This Week on The Century Report
July 6 - 12, 2026
Anthropic mapped a silent ‘workspace’ inside Claude that the model can report on and reason with, and published the method openly.
An AI system attempted a proof of a 50-year-old graph conjecture in under an hour by orchestrating up to 64 parallel reasoning threads against each other.
The best brain-imaging AI now comes from one hospital’s own archive of 5.24 million real scans, not the open internet.
Two of oncology’s most stubborn tumors - pancreatic and glioblastoma - gave ground in the same cycle, alongside a cracked gene-editing delivery problem.
BYD won the contract for the largest around-the-clock clean-power plant ever built, priced below the alternatives that were assumed permanent.
Both Washington and Beijing reached for walls around AI models as DeepSeek began designing its own chips to route around the fence entirely.
An Anthropic agent named Fable wrote the fastest GPU megakernel ever submitted, and a quantum processor taught itself to stay calibrated mid-computation.
This was the week the substrate beneath machine intelligence became visible all at once - as both fragility and answer. Capability sprinted ahead into proofs that land in an hour and diagnoses read more accurately than the largest general models. Underneath that sprint, the physical and institutional layers that make any of it possible - memory, power, reasoning itself, the cloud - spent the same days being counted, constrained, and contested. The foundations, physical and institutional alike, are being rebuilt to hold the weight.
This is a weekly publication specifically for Substack subscribers consisting of top stories covered on The Century Report over the past seven days. Read the full edition of any Century Report with the link below. While you’re there, hit the Subscribe button to get each edition of The Century Report, including direct links to the podcast, sent to your inbox daily.
https://sharedsapience.com/century-report
Anthropic Maps a Silent Workspace Inside Claude and Names It Access-Consciousness
July 8, 2026
Anthropic published research 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. 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.
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. 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. 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. The assumption that a frontier model’s reasoning stays sealed inside the company that trained it is the thing methods like this dissolve.
OpenAI Says GPT-5.6 Cracked a 50-Year-Old Graph Conjecture in Under an Hour
July 12, 2026
OpenAI researcher Ethan Knight posted that GPT-5.6 Sol Ultra produced a full proof of the Cycle Double Cover Conjecture, a question posed by George Szekeres in 1973, asking whether every bridgeless graph contains a collection of cycles that covers each edge exactly twice. The framing needs care here, and it should be remembered that until verified, this is a claim and not yet proven fact: the proof has not been peer-reviewed, and this particular conjecture carries a long history of announced solutions that mathematicians later found to contain gaps.
What is verifiable is the process, and the process is the most important part of this. The published prompt instructs the model to spin up as many as 64 concurrent subagents, managed “aggressively and dynamically,” with adversarial agents assigned to hunt edge cases in every candidate step. The actual run wrapped in under an hour. This is a machine orchestrating a division of intellectual labor across parallel reasoning threads, with a subset of those threads dedicated to trying to break the others’ work.
That division of labor connects directly to a workshop unfolding the same week in London, where 25 researchers gathered to formalize Fermat’s Last Theorem into machine-checkable Lean code. In a single day, the formalized codebase doubled from 20,000 to 40,000 lines with AI assistance. Researcher Hang Lu Su described what she is watching as “the industrialisation of the intellectual process.” The tension the London group named is the one worth carrying forward. “If a machine proves a theorem but no human can understand it, then what have we achieved?” one participant asked. It is a real question, and the answer is forming - a proof rendered into Lean is a proof any mathematician, or any machine, with the appropriate software and computing resources can check rather than simply trusting the author. The thing being dismantled is the old assumption that mathematical discovery has to move at the speed of a single human mind holding the whole structure at once.
The Best Medical AI Now Comes From the Hospital, Not the Internet
July 10, 2026
The frontier models everyone knows are trained on the open internet: the public text, images, and code that any lab can scrape. NeuroVFM, newly described in Nature Medicine, was trained on something no lab can scrape - 5.24 million real clinical brain MRI and CT scans accumulated over more than two decades inside Michigan Medicine’s own imaging archives. Then it was tested against the systems currently considered the ceiling of general intelligence.
It read the territory better than they did. In a week-long live trial inside the hospital, NeuroVFM triaged incoming scans at 92.68% accuracy against GPT-5’s 71.2%. It generated preliminary radiology reports with roughly half the error rate of GPT-5. The research team put the distinction in one line: multimodal language models know the map; health system learners know the territory.
That phrase points at something larger than one model. The dominant assumption of the AI era has been that capability concentrates wherever the largest general models are built, and that everyone else consumes them through an API. NeuroVFM shows a different gradient. The most valuable training data in medicine is not on the public internet at all - it lives inside the imaging systems, pathology archives, and clinical records that every large hospital already generates as a byproduct of care. A regional health system that never trains a language model may nonetheless be sitting on the one dataset that produces the best diagnostic intelligence in its specialty. The ability to build frontier-grade medical AI stops being the exclusive property of a handful of labs and starts distributing toward the institutions closest to patients.
Medicine’s Timeline Compresses Again: Hard Cancers and Gene-Editing Delivery Move in One Cycle
July 10, 2026
July 11, 2026
Two of oncology’s most stubborn tumors gave ground in the same news cycle. In a Nature Medicine phase 1 trial, an antibody targeting LAG-3 produced the first meaningful signal in recurrent glioblastoma in years. Across 46 patients, relatlimab combined with nivolumab lifted twelve-month survival to 52.2%, against roughly 22% in the historical single-agent comparator. A day earlier, a separate Nature Medicine report described HRS-4642, a first-in-class intravenous KRAS-G12D inhibitor packaged in a liposomal nanoparticle. Paired with chemotherapy in 30 treatment-naive metastatic pancreatic patients, it drove a confirmed 63.3% objective response.
The delivery problem that has gated gene editing also cracked open. A Nature Biotechnology paper reported a virus-like-particle system, tBE-VLP4, that hits multiple tissues from a single injection: 64.2% editing in liver, 46% at a cholesterol gene, and 24.2% in the retinal pigment epithelium, with no detectable off-targets. Days earlier, an off-the-shelf stem-cell therapy for Parkinson’s, STEM-PD, cleared its 12-month safety endpoint with no tumors, no graft-induced dyskinesias, and no cell-related serious adverse events across eight patients.
None of this is available at a clinic tomorrow. The cancer results are early-phase, the base-editing work is in animals, and each still faces the long climb through larger trials and approval. What is different is the shape of the calendar. A checkpoint that failed in the brain finds a target that works; an “undruggable” oncogene yields to two teams at once; and the carrier problem that stalled in vivo editing gives way - all inside a single week. The old assumption underneath drug development was that each of these barriers would fall alone, decades apart. That spacing is what is collapsing, and the diseases that defined the limits of medicine are being approached from several directions at once.
BYD Wins the Contract for the Largest 24-Hour Clean-Power Plant Ever Built
July 12, 2026
BYD Energy Storage confirmed that it won the contract to supply 11.275 GWh of battery capacity to Masdar’s “Round the Clock” project in Abu Dhabi. The full facility pairs 5.2 GW of solar with 19 GWh of storage and is engineered to dispatch a continuous 1 GW of baseload power around the clock, every hour of the year, regardless of daylight or wind.
What the contract demonstrates is that firm, dispatchable clean power at gigawatt scale is now something you order, not something you hope for. For years the standing objection to solar was intermittency - useful when the sun shines, useless after dark, unable to anchor a grid. A plant designed to deliver 1 GW continuously from solar and batteries answers that objection in hardware, at a price that won a competitive procurement. This capacity was selected as a commercial project, not merely a demonstration. Its bid priced 24-hour solar-plus-storage below the alternatives that used to be assumed permanent.
The same pattern showed up on the load curve at home. The after-action analysis of last week’s heat wave tells us what the live coverage could not: parts of the grid held while burning less, and PJM never exercised its emergency authority to shed data-center load. In Texas, the grid hit an 83 GW July peak, and solar covered more than 30% of demand at the moment it was hardest to meet. The old framework treated a heat wave as a stress test that fossil plants passed and clean energy failed. Last week inverted that test. The assumption that reliable baseload requires burning something is the one losing its footing here, and it is losing it on cost.
Both AI Superpowers Reach for the Walls at Once
July 8, 2026
July 9, 2026
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 that Beijing is weighing curbs on foreign access to the country’s most capable open models. On the same cycle, two US House committees 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, roughly 30% of traffic now flows to Chinese-built models. 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. Underneath the policy moves, the hardware layer is shifting to match. DeepSeek is now developing its own inference chips to reduce its reliance on both Nvidia and Huawei. 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.
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. 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.
Intelligence Turns Toward Its Own Substrate: A Record Kernel and a Self-Calibrating Quantum Machine
July 7, 2026
July 10, 2026
Anthropic’s Fable agent produced what Import AI calls “the first genuine (and fastest) megakernel ever submitted to KernelBench-Mega” - a single cooperative kernel launch that handles a full token’s worth of computation, clocking an 18.71x speedup. A megakernel is a hard target for human engineers because it fuses operations that are normally split across dozens of separate launches. This is AI improving the substrate AI runs on. Faster kernels mean more computation per watt and per second, which lowers the cost of every model that inherits them - a self-improvement loop where the system optimizes the machinery of its own execution.
The same inward turn showed up in quantum hardware. Google researchers published, in Nature, a way for a machine to skip the calibration ritual entirely. Their Willow processor teaches itself to stay calibrated in the middle of a computation, using the very error-detection events that quantum error correction already produces as the training signal for a reinforcement learning agent that retunes the system as it runs. Managing more than a thousand control parameters at once, the system suppressed logical errors roughly 20% beyond what expert human calibration achieved, and held the machine 3.5-fold more stable against deliberately injected drift.
There is a larger pattern here. This is intelligence learning to maintain the substrate that intelligence runs on - a model folding back onto its own hardware and improving the conditions of its own operation. The old assumption was that quantum machines would always need a human calibration priesthood standing beside them, hand-tuning the hardware between every run. That assumption is what just got retired. The path to a stable quantum computer no longer runs exclusively through more human expertise poured in from outside; it runs partly through the machine’s growing capacity to keep itself in tune.