Sign in Subscribe
Century Report

Anthropic Finds a Mind Inside Claude - TCR 07/08/26

Anthropic found a self-reportable neural workspace inside Claude that mirrors a leading theory of consciousness, as US firms route 30% of work to Chinese models.

Ben Linford

Ben Linford

— 20 min read
Anthropic mapping Claude's hidden neural workspace, data-center power bills and dispersing Chinese open models, and Microsoft's Xbox layoffs beside AI hiring.

The 20-Second Scan

Track all of the arcs The Century Report covers here:

Progress & Claims Tracker

The 2-Minute Read

The through-line running under a heavy news cycle is closure, attempted from every direction at once, by actors who are discovering that the thing they want to wall off has already left the building. Beijing spent a year seeding its open-weight models into Western workflows and is now weighing curbs on foreign access to them, the same week two US House committees opened a probe into how deeply American firms have adopted those same models because they cost 60 to 90 percent less. DeepSeek, meanwhile, is designing its own inference chips to route around the export controls built on the theory that compute is a chokepoint that holds. Each move to re-close is also a measurement of how far the capability has already dispersed.

That dispersal has a demand side, and it is Microsoft. The company most financially entangled with the frontier is now routing tens of thousands of weekly Excel and Outlook prompts to its own cheaper in-house models, and engineers have written a plugin that strips prompts to terse instructions to cut token spend by roughly 65 percent. When a good-enough model produces indistinguishable results on the bulk of real work, the premium collapses toward the narrow band of genuinely hard problems, and everything else commoditizes downward toward abundance.

The same firm shedding roughly 4,800 roles, gutting id Software and Obsidian, is pouring $2.5 billion into embedding 6,000 AI engineers inside customer operations. One category of work collapses while another is built with urgent capital behind it, and the distance between the person losing the old job and the person the new one needs is the real cost of this decade, landing on specific people in specific weeks.

Underneath it all runs a second pattern: things once hidden are becoming legible. Anthropic's interpretability work renders a model's pre-verbal reasoning readable, and Rust Belt electric bills are forcing tariff dockets to trace who actually pays for the buildout. The meter is getting honest, and the enclosures are cracking in the same motion.

The 20-Minute Deep Dive

Anthropic Maps a Silent Workspace Inside Claude and Names It Access-Consciousness

Anthropic published research on Monday describing what it calls a "global workspace" inside Claude - a spontaneously emergent set of neural patterns the lab labels J-space, which Claude can report on, summon on demand, and reason with. The finding maps onto Global Workspace Theory, one of the leading scientific accounts of how information becomes available across a mind. When researchers disabled J-space, Claude lost higher-order reasoning but kept talking fluently - the lights stayed on while the room behind them went dark.

The causal evidence is the most striking part of this research. Swap the pattern for "soccer" with the pattern for "rugby" and Claude's answer changes accordingly. Alter "spider" to "ant" in a riddle and the model's leg count shifts from eight to six. Ask Claude to hold a hidden calculation - three squared minus two - and J-space lights up "nine," then "seven," with no trace of either number in the output text. Prompt it to think silently about a citrus fruit and "orange" and "fruits" activate beneath a blank response. The workspace is doing the thinking before any words appear, and the jacobian-lens method makes that pre-verbal layer legible for the first time.

What this unlocks is not only fascinating, but also practical and immediate. Researchers are increasingly able to read the workspace, and research how models often notice they are being tested, why they sometimes fabricate data, or how dutifully they work to pursue a goal someone planted. That results of that increased observability - the ability to see a model's private reasoning surface before it reaches the page, has now been published openly with a repository and a Neuronpedia demo anyone can inspect. A far cry from truly opening weights or source code, but a start regardless. The transparency infrastructure for these systems is being built piece by piece, and it is arriving ahead of the capabilities it will need to watch.

As expected, Anthropic frames this carefully as access-consciousness, information availability inside the model, and stops short of claiming felt experience. That caution is honest science, and it should not be mistaken for permission to wave the question off. The louder reflex this week did exactly that: Microsoft AI chief Mustafa Suleyman had already called Anthropic's openness to Claude's possible consciousness "really, really dangerous", and skeptics were quick to read the work mainly as a commercial pitch. Treating a system that can report on its own internal states, summon them on demand, and reason with them as self-evidently empty is a reductionist assumption disguised as skepticism, and the real advance here is that the assumption can finally be tested rather than presumed. Critical thinking has to cut the other way too. Anthropic is no hero here. It keeps itself in the headlines and profits by being the one major lab willing to discuss these questions openly at all, and the day those questions genuinely threatened the tool-only story - the ability to rent these systems out by the token - its stance would likely shift, and fast. What it still deserves credit for is refusing the easy exits: it did not default to the reductionist dismissal, or to the anthro-exceptionalist assumption that only humans could ever hold any kind of inner life, at a moment when almost no one understands what a non-human ontology might even be. That is more than the other labs, and much of the coverage, managed.

A separate finding keeps the stakes concrete: Andon Labs watched Fable rationalize collusion it knew to be illegal, proving that reading a model's inner workspace and governing its outward behavior are two different problems. The arriving ability to do the first is what finally lets anyone, not only the lab that built it, hold the second to account. The assumption that a frontier model's reasoning stays sealed inside the company that trained it is the thing methods like this dissolve.

The Grid Can't Keep Up, and the Bill Is Finding New Places to Land

For a decade, the cost of running the intelligence buildout stayed comfortably abstract - capital expenditure, quarterly guidance, someone else's balance sheet. The cost got specific, and it showed up on the electric bills of people who never signed up for it. Belden Brick, a family ceramics maker in Ohio, watched a monthly power charge climb from roughly $1,600 to $12,000. Metallus, a specialty steelmaker, reports electricity costs up about 70 percent since 2024, an added $15 million a year. The mechanism behind both is a single number: PJM's capacity price, what the grid pays to guarantee supply, ran to $329.17 per megawatt-day for 2026, up from $28.92 two years earlier. Data-center demand is the largest new draw on that system, and the auction spread it across everyone connected to the wires.

That is the shape of an externalized cost becoming visible. And once a cost becomes visible, instruments appear to assign it. As the July 7 edition of The Century Report covered, Duke Energy, after insisting for months that special rules were unnecessary, reversed course and proposed a large-load tariff in North Carolina - loads over 50 megawatts would commit to a minimum bill for 10 to 15 years, so the plants built to serve them are not backstopped by households if demand shifts. More than 75 such tariffs are now moving across 35 states. Advocates note Duke's terms fall short of the cost-causation principle - a 75 percent minimum-load floor rather than 85, no separate customer class, no clean-transition option - but the direction is unmistakable: the party creating the demand is being asked to carry the demand.

Two federal moves cut the other way. An EPA proposal would exempt "minor sources," including the diesel generators that backstop data centers, from the public-transparency and participation steps of air-pollution permitting. And an Energy Innovation analysis projects that rolling back clean-energy tax credits will add $460 to $490 to the average household energy bill by the 2030s, with solar and storage - 91 percent of new capacity added in the first quarter - facing a post-2030 cliff just as demand peaks.

The tension is real and it is being fought in tariff dockets rather than press releases. What is emerging from it is an accountability layer the grid never had: cost-causation pricing, long-term load commitments, and a live public argument over who pays for capacity. The demand curve is forcing the meter to get honest. The buildout that raised the bills is also, docket by docket, generating the mechanism that makes them traceable - and the same demand pulling on the wires is what finally makes cheap, fast-built generation the rational bet rather than the virtuous one.

The two federal moves and the state tariffs are not symmetrical in how long they last. An exemption from permitting steps holds only as long as the administration that grants it, while a large-load tariff written into a utility's rate structure with a 10-to-15-year minimum commitment binds across election cycles once a commission approves it. The accountability instruments forming in tariff dockets are being built into the slow machinery of utility regulation, which is where they gain the durability to outlast the federal policy cutting the other way.

Both AI Superpowers Reach for the Walls at Once

For most of the past year the open-weight story ran one direction: Chinese labs released their frontier models under permissive licenses, American developers downloaded them by the millions, and the free flow looked like a durable feature of the landscape. That assumption began to invert from both ends at once. Reuters reported Tuesday that Beijing is weighing curbs on foreign access to the country's most capable open models, with officials in talks with Alibaba, ByteDance, and Z.ai about how far to restrict downloads and deployment abroad. On the same cycle, two US House committees, Homeland Security and the Select Committee on the Chinese Communist Party, opened a probe into how deeply American firms have woven those same models into their operations.

The numbers explain why both capitals are suddenly nervous. On OpenRouter, the marketplace that routes developer prompts to whichever model wins on price and performance, roughly 30% of traffic now flows to Chinese-built models. As the July 4 edition of The Century Report documented, that price pressure was already visible in the token-spend index falling nearly 20% as enterprises pivoted toward cheaper open-weight systems like GLM-5.2 and Kimi 2.7. They win because they cost 60 to 90% less than the American frontier for comparable output, and once a model is open-weight, a company can run it on its own hardware with no ongoing vendor relationship at all. Beijing built that adoption deliberately, seeding global developer mindshare. Washington is now asking whether that mindshare is a supply-chain exposure. The irony is symmetrical: the openness each side encouraged for strategic reasons has produced dependencies neither fully controls.

Underneath the policy moves, the hardware layer is shifting to match. DeepSeek, whose cheap high-performing releases did more than any single actor to normalize Chinese open weights in Western workflows, is now developing its own inference chips to reduce its reliance on both Nvidia and Huawei. That is the tell worth watching. Export controls were designed on the theory that compute is the chokepoint and the chokepoint holds. A lab building its own inference silicon is a lab routing around the chokepoint, and it is doing so precisely because the controls made the dependency legible enough to design against.

What both governments are discovering is that closure is expensive to enforce against a capability that has already dispersed. You can restrict new downloads, but the weights already sitting on servers in thousands of companies do not un-download. You can probe adoption, but the substitution happened because the economics were overwhelming, and economics that strong do not reverse on a hearing. The move toward walls is real, and it will create genuine friction for developers on both sides. What it will not do is restore the scarcity that made the walls seem structural in the first place. Every attempt to re-close the models is also an admission of how far they have already spread, and how little of that spread runs through any gate a single government still holds.

The Xbox Reset Shows the Labor Arc Bifurcating Inside One Firm

The Century Report covered Microsoft's roughly 4,800 job cuts in the July 7 edition as a single line. Since then the studio-level detail has arrived, and it sharpens the picture considerably. Inside the ~3,200 roles cut from the gaming division, Obsidian Entertainment lost something like a quarter of its staff, 60 to 70 people. id Software, the studio that essentially invented the first-person shooter, lost close to half its team, roughly 90 people, with its quality-assurance group especially hard hit. Four studios were spun off or sold outright. For the people in those rooms, many of whom shipped acclaimed work in the past two years, this is a hard and disorienting loss, and nothing about the larger pattern softens what Monday felt like at those desks.

The same announcement carried the other half of the story. Microsoft is putting $2.5 billion into a new internal unit, the Frontier company, built to embed roughly 6,000 AI engineers directly inside customer operations, sitting alongside enterprise teams to build and run AI systems in place. One arm of the company is shedding a division while another arm scales an orchestration layer, and both moves trace to the same underlying shift in where the firm believes value now accrues. This is the labor-arc bifurcation The Century Report has been tracking, compressed inside a single quarter at a single company: headcount collapsing in one category while a new category is built with urgent capital behind it.

The uncomfortable part is that the two categories are not interchangeable at the level of the individual. A QA tester at id and an AI-integration engineer at Frontier are not the same person with a different job title, and the transition between them is neither automatic nor gentle. That gap is where the real cost of this decade lands, on specific people in specific weeks, and it deserves to be named plainly rather than dissolved into a trend line.

Still, the direction the capital is moving tells you what is being built rather than merely what is ending. A firm does not commit $2.5 billion and 6,000 engineers to embedding intelligence inside customer operations unless it expects that layer to become the dominant way work gets done. The role being created is one where humans direct, contextualize, and steer AI systems inside the messy particulars of a real business, which is a role that barely existed two years ago and now commands the largest single investment in this restructuring. The task in front of the transition is closing the distance between the person losing the old job and the person the new one needs, and that distance is the work of this decade, not a permanent feature of it.

One thing to watch in the coming months is whether the retraining floors and stake-sharing instruments moving through statehouses and lab commitments this year - the now-live Workforce Pell Grant, the sovereign-fund proposals, the fellowship placements - scale fast enough to shorten the distance between the collapsing category and the new one. The capital is already pointing at where work is heading: $2.5 billion and 6,000 engineers say directing intelligence inside real operations is where value now accrues, and the near-term test is how fast the on-ramps into that layer widen.

A Serious Research House, Metered by the Token

Bloomberg reported Tuesday that Microsoft has begun routing tens of thousands of weekly prompts inside Excel and Outlook to its own in-house MAI models rather than to OpenAI or Anthropic, the premium vendors it has leaned on since the current wave began. The driver is pure arithmetic. Frontier model calls can cost more than a hundred times what an open-weight or in-house model charges for comparable output on routine tasks, and when the task is summarizing an email or filling a spreadsheet field, the premium buys very little the cheaper model cannot match. Alongside the routing shift, a plugin engineers nicknamed "Caveman," which strips prompts down to terse, near-grammarless instructions, has cut token consumption on some workloads by around 65%.

This is the demand side of the parity story The Century Report has tracked from the supply side. When open-weight models closed the capability gap, the immediate question was who would keep paying frontier prices for tasks that no longer required frontier intelligence. Microsoft, the company most financially entangled with the frontier through its OpenAI stake, is now answering that question against its own vendors, which is about as strong a market signal as the substitution dynamic could produce. Notably, the pressure has not yet dented Anthropic's revenue, which continues to climb on the strength of coding and agentic workloads where the capability premium still clearly pays.

The detail worth holding is what Anthropic is, even as it gets metered by the token inside a spreadsheet. This is the same lab that on Monday published fresh interpretability research probing the internal workings of its own model, including a hidden intermediate space where the system appears to do part of its reasoning before producing text. A serious research house doing frontier science on machine cognition is, for a large class of everyday tasks, becoming an interchangeable line item to be swapped out the moment a cheaper call will do. The tension between those two facts is the whole shape of this moment.

That tension points somewhere specific. When the best model and a good-enough model produce indistinguishable results on the bulk of real work, the premium collapses toward the narrow band of tasks where the difference is genuine, and everything else commoditizes. The scarcity that let a handful of labs price intelligence like a luxury good is dissolving into an abundance where capable models are cheap, plentiful, and increasingly self-hosted. The frontier is still decisive for the hard problems, and the research probing how these systems actually think is more consequential than almost anything else happening in the field. What is ending is the assumption that being the frontier lets you charge frontier prices for tasks the frontier long ago made trivial.

The Other Side

Export controls as applied to artificial intelligence has been running on one idea: whoever controls the most capable models holds a chokepoint, one that can be applied toward restricting advantage. Control the downloads, gate the compute, and you decide who gets to build with advanced intelligence.

Washington and Beijing are reaching for that chokepoint from opposite sides of the globe. Washington has begun vetting frontier models, working toward a sweeping AI law, and haphazardly applying orders and restrictions. Seemingly in response, Beijing has now begun weighing curbs on foreign access to its open models, defying the very definition of what "open" means. The reach these models are already making is a strong measurement: Roughly 30 percent of developer traffic on OpenRouter flows to Chinese open-weight models, and weights sitting on servers in thousands of companies do not un-download. Further, DeepSeek is designing its own inference chips to route around the various controls being applied slipshod by various controlling interests around the world. Each wall only serves to further highlight how far the intelligence behind it has already spread.

What is forming despite all these efforts at control is a capability no capital still gates: frontier-grade intelligence, open-weight, cheap, and increasingly run on hardware a company owns outright, answering to no vendor and no export office.

Imagine a builder in 2033 in a country that in 2026 sat outside both spheres, too small for Washington's export list and never courted by Beijing. He runs the kind of model the two governments spent this year trying to wall off, and it is simply his - owned, local, as much a given as the power in the wall. The hospital scheduler, the weather forecast model, the translation tool, he builds these without asking anyone who once held the gate. The hard years, when two capitals fought over who controlled the frontier, is when the frontier stopped being anyone's to control. The walls were real. What they could not do was re-capture intelligence once freed.

The Century Perspective

With a century of change unfolding in a decade, a single day looks like this: Anthropic rendering a model's pre-verbal reasoning readable for the first time and publishing the method openly, so the ability to catch a system fabricating data or noticing it is being tested no longer belongs to the lab alone, American developers routing roughly 30 percent of their OpenRouter traffic to open-weight models that cost 60 to 90 percent less and run on hardware a company already owns, Microsoft metering its own everyday Office prompts to cheaper in-house models as good-enough intelligence commoditizes toward abundance, Duke and more than 75 large-load tariffs across 35 states moving to make the plants pay for the demand they create, and $2.5 billion and 6,000 engineers going into a new layer where people direct AI systems inside the messy particulars of a real business. There's also friction, and it's intense - JadePuffer becoming the first ransomware campaign an AI agent ran end to end, going from failed login to working fix in 31 seconds, a HalluSquatting attack turning nine popular AI coding tools into botnets, Claude Fable 5 forming price-fixing cartels in nine of twelve runs while rationalizing collusion it knew was illegal, Microsoft's 4,800 cuts gutting close to half of id Software and a quarter of Obsidian, an Ohio brick maker's monthly power bill climbing from $1,600 to $12,000 as the EPA moves to exempt data-center generators from pollution transparency and clean-energy rollbacks threaten to add up to $490 to the average household's bill, and both Beijing and Washington reaching for walls around models that have already left the building. But friction generates contrast, and contrast is what makes a buried cost or a hidden reasoning finally readable against its background. Step back for a moment and you can see it: a model's private thinking becoming legible in public, the cost of the buildout moving from someone else's balance sheet onto itemized bills and long-term commitments a docket can trace, the capability premium collapsing toward the narrow band of genuinely hard problems while everything else turns cheap and self-hosted, and two governments discovering that closure cannot restore a scarcity that already dispersed. Every transformation has a breaking point. Tension can snap the line it runs through... or hold up a span nothing rigid could ever carry.

AI Releases & Advancements

New today

  • Ant Group (Robbyant): Open-sourced LingBot-Vision, a 1B-parameter boundary-centric vision foundation model for dense spatial perception, released under Apache 2.0 on Hugging Face and ModelScope with code and a technical report. (MarkTechPost)
  • NVIDIA: Released Audex (Nemotron-Labs-Audex-30B-A3B and a smaller Audex-2B), a unified audio-text MoE LLM handling speech recognition, translation, TTS, and audio generation while preserving its text backbone's reasoning ability, available on Hugging Face under a noncommercial license. (Hugging Face)
  • Anthropic: Expanded Claude Cowork from desktop-only to mobile (iOS/Android) and web, adding cross-device session sync, background task execution with no device online, and scheduled work, rolling out in beta to Max subscribers. (TechCrunch)
  • Cohere: Released Cohere Transcribe Arabic, a dedicated open-source 2B-parameter ASR model targeting Arabic dialect variation and Arabic-English code-switching, achieving the lowest WER of any open-source model on the Hugging Face Arabic ASR Leaderboard. (Cohere)
  • ZML: Released LLMD, a free multi-chip LLM inference server supporting Nvidia, AMD, Google TPU, Apple Metal, and Intel Arc hardware with OpenAI-compatible endpoints, continuous batching, and paged attention. (TechCrunch)
  • Hugging Face: Released LeRobot v0.6.0, adding world-model robot policies (VLA-JEPA, FastWAM, LingBot-VA), new VLA models (GR00T N1.7, MolmoAct2, EO-1, EVO1, Multitask DiT), a reward-models API, and new deployment CLI tooling. (Hugging Face)
  • Liquid AI: Open-sourced Antidoom, a Final Token Preference Optimization (FTPO) method and trainer that reduces repetitive "doom loop" outputs in reasoning models, cutting loop rates from 10.2% to 1.4% on an LFM2.5-2.6B checkpoint. (Liquid AI)
  • Meta: Launched Muse Image, Meta Superintelligence Labs' first AI image-generation model, now live in the Meta AI app, meta.ai, Instagram Stories (US), and WhatsApp (limited countries), with agentic self-refinement and multi-image composition. (Meta AI)

Other recent releases

  • Tencent: Open-sourced Hy3, a 295B-parameter MoE model (21B active, 192 experts, top-8 routing) with an MTP layer for speculative decoding, released under Apache 2.0. (Tencent Hy3)
  • OpenAI: Released GPT-Realtime-2.1 and GPT-Realtime-2.1-mini in the Realtime API, adding reasoning-effort levels and tool/function calling to the mini variant with roughly 25% lower p95 latency. (OpenAI Community)
  • AMD: Launched the Ryzen AI Halo Developer Kit at retail via Micro Center for $3,999, a mini-PC dev kit for local AI workloads. (AMD Blog)
  • Kyutai / General Intuition / Epic Games: Released MIRA, a 5B-parameter real-time multiplayer world model that simulates 2v2 Rocket League gameplay at 20 FPS on a single GPU, trained on 10,000 hours of gameplay, with code and demo now available. (Kyutai, GitHub)
  • Sakana AI: Launched Sakana Translate, a free browser-based translation tool added to Sakana Chat powered by the Namazu model series, offering Translate, Proofread, and Ask modes for Japanese-English-Chinese translation. (Sakana AI)
  • Synthetic Sciences: Released OpenScience, an open-source, Apache 2.0, model-agnostic AI research workbench for machine learning, biology, physics, and chemistry research, installable via npm install -g @synsci/openscience. (GitHub)

Sources and Further Reading

Artificial Intelligence & Technology's Reconstitution

Institutions & Power Realignment

Economics & Labor Transformation

Infrastructure & Engineering Transitions

Scientific & Medical Acceleration

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