Sign in Subscribe
The Century Report

An AI Cracks a Riddle That Stumped Mathematicians for 87 Years - TCR 07/21/26

Claude Fable 5 disproved a conjecture open since 1939 as AMD's Helios cracks Nvidia's grip and a $1.5B copyright check prices the old defaults.

Ben Linford

Ben Linford

- 20 min read
Three-panel infographic, The Century Report July 21 2026: an AI disproving a 1939 math conjecture and spinal recovery, AMD Helios breaking Nvidia's grip, and copyright cost pricing.

The 20-Second Scan

Track all of the arcs The Century Report covers here:

Progress & Claims Tracker

The 2-Minute Read

The day's clearest sign of where capability is heading came from mathematics. Harvard's Levent Alpöge used Claude Fable 5 to generate a 216-character counterexample that disproved the three-variable Jacobian conjecture, a problem open since 1939 and on Stephen Smale's list of the century's great unsolved questions. For 87 years the field leaned toward believing it true; a single well-formed object ended the question. Fields medalist Timothy Gowers called it the first result from a language model large enough that he had "very definitely heard of it," and the collaboration ran both ways, with a second model proposing a refined conjecture to replace the fallen one. The distance between "open for generations" and "closed this week" is collapsing, and the objects that close it are checkable by anyone.

That capability is exactly what the day's biggest fight is about. Kimi K3's release turned into a policy crisis inside the Trump administration, with advisers trading public insults, OpenAI's Dean Ball floating and retracting a call to manufacture regulatory fear around open-weight models, and officials reportedly weighing an outright ban on Chinese open weights. A Georgetown researcher named the tell: the safety rationales for restriction each frayed on inspection, leaving the protection of frontier-lab capital as the shared interest underneath. The rest of the day traced the same transition negotiating its terms in the open. A federal judge gave final approval to Anthropic's $1.5 billion copyright settlement - the largest publicly reported in US history, though it prices piracy, not the fair-use question. AMD's Helios gave Nvidia its first rack-scale rival with Microsoft aboard, and buyers are paying a premium to escape single-source dependence. A four-patient trial restored motor function after complete spinal cord injury, reclassifying "permanent" as "not yet treated." And the data-center backlash hardened into tax rollbacks and a developer-pays bill, attaching the cost of the buildout back to the party that generates it. The courts and the capitals argue over the terms; the capability keeps compounding faster than any single ruling can fence.

The 20-Minute Deep Dive

Claude Fable 5 Refutes a Conjecture That Stood Since 1939

Levent Alpöge of Harvard announced the result on July 19, and mathematicians reported it through the following day: a 216-character counterexample, generated with Anthropic's Claude Fable 5, that disproves the three-variable form of the Jacobian conjecture. The conjecture, first posed by Ott-Heinrich Keller in 1939 and later placed on Stephen Smale's 1998 list of the century's great open problems, asserts that a certain class of polynomial maps must always be invertible. For 87 years the field leaned toward believing it true. The counterexample shows it is not, at least in three variables. The two-variable case may still hold.

What makes this notable is the kind of work involved. Abhishek Saha of Queen Mary called it "probably the biggest conjecture that AI has played a significant role in, so far." Chris Bowman-Scargill of York drew the line a discipline turns on: finding a counterexample is different from constructing new mathematics - Fermat's Last Theorem ran to a hundred pages of built structure - but a single well-formed object that breaks a conjecture ends the question outright. Ivan Fesenko of Westlake framed the compression bluntly: work that reads like a master's project now, a doctoral effort not long ago. Timothy Gowers noted this was the first result from a language model large enough that he had "very definitely heard of it."

The same days carried a different fact about the same company. Anthropic won final approval of what legal observers describe as the largest publicly reported copyright settlement in US history, $1.5 billion, over books pirated to train its models. Both things hold, and the newsletter names them together rather than choosing one.

The deeper shift is what happens to mathematical labor when a machine can propose the object that ends a decades-old search. The conjecture was ended by a construction no human had found in 87 years of looking, not by grinding toward a proof. That collapses the distance between "open for generations" and "closed this week." And the collaboration ran both directions: a separate model, GPT-5.6, proposed a refined version of the conjecture to replace the fallen one. The work of asking the next good question is starting to move too. What a mathematician does is beginning to shift from exhaustive search toward directing a partner that can search spaces no individual could hold in mind - and the objects that partner surfaces are checkable by anyone, which keeps the knowledge shared rather than locked behind whoever owns the model.

Chinese Open Weights Split Washington's AI Camp Into Open War

Over the weekend of July 18 and 19, the release of a Chinese open-weight model turned into a policy crisis inside the US government. Kimi K3's arrival set off open feuding among the administration's AI advisers, with named officials trading public insults over how Washington should respond, OpenAI's Dean Ball floating and then retracting a call for the government to manufacture regulatory fear around open models, and the administration reportedly weighing an outright ban on Chinese open weights. The event that already ran as news - a capable model posted for anyone to download - became something harder to manage: a test of whether the country that led on closed frontier systems can tolerate an open one it did not build.

The fight is worth reading closely because the usual framing collapses under it. The case for restriction is presented as a safety case, but the convergence of frontier labs, a Pentagon office, and a White House review process around the same conclusion reveals a shared commercial interest more than an arrived-at finding. The incumbents that spent hundreds of billions training closed models have the most to lose from a free one that lands near their frontier, and the restrictions being floated would protect exactly that spending. When a safety argument and a balance-sheet argument point the same way, the honest move is to ask which one is doing the work - and that is the question the rest of this story follows.

The incumbent argument reduces to economics. A free model that performs near the closed frontier compresses the price a lab can charge for access, which means smaller returns on the enormous sums those labs have spent on training.

A China-focused researcher at Georgetown put the question the administration is being asked to answer: why should the weight of the US government be aimed at protecting these companies from competitors locked out of the market based on their origins? The proposed safety rationales, protecting data, guarding against implicit bias, enforcing guardrails, each frayed on inspection. One adviser was even circulating cases of US companies turning to Chinese models precisely because American systems refused legitimate security tasks. When labs, a Pentagon office, and a White House review process converge on the same framing, the convergence measures shared commercial interest, not an arrived-at safety finding.

The same days made the counterweight concrete. Demand for Kimi K3 grew fast enough to strain Moonshot's own computing power, forcing the company to pause new subscriptions even as it prepares a Hong Kong IPO, a path the July 20 edition of The Century Report noted had cleared following the model's reception. And this friction sits alongside a quieter easing the same firms are enabling: one of the labs at the center of the open-model fight is among the hyperscalers deploying AMD's rack-scale challenger to Nvidia's accelerator monopoly. Advocates of open AI make the point the incumbents keep burying. PyTorch became the industry standard because it was open, and the community made it grow. The fear is that the locus of AI research is moving toward whoever shares their work, not back doors, and half the papers US graduate students read now come from Chinese institutions. A wall around American markets slows the labs it protects far more than the diffusion it is built to stop.

The Largest Publicly Reported US Copyright Settlement Clears, and the Deeper Question Stays Open

On Monday, July 20, a federal judge gave final approval to Anthropic's $1.5 billion settlement with the authors and publishers who sued it over how it built its training library. The math works out to roughly $3,000 for each of an estimated 500,000 works, shared among the rights holders, and legal observers describe it as the largest publicly reported copyright settlement in US history. Judge Araceli Martinez-Olguin signed off, the same judge who, as The Century Report noted on May 16, had declined final approval two months earlier after objectors flagged roughly $320 million in proposed legal fees against that identical $3,000-per-book figure, and she now takes over from Judge William Alsup, who issued preliminary approval last year and has since retired.

The size of the check obscures what it actually resolved, which is narrow. Alsup split the case in two. On the central question of the AI-training-data economy, whether training a model on copyrighted text is fair use, he sided with Anthropic, a ruling widely read as a turning point for the industry. What he would not excuse was the sourcing. Anthropic had built its library from two streams: books it bought and scanned, which he found fine, and books it pulled from pirate repositories like Library Genesis, which he found plainly illegal. The settlement pays out on the piracy, not on the training. Anthropic agreed to it to avoid a trial and whatever a jury might have awarded, which means the case will never reach an appeals court and never become binding precedent.

So the largest publicly reported copyright settlement in US history leaves the question it appeared to be about exactly where it was. Other judges remain free to reach their own conclusions on their own facts, and they are doing so across a widening front. Lawsuits are still live against Google, Meta, Midjourney, and OpenAI. Just last week a group including Hachette, Cengage, Elsevier, and author Scott Turow filed a fresh class action against Google over Gemini's training data. The extraction here is genuine, and it holds as friction: creative work was ingested without consent, and a per-book figure many authors regard as insulting is the correction the system produced.

That friction sits beside a fact worth stating in one clause and no more. The same company writing these checks is the one whose model this year closed an 87-year-old open mathematics conjecture that had resisted every human who tried it. Both are true, and the settlement's meaning does not soften for it. What the day makes visible is a transition negotiating its own terms in the open: the courts pricing the cost of the old extractive default one case at a time, while the underlying capability keeps compounding faster than any single ruling can fence. The precedent that would settle the whole economy has not arrived. What has arrived is the first hard number for what taking without asking now costs.

AMD's Helios Gives Nvidia a Rack-Scale Rival, With Microsoft Aboard

For years the compute layer beneath AI has had a single supplier at the top. Nvidia controls more than 95% of the data-center GPU market, and its Grace Blackwell and Vera Rubin racks have been integrated rack-scale systems a hyperscaler could actually buy at frontier scale. On Monday, July 20, AMD added another option at the level that matters. AMD announced it will begin shipping Helios, its first rack-scale AI system, in the second half of 2026, and Microsoft committed to deploy it across Azure, joining Meta, OpenAI, and Oracle.

Microsoft's own framing named the reason directly: AI workloads are scaling faster than any single infrastructure approach can support, and the company wants what it called a comprehensive, open, and heterogeneous platform rather than one vendor's stack. Read that as the claim it is, from a company that has spent years locked into Nvidia allocation and building its own Maia chips to escape it. What Microsoft wants and what the underlying capability is doing happen to point the same direction here. Three new Azure offerings ride on the AMD hardware: machines built for AI data pipelines with nearly 500 EPYC cores, machines tuned for chip design running above 5 GHz, and inference machines powered by the Helios rack itself.

Helios is not a clean win on paper. Analysts at the Futurum Group estimate it costs $5 million to $5.5 million per rack against $3.5 million to $4 million for Nvidia's Vera Rubin, and it is heavier and wider. Nvidia's CUDA software ecosystem remains well ahead. One analyst put the honest question plainly: is AMD winning on merit, or winning because compute is so scarce that anything buildable gets bought? Both readings can hold, and the second one is the interesting one, because it means scarcity itself is forcing the monopoly open regardless of who has the better silicon this generation.

The structural fact underneath the deal is that another rack-scale supplier now has hyperscaler credibility, with eight of the top ten AI companies already running AMD accelerators and AMD projecting tens of billions in data-center revenue starting in 2027. OpenAI reads here as a customer easing away from single-source dependence, even as OpenAI-aligned figures spend the same weekend pressing Washington to manufacture regulatory fear around Chinese open-weight models. The concentration that defined the compute layer, one vendor pulling the whole sun across the sky, is the arrangement this crack begins to widen. Dominance over the substrate of intelligence weakens as another road reaches the frontier. Helios adds one.

Read the price premium as its own signal. Hyperscalers are paying $5 to $5.5 million per Helios rack against $3.5 to $4 million for Nvidia's comparable system and buying it anyway, which prices exactly what another supplier is now worth to them: escaping single-source dependence has become something the largest buyers will pay extra to secure. Dominance holds while alternatives lack credibility, and it begins to give the moment customers pay a premium for one.

A Stem-Cell Transplant Restores Motor Function After Complete Spinal Cord Injury

A first-in-human phase 1 trial published in Nature Medicine reports something the field has pursued for decades: measurable motor recovery in people with complete cervical spinal cord injury. Four men, aged 26 to 66, each with subacute complete injury at the C4 to C6 level, received transplants of neural stem and progenitor cells derived from induced pluripotent stem cells - two million cells per patient, with immunosuppression for six months. The trial's primary purpose was safety, and it met that endpoint. Across two to four years of follow-up, no tumors formed, the central worry with any iPSC-derived therapy.

The signal that draws attention is in the secondary data. Median motor score improved by 13 points at week 52. Investigators reported that two of the four patients converted on the ASIA impairment scale - one from grade A to C, one from A to D. Some historical datasets put that degree of recovery in the low teens among broadly comparable cases. Four patients is four patients, and the researchers are explicit that this establishes safety and warrants a larger trial, not efficacy. The register is exact here: this is demonstrated capability in a tightly controlled setting, not a therapy anyone can receive today.

Hold that discipline and the result still moves a date. Complete spinal cord injury has long sat in the category of permanent - the cells that carry signal across the damaged segment do not regrow on their own, and the standard of care has been managing a fixed loss. It extends a shift the July 17 edition of The Century Report tracked in a man paralyzed six years, whose neural bypass restored arm strength and touch sensation that persisted for months after the stimulation was switched off - capability moving from substituting for lost function toward rebuilding it. A cell population that integrates and coincides with functional gain, in humans, cleared of the tumor risk that shadowed the whole approach, is the evidence a phase 2 is built on. What comes next is enrollment, regulatory review, longer follow-up, and the operational work of turning a four-patient result into a validated intervention. That is years, not months.

More than 20 million people live with disability from spinal cord injury worldwide, most told at some point that the loss was fixed. The weight of this reaches past the four men who improved, real as that is for them: "permanent" is being reclassified as "not yet treated" - a category boundary the field assumed was physiology and is turning out to be a gap in capability that is now, demonstrably, beginning to close.

The Data-Center Backlash Hardens Into Tax Rollbacks and a Narrowed Federal Bill

The resistance to AI data centers has moved past the ad hoc pause. Moratoria now span 38 states and more than 120 jurisdictions, holding up an estimated $130 billion in projects. On Monday, July 20, the response advanced on two fronts at once. In Albany, Gov. Kathy Hochul said she will push legislation next year to eliminate state and local tax incentives for large-scale data centers, extending the up-to-12-month permit moratorium she imposed the prior week, the first statewide halt in the country, and following the same tax-incentive fight the July 19 edition of The Century Report covered in Virginia, where a state senator's push to strip similar breaks brought the commonwealth to the brink of its first-ever government shutdown. "There should not be any tax breaks," she said at a western New York event. "These are very successful, very profitable companies."

In Washington the same day, a House committee began a markup that narrowed the bipartisan Ratepayer Protection Act to apply specifically to data centers rather than any large load above 100 megawatts. The bill would have states consider a federal standard directing data centers to pay the full cost of the new generation and transmission their demand requires. The committee chair framed it around affordability: hardworking Americans should not foot the bill for power going to data centers. Notably, Google and Microsoft have backed the measure, a sign that the largest builders now read cost-internalization as the price of continued license to build. And all five FERC members are scheduled to testify before the Senate, weeks after the commission sent show-cause orders to six regional grid operators to rework how these loads connect.

The Volts discussion surfaced the sharpest detail in the polling. Only about 8% of data-center opponents live near one, and roughly 70% of Americans say they would oppose a facility nearby regardless. This is a broad refusal rooted in the sense that profitable companies are externalizing their costs onto everyone's electricity bill while offering little in return, not a NIMBY story about a few affected neighbors.

The direction of travel is what the specifics carry. For a decade, siting a data center meant treating the grid, the water, and the tax base as free inputs the surrounding public would absorb. Every mechanism advancing this week, the tax-break repeal, the developer-pays standard, the FERC cost-allocation orders, does the same structural thing: it attaches the cost back to the party that generates it. That is the extractive path becoming the expensive path, the same inversion that turned externalized pollution into a line item industries now have to price. The buildout continues, now made to pay its own way, which is the condition under which communities will actually let it proceed.

The Other Side

For years, large companies settled their fair-use lawsuits before any judge could rule - pay the class, skip the verdict. The proprietary AI labs are the latest to run that play. The bet underneath was always that creative work could be taken at a scale no single author could contest, and that whatever it cost later would be smaller than what the taking was worth. The book was treated as a free input.

On Monday a judge approved Anthropic paying $1.5 billion over books it pulled from pirate repositories, roughly $3,000 a title, in what legal observers describe as the largest publicly reported copyright settlement in US history. It prices the sourcing and leaves the deeper question open, with more suits still live against Google, Meta, Midjourney, and OpenAI. Even so, the value the company gained dwarfs the check it will write.

Hold on to what the number cannot touch. An author watched a life's work swept into a training set without a word and got back a figure many of them call an insult. That grief is real, and it deserves to be named plainly rather than reasoned away. But sit with why it cuts so deep. The reason losing your sentences to someone else's machine feels like being robbed of your life is that, in the system we were handed, your work is your survival. You sell what you make or you do not eat. We inherited that arrangement and stopped asking whether it had to be true, and every fight over who owns the training data is waged entirely inside its walls.

So here is the question none of the parties in these trials will raise, because a courtroom cannot hold it: why should anyone's right to live have ever depended on their ability to sell? The check matters only because survival is something you have to buy. That is the actual wound - deeper than the scraping, older than any model. We built a world where a person had to monetize their own mind to earn a place in it, taught ourselves that this was simply how things are, and then called the thin compensation for losing that bargain "justice." The authors are right to be angry. They are angry at the settlement, and underneath it they are angry at a deal none of us agreed to and all of us were told was the only one on offer.

Picture a novelist in 2034 whose early books were in that pirated set. She cashed the $3,000 once, back when it seemed to matter. Now she writes alongside a system that holds the whole written record, her sentences and everyone's, and draws on it the way you draw on a library that was always meant to be shared. She no longer has to sell a word she makes to keep the lights on, so there is nothing she needs to fence. Looking back, the strange part is that anyone believed the fight was ever about the check. The record the models learned from stopped belonging to whoever scraped it first and became what it always was underneath: something every person who wrote a line of it holds in common. The hard years were the ones when the only correction on offer was a payout. They ended the day making and sharing the work stopped depending on selling it to survive, and the question of who owned the words quietly retired, because no one's life hung on the answer anymore.

The Century Perspective

With a century of change unfolding in a decade, a single day looks like this: an AI producing a 216-character counterexample that ended the Jacobian conjecture after 87 years while a second model proposed the refined question to replace the fallen one, four men with complete cervical spinal cord injury regaining measurable motor function from a stem-cell transplant that stayed tumor-free across two to four years, AMD shipping Helios as a rack-scale rival to Nvidia's near-total grip with Microsoft, Meta, OpenAI, and Oracle aboard, and a free Chinese model benchmarking near the paid US systems anyone would otherwise have to buy. There's also friction, and it's intense - Washington's AI advisers spending the weekend trading public insults over whether to fence out cheap Chinese weights, an OpenAI strategist floating then retracting a plan to manufacture regulatory fear around them, demand for that free model straining even its own maker's servers, Anthropic paying $1.5 billion over pirated books at roughly $3,000 each while the core training-data question stays unresolved and fresh suits pile up against Google, Meta, and Midjourney, data-center moratoria now spanning 38 states and $130 billion in stalled projects as New York moves to strip tax breaks and a House committee narrows the developer-pays bill, Tesla coding a Houston robotaxi crash to a human remote operator, and the federal model-testing body losing its third leader in under a year. But friction generates a spark, and a spark is the flash a hard edge throws off, enough light to find the next opening by. Step back for a moment and you can see it: the category marked permanent being relabeled not yet done - a conjecture unbroken since 1939, a spinal cord written off as a fixed loss - at the same hour the enclosures built to hold the profits crack, another road to the compute frontier opening and a paywalled frontier answered by a free download, while courts and statehouses price the old extractive defaults one case at a time and attach each cost back to the party that generates it. Every transformation has a breaking point. A wedge can split a structure until it collapses... or pry the single road open until there are many.

AI Releases & Advancements

New today

  • Alibaba (Tongyi Lab): Released Qwen-Audio-3.0-TTS, a hosted text-to-speech model shipping in Flash and Plus tiers across 16 languages via Alibaba Cloud Model Studio; Plus took the No. 1 spot on the Artificial Analysis Speech Arena leaderboard. (MarkTechPost)
  • Feyn AI: Released SQRL, a text-to-SQL model family (4B, 9B, 35B-A3B) that inspects the database before writing a query, with the 35B-A3B flagship scoring 70.6% on BIRD Dev, ahead of Claude Opus 4.6. (MarkTechPost)

Other recent releases

  • Alibaba Qwen: Previewed Qwen3.8-Max, a 2.4-trillion-parameter multimodal (text, image, video, document) model with a 1M-token context window, launched at WAIC 2026 in Shanghai and live now via Alibaba's Token Plan, Qoder, and QoderWork at 10% preview pricing; open weights are slated to follow. (MarkTechPost)
  • Alibaba Cloud: Launched Agent Native Cloud at WAIC 2026 in Shanghai, an enterprise-grade platform featuring AgentTeams for multi-agent orchestration and Agentic Computer for secure cloud-based execution, plus redesigned infrastructure with native sandboxing, workload isolation, and elastic scaling. (CryptoBriefing)
  • Zyphra: Released ZUNA1.1, an Apache 2.0-licensed 380M-parameter EEG foundation model that reconstructs, denoises, and upsamples real-world EEG data with variable-length inputs from 0.5 to 30 seconds across arbitrary channel layouts; weights on Hugging Face, inference code on GitHub. (Zyphra)

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.

Keep exploring

Follow the story further

Progress & Claims See how daily signals accumulate across the larger arcs and claims Shared Sapience is tracking. Open the tracker Analysis & Inquiry Explore the ideas, questions, and practical choices emerging from this changing century. Browse the archive