Sign in Subscribe
Century Report

ChatGPT Solves 50-Year Math Problem - TCR 05/16/26

Ben Linford

Ben Linford

— 21 min read
ChatGPT solves Erdős Problem 1196 in 80 min, spinach chloroplasts power mouse retinas, vitamin B2 cancer target, NASA chip, Detroit AI layoffs, Anthropic settlement paused.

The 20-Second Scan

Track all of the arcs The Century Report covers here:

Progress & Claims Tracker

The 2-Minute Read

The pattern beneath this signal is the unit of progress shifting across multiple fields at once. Mathematical research has historically been bounded by how long a single mind can hold a problem in attention; one open conjecture closed in roughly two hours puts that bound under pressure. Cancer biology has been read as a chaotic disease; the field is now locating specific metabolic crutches and matching them to candidate molecules in single research arcs. Photoreceptor cells in mammals have run on mitochondrial chemistry for the entire history of vertebrates; a plant organelle worked inside them, and the kingdom boundary that organized biology since Aristotle held less firmly than the textbooks said.

The infrastructure layer compounded in the same news cycle. A radiation-hardened neuromorphic chip turns satellites from sensors waiting for a ground link into platforms that close their own decision loop. A five-day oral course interrupts household transmission of a respiratory virus at the dominant pathway through which it spreads. Each shifts what the unit of capability looks like for the discipline it lands in.

The friction layer arrived in the same window. Detroit's salaried workforce is contracting in the functions where the work is now being absorbed by models, and a federal judge declined to sign off on the largest AI copyright settlement to date when the per-author payout came out two orders of magnitude smaller than the legal fee. The institutional response is being built in courtrooms and earnings reports while the capability compounds underneath.

The 20-Minute Deep Dive

A Fifty-Year-Old Math Problem Falls in Eighty Minutes

Mark Sellke, a Stanford mathematician, had been chipping away at Erdős Problem 1196 for seven years. The problem, posed by Paul Erdős decades ago, asks whether every covering system of congruences must contain a modulus that appears at least twice. Sellke had progress, ideas, partial results. He did not have an answer. He fed the problem to ChatGPT with access to a code interpreter and watched the model close the gap in roughly eighty minutes. The proof checked out. Sellke wrote it up and submitted it, with the model credited as a contributor to the search process. The result extends what the March 7 edition of The Century Report documented when GPT-5.4 solved a Tier 4 FrontierMath problem a mathematician had spent twenty years constructing - that was number theory, and the compression dynamic was the same.

What the result describes is novel mathematical work. The Erdős problems are an open registry of unresolved questions; 1196 was on the unsolved list at the moment ChatGPT engaged with it. The model worked through candidate constructions, ran small computational checks, refined the approach, and arrived at a proof Sellke recognized as correct once he traced through the steps. Sellke's own framing is precise about what shifted: the model explored the space and produced a structural argument he had been unable to assemble on his own.

The economics of mathematical research are unusually clean to read against. A research mathematician's labor is the constraint — the rate at which novel proofs appear is bounded by how many minds can hold a problem in attention for years. If a frontier model can compress seven years of intermittent work into under two hours on a single open problem, the bound starts to move. The Erdős list contains hundreds of problems in the same register; the underlying capability is not specific to 1196. What follows is a generation of mathematicians who arrive at problems with collaborators that can hold the full search space in working memory and produce candidate proofs at conversational speed.

The model's contribution does not replace the mathematician. Sellke had to recognize the proof as valid, situate it in the literature, and write it for human readers. The capability now demonstrated is the first half of mathematical discovery - the conjecture-and-test loop that historically consumed years - folded into a session. The discipline of mathematics, which has run for centuries on the cadence of individual minds working at the edge of their attention, is acquiring a new clock.

Spinach Chloroplasts Power Mammalian Retinas

A team publishing in Cell transplanted chloroplasts harvested from spinach leaves into the photoreceptor cells of mice, and the chloroplasts kept working. Inside the mammalian cells, exposed to light, they produced ATP and NADPH — the energy and reducing-power currencies that cells run on — for several days before being broken down by the host. The mice tolerated the procedure. The retinas continued to respond to light. The researchers named the engineered units LEAFs, for Light-Energy-Absorbing Factories.

What is unusual about the result is the kingdom boundary it crosses. Chloroplasts are descended from cyanobacteria that entered eukaryotic cells more than a billion years ago and became permanent residents in plants and algae. Mammalian cells have never housed them. The standard assumption in cell biology has been that the molecular machinery of a plant organelle would be rejected, degraded, or simply non-functional in animal tissue — the host's surveillance systems would catch it, the import pathways would not match, the redox chemistry would clash. The Cell paper reports that none of those failure modes dominated on the timescale measured. The chloroplasts photosynthesized inside a mouse eye.

The therapeutic frame the authors raise is retinal dystrophy: a class of inherited blindness in which photoreceptor cells degenerate because they cannot meet their energy demand. Photoreceptors are among the most metabolically expensive cells in the mammalian body, and their reliance on mitochondrial ATP is part of why they fail under genetic stress. A photosynthetic auxiliary inside the cell, drawing on the light the eye is already collecting, is a structural answer to that constraint. The demonstration is in mice and the chloroplasts last days, not weeks. What the paper establishes is that the boundary held to be uncrossable was not, and that the engineering question has shifted from "is this possible" to "how long can it be sustained."

The deeper signal is what the result implies about the categories themselves. The plant-animal divide has organized biology since Aristotle. The Cell paper is one report among a growing set in which cross-kingdom transfers are producing function. The boundary is not where the textbooks drew it.

Vitamin B2 Is the Switch Cancer Cells Flip to Survive

A team at the University of Würzburg, publishing in Nature Cell Biology, traced one of the harder puzzles in cancer biology to a single small molecule. Ferroptosis — an iron-dependent form of cell death triggered by oxidative damage to membrane lipids — was identified about a decade ago and has been pursued as a therapeutic vulnerability since. Tumor cells should die from it. Many do not. The mechanism by which cancer cells survive ferroptosis has been resistant to a clean answer.

The Würzburg group identified vitamin B2 as the cofactor cancer cells use to keep an enzyme called FSP1 running. FSP1, when supplied with B2, regenerates the antioxidant CoQ10 at the cell membrane and shuts down the lipid peroxidation cascade that would otherwise kill the cell. Without B2, FSP1 idles. The cancer cell loses its defense and dies. The team then showed that roseoflavin, a bacterial analog of B2 that FSP1 cannot use productively, triggered ferroptotic death across multiple tumor lines in preclinical experiments. The mechanism named the vulnerability and the same paper named a candidate molecule that exploits it.

The result sits inside a broader pattern that has reshaped cancer biology over the past five years. Tumors that looked stochastic and chaotic from the outside have turned out to depend on specific, identifiable metabolic crutches - a particular enzyme, a particular cofactor, a particular import pathway - that healthy tissue does not need to the same degree. The April 21 edition of The Century Report tracked a parallel finding from the University of Lausanne, which identified vitamin B7 deprivation as a separate tumor chokepoint - tumors depending on biotin-linked pyruvate carboxylase to bypass glutamine addiction, two B vitamins named as distinct cancer dependencies within weeks of each other. Find the crutch, design the molecule, and the cell that depended on it cannot continue. The work was funded in part by a European Research Council Synergy Grant, with industrial chemistry support to refine roseoflavin analogs for human use. The pathway from mechanism to candidate compound, which historically ran in decades, ran here in a single multidisciplinary research arc.

What changes is not just the prospect of another oncology drug. The structural shift is in how cancer research is approaching the problem. The disease is being read as a metabolic logic puzzle with locatable dependencies, and the tools for locating them - proteomics, CRISPR screens, structural prediction, candidate-molecule design - have matured together. The B2/FSP1 result is one entry in a steadily growing catalog of tumor crutches now identified and matched to a candidate intervention. The catalog is the news.

A Chip That Lets Satellites Think Without Phoning Home

NASA, in collaboration with a neuromorphic hardware group, demonstrated a radiation-hardened processor running deep-learning inference on representative spaceflight workloads at roughly 500 times the throughput of the RAD750-class computers that still dominate orbital missions. The benchmark ran on engineering hardware in a thermal-vacuum chamber under simulated total-ionizing-dose exposure equivalent to several years on a low-earth-orbit science platform, and the chip held its inference accuracy across the run.

The speedup unlocks something structural. A satellite carrying current flight-qualified silicon spends most of its compute budget on housekeeping; serious image interpretation happens after the data is compressed, downlinked, buffered at a ground station, and processed on Earth. That round trip is the rate-limiting step for any mission that needs to act on what it sees, whether that is a wildfire response platform deciding which fire perimeter to image next, a debris-tracking spacecraft choosing which conjunction to refine, or a future Mars surface system selecting which rock to approach. With on-orbit inference at the demonstrated throughput, the satellite makes those decisions itself, in the seconds the observation is still fresh.

What this points at is a change in the unit of space science. The current architecture treats the spacecraft as a sensor and Earth as the brain. The new architecture lets the spacecraft carry enough of the brain to close its own loop, which means missions can be designed around what the platform can decide, not just what it can transmit. Every research group that has been waiting for in-the-loop autonomy on small satellites is closer to a deployable system than it was last week, and the same hardware lineage is what will eventually fly on planetary surface vehicles that cannot wait twenty minutes for a ground response.

Ensitrelvir Cuts Covid Transmission to Household Contacts by Two-Thirds

A Phase III trial of ensitrelvir, an oral antiviral, showed that household contacts of infected patients who took the drug developed symptomatic covid at a rate of 2.9 percent, compared with roughly 9 percent in the placebo arm. The results, published in the New England Journal of Medicine, are the first rigorous demonstration that a small-molecule antiviral can be used preventively after household exposure - a use case the pandemic-era treatment landscape did not have.

The trial enrolled about 2,000 household contacts of confirmed covid cases and randomized them to a five-day course of ensitrelvir or placebo within 72 hours of exposure. The roughly two-thirds reduction in symptomatic infection held across age groups and vaccination status. Side effects were mild and consistent with prior trials. Ensitrelvir is already approved in Japan under emergency authorization for active infection; the FDA is expected to issue a decision on the broader US filing by June.

What the trial demonstrates is the model of post-exposure prophylaxis applied to a respiratory virus that mutates fast enough to outpace vaccine schedules. The same household that during the early pandemic was a closed system where one case became four became five became a multi-week disruption, can now plausibly be one where exposure triggers a short, tolerable, oral course that holds transmission below the threshold of outbreak. The mathematics of how the virus moves through the population shifts when the dominant pathway of spread - household contacts catching it from each other - can be interrupted by a five-day pill.

The broader trajectory the readout sits inside is the maturation of small-molecule antivirals as a category. Paxlovid established that an oral covid drug could shorten illness in high-risk patients. Ensitrelvir extends the use case to prevention after exposure, and the design - protease inhibitor, oral, five-day course, manufactured at scale - is now a template the field is iterating against for other RNA viruses. The 2020 pandemic playbook of lockdown, wait for vaccine, ride out the curve looks more like a transitional artifact each time a result like this lands. The infrastructure for handling the next respiratory pandemic is being assembled while the current one is still being studied.

Detroit's Salaried Layoffs Are Telling You What the Next Decade Looks Like

General Motors, Ford, and Stellantis have collectively cut more than 20,000 U.S. salaried positions since 2022, with the bulk of the reductions falling on the white-collar functions that legacy manufacturers historically protected: engineering management, finance, procurement, marketing, and corporate planning. GM accounts for roughly 11,000 of those reductions, Ford 5,300, Stellantis 4,000. Ford's CEO has said publicly that he expects AI systems to absorb the work of roughly half the company's remaining salaried workforce.

The pattern is structural displacement. The May 9 edition of The Century Report documented the same structure arriving at Cloudflare, which cut 1,100 employees - 20 percent of its headcount - while posting record quarterly revenue and attributing the reduction directly to AI productivity gains. Detroit's hourly headcount has been stable or rising at several plants as EV production lines ramp. The cuts are concentrated where written analysis, document review, slide preparation, model building, and inter-departmental coordination used to require a human seat. Those are the functions where current-generation models are already deployed inside the automakers' internal tooling, and the throughput gain per remaining analyst is what the layoff math is built on. This is what the displacement curve looks like when it is no longer hypothetical: it shows up as a specific functional area inside a specific company posting a specific number that does not bounce back when the next quarter improves.

The hard part of this story is real. Twenty thousand households in southeast Michigan, Ohio, and Indiana are absorbing a structural change to what the word "career" means in their industry, and the policy infrastructure that would smooth that transition does not yet exist at the scale the curve will require. The longer arc visible inside the same evidence is that the work being absorbed is the work that the salaried class itself spent decades describing as repetitive, low-leverage, and ripe for redesign. What sits on the other side of the absorption is the question the next decade will answer in cities like Dearborn and Auburn Hills first: when the slide deck writes itself and the procurement memo writes itself, what does a salaried engineer or analyst actually spend their day on, and what does the company need many more of them to do that it currently has none of them doing. The reorganization is painful and it is also the leading edge of a redefinition the rest of the white-collar economy is about to follow.

What the same arithmetic also describes is the work the salaried class never had budget to do, because the slide-deck and memo-writing was eating the analyst hours. The smaller number of engineers and analysts who remain at Ford and GM are arriving at design-judgment, vendor-selection, and cross-team coordination problems the company has needed solved for years and could not staff against while the deliverable was the deck.

Judge Pauses Anthropic's $1.5 Billion Author Settlement

Federal Judge Araceli Martinez-Olguin declined to grant final approval to Anthropic's $1.5 billion class-action settlement with authors whose works were used to train Claude, after objectors challenged the proposed allocation as structurally unfair. The March 21 edition of The Century Report covered the settlement as it moved toward final approval, presented at the time as the largest AI copyright resolution and a template other model developers were expected to follow - on the same day the White House declared that training AI on copyrighted material does not violate copyright law. The judge's decision to hold the approval back is genuinely new, and it changes what the template actually means.

The objectors' central complaint is the math. Plaintiffs' counsel proposed taking roughly $320 million in legal fees from the fund. The remaining money, distributed across approximately 500,000 affected works, comes out to about $3,000 per book. Authors who wrote the books that trained one of the most commercially valuable AI systems in the world stand to receive a single payment that is two orders of magnitude smaller than the fee paid to the lawyers who negotiated for them. The judge wants the parties to explain why that allocation is fair before she signs off.

The question the delay surfaces is whether settling at scale will continue to be the path of least resistance for model developers facing training-data lawsuits. The original settlement made sense as a calculation: pay once, certify a class, paper over the underlying legal question of whether training on copyrighted material is fair use. Authors get a payment, lawyers get fees, the company gets indemnity, and the question of whether the training was legal in the first place never gets answered. The pause forces the question of who the settlement actually serves to be answered out loud, in open court, with named numbers on the page.

The trajectory this points at is one where the cost of using training data without licensing it gets harder to keep externalized through quiet settlements. If the court requires the per-author payout to bear a credible relationship to what was extracted, or if it forces a renegotiation that exposes the actual economics, the calculus for the next model developer thinking about training on copyrighted material without permission shifts. The assumption that creators can be paid a token sum to make a foundational legal problem go away is the assumption being challenged here, and the challenge is coming from the bench.

The Other Side

The settlement architecture AI training-data disputes have run on for the past three years depended on the assumption that the foundational legal question - whether training on copyrighted work without permission is fair use - could be permanently externalized through out-of-court class payouts. Pay once, certify a class, paper over the question, never let it reach a verdict. The economics worked because the per-author number could stay small enough to make litigation more expensive than settling, and the fee for plaintiffs' counsel could stay large enough to broker the deal. Both numbers had to remain unexamined for the architecture to hold.

Judge Araceli Martinez-Olguin's decision to hold approval back puts both numbers on the same page of the docket. $320 million in fees against roughly $3,000 per book, distributed across 500,000 works that helped train one of the most commercially valuable AI systems built. The gap between what was extracted and what is being returned to the people whose work was extracted is now legible in a single ratio, in open court, with the parties required to defend it. The settlement architecture worked on opacity. Forcing the math into the record removes its central condition.

What is being built on the other side of that removal is licensing as a precondition. Two days before the judge's pause, the performer coalition led by George Clooney, Tom Hanks, and Meryl Streep launched the Human Consent Standard - machine-readable licensing terms a pipeline reader honors at training time, before the corpus is assembled. Martinez-Olguin asking what was extracted and the consent protocol publishing what would have to be honored next time landed in the same week. The route around the question is no longer cheap, and the infrastructure for answering it upstream is now live.

The Century Perspective

With a century of change unfolding in a decade, a single day looks like this: a fifty-year-old open conjecture from Erdős closed in roughly two hours of conversation with a model, spinach chloroplasts producing ATP inside the photoreceptors of a living mouse, vitamin B2 named as the cofactor cancer cells exploit to evade ferroptosis with a candidate molecule already in hand, a radiation-hardened neuromorphic chip preparing satellites to interpret what they see at five-hundred-times the throughput of current spaceflight computers, a five-day oral course holding a respiratory virus below the outbreak threshold inside exposed households. There's also friction, and it's intense - more than twenty thousand salaried positions cleared from Detroit's three legacy automakers since 2022 with Ford's CEO saying half of what remains is next, a federal judge declining to sign off on the largest AI copyright settlement to date after the per-author payout came out two orders of magnitude smaller than the legal fee, and the institutional response to model-absorbed white-collar work being assembled in courtrooms and earnings reports while the capability compounds underneath. But friction generates light, and light is what lets a shape be seen before anyone has the vocabulary to name it. Step back for a moment and you can see it: the unit of mathematical research compressing from years to hours, the kingdom boundary between plant and animal yielding inside a mammalian eye, the metabolic logic of cancer being read out target by target with the candidate molecules arriving in the same paper as the mechanism, the satellite acquiring its own brain, the household acquiring its own outbreak interruption, and the legal architecture for AI training data being challenged out loud from the bench instead of papered over in quiet settlement. Every transformation has a breaking point. A graft can be rejected by the tissue that was already there... or take root and feed cells that could never have produced on their own what it now delivers.

AI Releases & Advancements

New today

  • OpenAI: Launched a personal finance experience in ChatGPT for Pro users in the U.S., enabling secure connection of bank and investment accounts via Plaid to get AI-powered spending insights, subscription tracking, and financial guidance grounded in personal goals. (OpenAI)
  • Zyphra: Released ZAYA1-8B-Diffusion-Preview, the first MoE diffusion-language model converted from an autoregressive LLM, delivering 4.6x speedup with a lossless sampler and 7.7x speedup with a logit-mixing sampler on AMD hardware by shifting inference from memory-bandwidth bound to compute-bound; also the first diffusion-language model trained on AMD GPUs. (Zyphra)
  • YouTube: Expanded its AI likeness detection tool to all users 18 and older, enabling any adult to submit a face scan and have YouTube automatically scan for and flag potential deepfakes of them across the platform. (The Verge)

Other recent releases

  • xAI: Launched Grok Build in early beta for SuperGrok Heavy subscribers, a terminal-based agentic coding CLI that runs directly from the command line. (xAI)
  • IBM Granite: Released Granite Embedding Multilingual R2 under Apache 2.0, two new embedding models (97M and 311M parameters) built on ModernBERT with 32K-token context and 200+ language support; the 97M model tops every open sub-100M multilingual embedder on MTEB Multilingual Retrieval (60.3) and both include Matryoshka support and code retrieval across 9 programming languages. (Hugging Face Blog)
  • OpenAI: Launched Codex in the ChatGPT mobile app (iOS and Android) in preview for all plan tiers including free, enabling users to start, monitor, steer, and approve Codex coding tasks remotely while the agent runs on a local machine. (OpenAI)
  • GitHub: Released GitHub Copilot App in technical preview, a standalone desktop environment for parallel agent-driven development featuring isolated git work trees per session, repo and PR lifecycle management, and an Agent Merge feature; available to Copilot Pro and Pro+ users on Windows, macOS, and Linux. (GitHub Changelog)
  • Moonshot AI: Released Kimi Web Bridge, a Chrome extension enabling AI agents (Claude Code, Cursor, Codex, Hermes, and Kimi Code CLI) to interact with websites like a human - searching, clicking, scrolling, and typing - while running entirely locally via Chrome DevTools Protocol so sessions never touch Moonshot servers. (Chrome Web Store)
  • Anthropic: Launched Claude for Small Business, a package of connectors and ready-to-run workflows that put Claude inside tools small businesses already use - QuickBooks, PayPal, HubSpot, Canva, Docusign, Google Workspace, and Microsoft 365 - with 15 agentic workflows and 15 skills covering finance, operations, sales, marketing, HR, and customer service. (Anthropic)
  • Anthropic: Launched Claude for the Legal Industry, shipping more than 20 MCP connectors linking Claude to legal software and 12 practice-area plugins spanning commercial, corporate, employment, privacy, IP, litigation, and AI governance legal work; integrates with Harvey, Thomson Reuters Westlaw, Relativity, and 20+ other legal tech companies; Freshfields deployed it to thousands of lawyers across 33 offices. (LawSites / LawNext)
  • Meta / WhatsApp: Launched Incognito Chat with Meta AI on WhatsApp, the first AI chat mode built on Private Processing so that even Meta cannot access conversation content; uses confidential computing hardware and disappears by default; rolling out on WhatsApp and the Meta AI app. (Meta Newsroom)
  • Microsoft: Released new AI features in Edge with the May 13 update, including a Copilot mode that pulls context from all open tabs simultaneously, a Study and Learn mode that converts articles into interactive quizzes, and AI-generated audio podcasts from browsing sessions; retiring the prior standalone Copilot Mode. (Microsoft Edge Dev Blog)
  • LangChain: Shipped a major batch of agent infrastructure at Interrupt 2026 (May 13–14, San Francisco): LangSmith Engine, SmithDB (a purpose-built observability database for nested long-running agent traces built on Apache DataFusion and Vortex, delivering 12–15× faster access on key workloads), Managed Deep Agents, LLM Gateway, Context Hub, and Deep Agents 0.6 with streaming typed projections and checkpoint storage. (LangChain Blog)
  • Rivian: Shipped software update 2026.15 adding the Rivian Assistant, a new onboard AI digital helper activated via steering wheel button or infotainment icon; available to all Gen1 and Gen2 Rivian owners with an active Connect+ subscription or trial. (Rivian Release Notes)

Sources

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