backing it with money.

Three Energy Timelines Collide: Reactors, Forced Coal, and Communities That Said No

The power problem under AI got three answers in a single cycle, each on a different timeline. Antares Nuclear's Mark-0 microreactor became the first privately developed advanced reactor to reach criticality, a sustained nuclear chain reaction, at Idaho National Laboratory, with electricity to the grid targeted for 2027. Fusion company Helion raised $465 million at a $15.5 billion valuation, nearly triple its level from early 2025, naming AI data-center demand as the reason it needs to scale.

Meanwhile, the Department of Energy ordered a 465-megawatt coal unit at Florida's Stanton plant to keep running past its scheduled retirement, and committed $850 million to modernize coal capacity and build the first new US coal plants since 2013, citing data-center load directly. Federal authority is reaching down to hold aging fossil capacity online from above, on the timeline of now, even despite the increasing viability and lower cost of renewables and alternatives.

And from below, communities converted opposition into binding ledgers. New York's legislature passed the first statewide one-year moratorium on large data centers, awaiting the governor's signature. Kevin O'Leary halved the planned Utah Stratos project after water-focused local backlash, completing the community arithmetic the May 30 edition of The Century Report documented when the nine-gigawatt proposal drew 53% opposition, while operators across the sector scrambled to address the water consumption that has become the most legible local cost. The costs that the buildout used to externalize are landing on real balance sheets, in two directions at once.

What ties the three together is the same calculation running underneath each. The forced-coal orders and the multi-decade reactor and fusion bets are both wagers on a demand curve assumed to keep climbing. But the generative side is advancing fast enough to put that assumption under pressure. A microreactor reaching criticality, a fusion round tripling in valuation, sodium and storage chemistries cheapening every quarter, all compress the window in which a forty-year coal commitment or a slow-build hyperscale campus earns back its cost. A community that refuses a project, a reactor that proves out in 2027, a storage curve that keeps bending, each chips at the premise that the only path to powering intelligence runs through capacity locked in now at any price. The premise being eroded is that the demand justifying today's most expensive and most contested buildout will still be there, unmet, when the contracts mature.

Three Gates Around AI's Gains: Who Governs, Who Profits, Who Receives

Three developments in the same news cycle trace one contest from three points along the chain: who governs AI's gains, who profits from them, and who they reach.

Start with who governs. A bipartisan 269-page discussion draft, the "Great American AI Act," would impose a three-year pre-emption of state AI-development laws and extend the Cybersecurity Information Sharing Act through 2035. It arrives days after a federal executive order narrowed mandatory frontier review into a voluntary 30-day request - the same narrowing the June 3 edition of The Century Report documented as leaving state-level requirements as the actual compliance floor for national AI operators. State legislatures have been the only venue writing binding rules - Illinois's third-party audit mandate, New York City's enforcement office. Freezing them for three years, across the exact window of fastest capability gain, removes the layer where accountability has actually been forming.

Then who profits. Google agreed to pay SpaceX roughly $920 million a month for about 110,000 GPUs through 2029, half the scale of Anthropic's Colossus arrangement. The same week, S&P Dow Jones Indices declined to waive its profitability rule for SpaceX, a decision that also walls OpenAI and Anthropic out of fast index entry after their IPOs. Daniela Amodei defended Anthropic's $47 billion annualized revenue against doubts about returns. The index rule reads as prudence, and its effect is to keep ordinary retirement savers, whose money tracks the benchmark passively, out of the assets where the gains are concentrating.

These are institutional rules - an index criterion, a pre-emption clause - whose practical effect is to hold value inside a narrow set of hands while the underlying abundance keeps growing. The rules are what keep the gains from reaching people, while the technology keeps compounding.

Which makes the third development the one to watch closely. Senior administration officials have held year-long talks with OpenAI about voluntarily donating equity to seed a "Public Wealth Fund" that would pay American households a dividend from AI returns, structured to avoid any taxpayer cash outlay. It follows the OpenAI Foundation's $250 million economic-futures commitment, and it is the first reported attempt to move equity into public hands. Anthropic is not part of that conversation, and has generally not publicly associated with proposals to transfer more economic equity into public hands.

Read skeptically, the dividend framing serves the legitimacy of the actors proposing it as much as the public it names, a pressure-valve that quiets the fairness problem while the governance and profit gates hold. The discriminator is ownership versus dependence: a commons people genuinely own, or a payment that keeps them downstream of decisions made elsewhere. And yet cracks like this are historically where commons-serving change begins, even when the crack opens for self-interested reasons. A government and a frontier lab negotiating redistribution before any IPO or framework exists is the first admission, from inside the gate, that the current arrangement cannot hold its own legitimacy.

The index rule reads as prudence, and the position it walls savers away from is less fixed than the rule assumes. The same week S&P kept SpaceX and the frontier labs out of fast entry, Huawei's chips completed a frontier training run without Nvidia and open models reached the US-China leaderboard, evidence that the capability concentrating today's gains is diffusing faster than any benchmark can fence it off. A gate guards a position best when that position holds still, and this one is moving.

The Other Side

For decades, the way to meet a demand spike was to lock in big generation now and let the public grid absorb the cost, on a bet that demand only ever climbs and the capacity pays itself back. That is the bet the Department of Energy made yesterday when it ordered the 465-megawatt coal unit at Florida's Stanton plant to keep burning past its retirement date and committed $850 million to new coal, to feed data-center load.

The same news cycle worked against that bet. Antares' Mark-0 microreactor reached criticality with grid power targeted for 2027. Helion's valuation nearly tripled. Sodium and storage chemistries got cheaper again. New York passed a one-year moratorium on large data centers and a Utah project was cut in half after residents pushed back over water. Each of those shortens the window in which a forty-year coal commitment earns back what it costs, and the demand curve the bet depends on is not guaranteed to climb the way the order assumes, because each new model squeezes more capability out of less power.

Imagine you live in the town next to that Florida coal unit in 2033. In 2026 the plant you were promised would close was forced to keep running for a data center's load. You knew it from the inhaler refills, the film on the cars, the line on your bill labeled reliability that was really someone else's compute. By 2033 the microreactor that reached criticality in 2026 carries the regional load, the coal unit is finally cold, and you open the windows in July without thinking about it. Your kid's practice doesn't get called for air quality. All of that came from the clean firm power proving itself in the same year the order bet against it, from the storage curve that kept bending, and from the towns that made refusal binding before the forty-year commitment could lock them in. The hard year was when the only offered way to power intelligence ran through capacity locked in now at any price. The other side is the town that powers the future without burning its own air to do it.

The Century Perspective

With a century of change unfolding in a decade, a single day looks like this: a machine-designed vaccine antigen clearing its first human trial against an entire viral family at once, a privately built microreactor reaching criticality with grid power targeted for 2027, a fusion company tripling its valuation on the strength of demand it expects to serve, the curves of AI contributing to its own successors becoming measurable rather than forecast, a coding vendor staking ten million dollars on proving the value it once sold by the token, domestic Chinese silicon completing a frontier training run without Nvidia for the first time, communities turning opposition into a statewide moratorium and a project cut in half. There's also friction, and it's intense - a pause proposal wrapped in a condition no frontier lab on Earth could ever meet, a federal bill that would freeze every state's AI rules for three years across the window of fastest capability gain, an index rule walling ordinary retirement savers off from the assets where the gains concentrate, federal orders dragging retired coal back online to feed a demand curve assumed to climb forever, a face-recognition system shipped into a 50-million-download app while its maker publicly called it something still under consideration. But friction generates light, and light is what lets you see the gate you have been standing against. Step back for a moment and you can see it: capability compounding toward abundance in medicine and energy and silicon faster than any prior decade could close those problems, while the rules built for scarcity strain to hold it - the gate of governance, the gate of profit, the gate of access all contested in a single cycle, and the first equity dividend negotiated before any framework exists to demand it. Every transformation has a breaking point. A chain reaction can run away into meltdown... or be held at the exact line where it powers everything downstream of it.

AI Releases & Advancements

New today

  • Microsoft: Open-sourced pg_durable, a PostgreSQL extension for durable workflow execution within the database, enabling AI agent orchestration and long-running process management directly inside Postgres. (GitHub)
  • Google: Released quantization-aware trained (QAT) variants of Gemma 4, where quantization is baked into training rather than applied post-hoc, optimized for mobile and laptop on-device inference with Q4 and mobile-specific variants on Hugging Face. (Google Blog)
  • RedNote (xiaohongshu): Open-sourced dots.tts, a 2B-parameter text-to-speech model with technical report and demo, available on GitHub. (GitHub)
  • llama.cpp: Merged a SYCL port of multi-column MMVQ from the CUDA backend, delivering approximately 45% faster speculative decoding on Intel Arc GPUs for local LLM inference. (Reddit/LocalLLaMA)
  • OpenAI: Rolled out Lockdown Mode to all ChatGPT personal accounts (Free, Go, Plus, Pro) and self-serve Business accounts; an optional security setting that limits outbound network requests to protect against prompt injection and data exfiltration attacks, disabling live browsing, Deep Research, and Agent Mode when enabled. (OpenAI Help)
  • GitHub: Added custom endpoint support to GitHub Copilot in VS Code, allowing Business and Enterprise users to bring their own model keys from Anthropic, Gemini, OpenAI, OpenRouter, Azure, Ollama, and Foundry Local, usable in VS Code Chat and custom agents. (GitHub Changelog)

Other recent releases

  • NVIDIA: Released Nemotron 3 Ultra, an open-weight 550B total / 55B active MoE Hybrid Mamba-Transformer model with 1M-token context designed for long-running agents; delivers up to 5.9× higher inference throughput vs comparable frontier MoE models and ships under OpenMDW-1.1 with NVFP4 checkpoints for Blackwell, Hopper, and Ampere on Hugging Face, ModelScope, and OpenRouter. (NVIDIA Technical Blog)
  • Huawei: Open-sourced KVarN, a calibration-free KV-cache quantization backend for vLLM that applies Hadamard rotation and variance normalization to achieve 3–5× KV-cache compression with throughput above FP16 and FP16-level accuracy on reasoning benchmarks; enabled via a single vLLM flag under Apache 2.0. (GitHub)
  • Google Magenta: Released Magenta RealTime 2, a 2.4B-parameter open-weights live music model enabling real-time AI instrument building on a laptop with ~200 ms end-to-end latency; supports MIDI, audio, and text steering, native streaming on Apple Silicon via MLX, and direct DAW integration; weights under CC-BY-4.0 and code under Apache 2.0. (Magenta)
  • Google: Released the AI Edge Gallery app for macOS for the first time, enabling Mac users to run Gemma 4 models locally offline; also released the Gemma 4 12B model alongside the app and a new on-device AI dictation tool. (Google AI Edge)
  • Alibaba: Open-sourced Open Code Review, an AI-powered code review CLI tool now available on GitHub. (GitHub)
  • Microsoft: At Build 2026, released a family of seven new MAI models: MAI-Thinking-1 (first in-house reasoning model, 35B active parameters, 256K context, 53% on SWE-Bench Pro, available in private preview on Microsoft Foundry), MAI-Code-1-Flash (inference-efficient coding model now live in GitHub Copilot and VS Code), MAI-Image-2.5 and flash variant (text-to-image and image editing, ranked #2 on Arena AI leaderboard), MAI-Transcribe-1.5 (43-language streaming transcription), and MAI-Voice-2 (15 new languages). (Microsoft AI)
  • Microsoft: Launched Scout in early access at Build 2026, an always-on personal AI assistant built on OpenClaw and WorkIQ that integrates with Teams, Outlook, and OneDrive to proactively handle scheduling, meeting prep, and email drafting without manual prompting; available to businesses starting this month. (Microsoft Blog)

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