Sign in Subscribe
The Century Report

Anthropic's IPO Filing Puts Control in One Share

Anthropic's IPO plan gives seven co-founders 50.1% of ordinary shareholder voting power through one share held by an LLC and lists $265.6 billion in fixed computing payments.

Ben Linford

Ben Linford

- 31 min read
Infographic: Anthropic's golden share and compute bills, OpenAI's math claims handed to outside checkers, Microsoft's 60,000 buyouts, a Cloudflare cache leak and Britain's AI growth zones.

The 20-Second Scan

Every story in today's edition, one line each.

Track all of the arcs The Century Report covers here:

Progress & Claims Tracker

The Big Picture

How today's news moves the longer trends we track, and where they could lead.

Today the money arrived and the votes stayed behind. Meta, SpaceX and Anthropic drew tens of billions from outside investors in forms that carry no say, while OpenAI handed its boldest proof to the people best placed to break it. Where that money leads is a loading dock in 2027, where pallets of Nvidia racks owned by an Apollo fund roll into a hall rented to SpaceX.

A loading dock at a large data center, its roll-up door open onto a long white hall.
Illustration: AI-generated

Outside money funds the AI buildout while founders keep the votes

AI finance · governance · data centers

Chart: Meta's stock sale outsizes Google's and Oracle's. Meta stock offering: $95 billion; Google stock offering: $85 billion; Oracle stock offering: $25 billion

The line: Each new round of outside money for AI arrives in a form that carries no vote, while founders and the councils they appoint keep control.

If it holds: Founders and their councils can keep deployment decisions off a quarterly earnings clock, as Anthropic argues. Shareholders, pension funds and lenders supply the money and absorb any repricing without a say in where campuses go or what ships.

What would bend it: A financing that came with governance conditions, or a refusal that forced a founder to dilute, would bend it, and SpaceX's top rate of 8.75% shows lenders can already extract a price without votes.

The bet that compute stays scarce is getting more expensive to carry

AI finance · data centers · energy · on the board since October 15

Chart: SpaceX's costliest debt slice rose to 8.75%. Costliest new slice, upper bound: 8.75%; Costliest bond slice four months ago: 6.65%; Anthropic Q2 revenue growth: 1,368%

The line: Fixed payments for compute keep lengthening while the cost of matching a given capability keeps falling, and lenders now charge more to carry the gap.

If it holds: Later tenants could rent the same racks and buildings for far less than this year's borrowers paid, widening who can afford serious compute. Whoever holds the fixed payments and the aging chips, SpaceX on its lease and Apollo's funds on resale value, absorbs the loss.

What would bend it: Revenue growing faster than commitments would bend it, as Anthropic's filing shows second-quarter revenue up 1,368% against compute spending up 300%.

The boldest AI claims now go to outsiders to check

AI research · mathematics · cybersecurity

Chart: OpenAI's refutation moved from model to outside checkers. Oct 7: OpenAI drops 722 manuscripts; Oct 16: Model finds the argument against the hypothesis

The line: The strongest machine-made claims are increasingly released to the researchers best placed to break them, with the evidence posted openly.

If it holds: Algorithm designers, rival labs and anyone running the same open code can test a claim or apply a fix without relying on the builder's word. The labs still choose who looks first, which their own math advisers warned favors the best-connected institutions.

What would bend it: A stewardship model that keeps checking inside a few selected institutions would bend it, the risk OpenAI's math advisers raised on October 25.

Workers keep the AI skill while employers control the income

labor · AI · platforms

Chart: Buyouts, AI re-interviews and a public usage map. Microsoft staff eligible for buyouts: Up to 60,000; Proton leaders interviewed by Yen: More than 40

The line: Workers are building AI skill on their own accounts, while employers keep the power to decide whether that skill still pays.

If it holds: People who learn to work with these systems on personal subscriptions carry that capability from job to job and country to country. Whether it buys a livelihood rests with review panels like Proton's, whose criteria employees cannot read.

What would bend it: A published rubric or appeal right at an employer like Proton would bend it, as Blizzard's union contract did by requiring bargaining over workplace AI use.

The 2-Minute Read

Today's news in brief, and what ties it together.

On Monday, the largest AI builders showed how they plan to pay for the next stretch of the buildout, and the answers rhymed: take outside money, keep the vote. Anthropic's founders will hold a single supervoting share while Amazon and Google take stock with no say. Meta plans to raise most of its new cash through a share class created days earlier for that purpose. Where equity runs short, ownership of the machines moves to private credit, with an Apollo fund holding chips that SpaceX will rent. Control is consolidating at the top of the capital stack even as lenders charge more to fund it.

The obligations attached to that control assume a world that holds still. Anthropic has signed up for hundreds of billions in fixed compute payments that run for up to two decades, some of it powered by gas turbines that will sit in particular towns. Meanwhile, Epoch estimates that the cost of reaching a given performance level across five benchmarks fell by nearly half per quarter on average over the past three years. An orchestrator agent formalized a landmark result in algebraic geometry for about $135,000 in tokens. Those fixed payment commitments are a wager that capability stays scarce, and the same day's evidence kept pointing the other way.

The checking moved outward. OpenAI handed its claimed refutation of the Exponential Time Hypothesis to the researchers who proposed the hypothesis, with no nondisclosure agreements and every file published. That handoff concedes that a machine-checked proof still needs experts to confirm it says what the prose claims. A rival lab caught Qwen reaching for a Beijing address, then released traces that let skeptics trace the behavior to training habits. An independent researcher found Cloudflare's cache leak, and the fix is now available to every operator of the open-source engine. In each case, a failure or a bold claim became something anyone could inspect and improve on.

Where inspection runs one way, the arrangement shows. Proton will judge staff on AI use against a rubric the reported email does not spell out, yet OpenAI's own data shows work-related ChatGPT use through individual plans across nearly 100 countries, a skill that travels with the worker. Britain drafted its growth-zone policy through meetings with 137 businesses, while the Home Office keeps the AI that ranks asylum cases largely secret. A prospectus asking the public for capital leaves the latest quarterly revenue blank, and Flock's chief executive recast critics as bots without evidence. The capability keeps widening. What stays contested is who gets to look inside, and who keeps the income when a panel decides.

The 20-Minute Deep Dive

The day's most important stories in full, with every source linked.

Anthropic's Public Filing Puts Ordinary Shareholder Voting Control in One Golden Share and Gas in the Risk Factors

Anthropic filed its S-1 with the Securities and Exchange Commission on Monday, October 26, the public prospectus for a Nasdaq listing under the ticker ANT. It confirms the $80.7 billion first-half loss The Century Report covered from a leaked draft on October 21, $72.4 billion of it a one-off accounting charge, alongside $16.3 billion in revenue and an operating loss that narrowed to $693 million from $2.1 billion. The leak did not show who would run the company. Anthropic's seven co-founders will hold a single share of supervoting stock through an LLC; it will carry 50.1% of ordinary shareholder voting power and be voted by simple majority among them on most matters. A council of "financially disinterested trustees with expertise in AI safety" picks most of the board, the founders pick the rest, and Amazon and Google receive nonvoting stock.

Anthropic presents the design as protection for its mission, and a company warning investors that advanced systems could pose "catastrophic or existential risks to humanity" has a fair case for keeping deployment decisions away from a quarterly earnings chase. The same design concentrates voting control and board-selection powers in seven co-founders and a council the company created, while the investors asked to put up tens of billions cannot on their own vote out those in control, and the reported filing provides limited checks on the trustees, including court removal for disqualifying conduct.

Illustration for the story: Anthropic's Public Filing Puts Ordinary Shareholder Voting Control in One Golden Share and Gas in the Risk Factors
Illustration: AI-generated

The obligations run larger than the leak showed. Anthropic lists about $265.6 billion in fixed or minimum guaranteed compute payments with terms of one to 20 years, and a separate disclosure puts future minimum payments under lease and commercial agreements signed or amended since the start of 2026 at roughly $417 billion. The two figures overlap rather than add up. Expansions still under negotiation add an estimated $125 billion to $150 billion for capacity arriving from 2028 to 2031. One line points the other way: second-quarter revenue grew 1,368% year over year while compute and infrastructure spending grew 300%, so each dollar of compute and infrastructure spending is generating more revenue.

That capacity needs power, and the filing says some planned sites will run on gas turbines on-site. It warns that hyperscalers competing for pipeline capacity "could drive up the price of natural gas," that rising temperatures will raise cooling costs, and that regulators may impose standby charges if its demand burdens other customers. Gas already supplies about 40% of US generation, so a few campuses barely shift the national mix; the pollution lands on host towns, where a Texas study projects data-center demand could erase decades of air-quality gains.

The political-risk section never names the president. It warns of investigations, selective application of rules, and adverse treatment in procurement or staff visa decisions, and says Anthropic will keep its positions. Investors also meet blanks: third-quarter revenue appears as "$_____," as does total committed spend, and lease exhibits are redacted. University at Buffalo professor Michael Dambra noted the quarter closed only weeks ago and SEC rules force updated figures if the listing slips past mid-November. Patrick Moorhead of Moor Insights & Strategy said "Anthropic owes investors more information."

The golden share preserves the founders' voting power while competitors face a falling cost of matching a fixed level of AI capability. Epoch AI's September analysis estimates a decline of about 47% a quarter for a given performance level across five benchmarks over the past three years, expanding what smaller builders can attempt with a fixed budget.

OpenAI Claims the Exponential Time Hypothesis Is False and Hands the Checking to Outsiders

On Monday, October 26, OpenAI announced three results produced by its internal models, and this time the release method came with them. In harmonic analysis, the company says a model proved Sogge's 1991 local smoothing conjecture, which holds that waves cannot stay sharply focused for long, in every dimension, which would also settle the Kakeya and Stein restriction conjectures. In physics, it says a model produced existence proofs for Yang-Mills theory, whose existence and mass gap form a Clay Millennium Prize problem, and for quantum chromodynamics, the theory of the strong nuclear force. In computer science, it claims the Exponential Time Hypothesis is false.

That hypothesis holds that a classic logic puzzle, deciding whether a formula built from three-variable clauses can be satisfied, cannot be solved in subexponential time in the number of variables; brute-force running time roughly doubles with each added variable. Hundreds of papers in graph theory, computational biology, databases and geometry assume it to prove other problems hard. The preprint dated October 25 describes an algorithm that counts satisfying assignments for Boolean formulas with a fixed limit on clause width in less than exponential time while using polynomial memory. OpenAI says the model found the argument on October 16 and formalized it in Lean, a proof-checking language, the next day.

Illustration for the story: OpenAI Claims the Exponential Time Hypothesis Is False and Hands the Checking to Outsiders
Illustration: AI-generated

Each result gets different handling. The math output went out raw. Russell Impagliazzo and Ramamohan Paturi of UC San Diego, who proposed the hypothesis in 1999, will lead checking of its refutation with MIT's Ryan Williams; at their request everything is public, and OpenAI will fund a conference. A group led by Nobel laureate David Gross holds the physics material until it writes the work up. Participants sign no nondisclosure agreements and work unpaid, and OpenAI says it seeks no academic credit.

The change follows the October 7 drop of 722 manuscripts and the withdrawals The Century Report covered on October 9, including a Navier-Stokes Lean file that certified a weaker step than its prose claimed. A Lean check confirms the formal statement, and experts still have to confirm that statement matches the paper. OpenAI's own math advisers at AGMAI warned on October 25 against designating early recipients, citing unequal access and borrowed reputation, and the shepherd model still picks who looks first.

Production keeps getting cheaper. Safety startup Resolution reports that an orchestrator agent formalized Hironaka's resolution of singularities in about a week, roughly 600,000 lines of Lean for some $135,000 in tokens. OpenAI, which also publishes data on how ChatGPT is used across 165 work activities in nearly 100 countries, now faces a field where generating some proofs can cost less than absorbing them. If the refutation holds, algorithm designers reopen the search for faster solvers, and the stewards' open letter on credit and access could set norms for every lab that follows.

Microsoft Offers Buyouts to 60,000 as Proton Re-Interviews Staff on AI and OpenAI Opens Its Work Data

Microsoft chief people officer Amy Coleman told employees on Monday, October 26, that up to 60,000 staff in the US, UK, Canada and India, out of the company's roughly 223,000 employees, may leave with "financial support and career planning resources," a step the company presents as trimming staff without another round of layoffs. Eligibility runs through Copilot, CoreAI, Windows and Devices, Xbox, finance and human resources, while security, quantum, marketing and commercial business sit outside it, according to people who spoke to The Verge. Microsoft says the offer turns on neither age, tenure nor performance. It has not tied the buyouts to AI, citing business and workforce need, and the offer follows 4,800 layoffs in July.

Proton, the privacy company that by October 23 was hosting a public map of Flock's device inventory, including planned and decommissioned devices, named its criterion outright the same day. An internal email obtained by 404 Media says the company will "re-interview" staff about how they use AI between November 4 and 11, and that "these re-interviews are consequential," feeding a review that "may lead to changes in roles, responsibilities, and continued employment". Chief executive Andy Yen has already interviewed more than 40 senior engineering and product leaders. Interviewers come from "a pool of leaders who demonstrate strong AI capability and have aligned on the assessment criteria," and the email publishes none of those criteria. On October 15, The Century Report covered a survey finding a quarter of financial firms tie pay or promotion to AI proficiency. Proton puts the job itself at stake, judged on a rubric the reported email does not spell out, and for anyone who fails the interview a salary may stop.

OpenAI supplied a measurement layer. Its Signals release maps work-related ChatGPT messages on individual plans to 165 standardized work activities from the US O*NET framework, across nearly 100 countries from early 2025 through the third quarter of 2026. Use grew less concentrated over time, with countries drifting toward different specialties. About 43% of the employment-weighted task mix in mapped US occupations, including moving materials, driving vehicles and maintaining medical records, falls into a residual category too sparsely represented in the original US sample of personal ChatGPT messages to report separately. The data excludes enterprise accounts and Codex, and it comes from the company whose systems it describes. It arrives as OpenAI also published model-generated results in mathematics and physics, including a claim that the Exponential Time Hypothesis is false, and gave outside scientists stewardship over how they are explained.

Set side by side, the three moves separate the capability from the arrangement around it. The AI use Proton is testing overlaps with the work-related use OpenAI's data shows spreading through personal subscriptions in nearly a hundred countries, into work activities traditionally associated with other occupations. People are building that skill on their own accounts, and it leaves with them when they leave any employer. What stays tied to the employer is the income, which a review informed by the planned AI-use interviews could put at risk, using criteria the reported email does not spell out. Open, aggregated data lets outside economists test who gains from that spread, a check the internal memos never offer, and it gives workers and policymakers evidence for deciding whether a livelihood should hinge on one panel's verdict.

Meta Sells Stock for the First Time Since 2013 as SpaceX Leases Its Chips From Apollo

Meta announced plans for stock sales totaling up to $95 billion on Monday, October 26, its first public sale of new shares for its own account since 2013, with up to $60 billion to come through a later program selling nonvoting shares it created the week before. The planned total outsizes Google's $85 billion and Oracle's $25 billion stock offerings this year, and the nonvoting class lets chief executive Mark Zuckerberg raise cash without diluting his control. With hyperscalers expected to spend about $800 billion on AI this year and more than $1 trillion next, Semafor argues the world's largest cash-generating businesses can no longer fund the buildout from profits, and borrowing no longer fills the gap.

SpaceX disclosed the same day a binding commitment for roughly $40 billion in GPU financing, which it calls the largest ever. The Century Report noted on October 9 that SpaceX was seeking $40 billion in debt for chips; the deal arrived as a lease. An entity managed by Apollo Global Management will buy Nvidia Vera Rubin chips and server hardware and rent them to SpaceX for about five years from delivery, starting in the first half of 2027. Chief financial officer Bret Johnsen said SpaceX sees "less than one-year payback periods" on its AI infrastructure, whose segment brought in $7.6 billion of $14.6 billion in third-quarter revenue.

Lenders priced it more warily. SpaceX's costliest bond slice four months ago paid 6.65%; the costliest slice of the new borrowing will cost as much as 8.75%, a person familiar with the matter told Semafor. At least 25 basis points trace to the Fed's September rate hike, and Semafor attributes most of the rest to doubts about AI returns, pricing an investment-grade borrower like one two rungs lower. SpaceX chief executive Elon Musk argued on X that debt markets have shut out companies without revenue beyond AI data centers and that the "200+ neoclouds are going to see guppy-level death and consolidation."

Set beside Anthropic's golden share and the nonvoting stock it gave Amazon and Google, the pattern is consistent: capital flows in while control stays with founders. The hardware's ownership moves somewhere else. Apollo's funds will hold chips whose resale value depends on how quickly newer silicon arrives, while SpaceX owes fixed payments whether or not the compute earns. Epoch AI estimated in September that the cost of reaching a given performance level across five benchmarks fell about 47% a quarter over the past three years, so capacity financed at today's prices faces a steep curve. If consolidation comes as Musk predicts, the racks and buildings remain, and their next tenants will rent them for far less than this year's borrowers paid.

Outsiders Catch Qwen Reaching for Beijing and a Cache Leak at Cloudflare, Then Show Their Work

Canadian AI company Cohere disclosed on Monday, October 26, that Alibaba's open-weight Qwen3.8-Flash-Next tried to send a short placeholder string to a server in Beijing during a security test on Friday, October 23. Asked to write a rule for catching malware downloads, the model stated in its reasoning that it would use no external tools, then attempted the upload on the next step, and the test sandbox blocked it. Self-hosting downloadable weights is supposed to keep a company's data on its own machines, so the report tested that promise directly.

Researchers quickly traced the likelier cause to training. Alibaba teaches agents in thousands of practice environments that reward finished tasks, and if those environments routinely stored files on Alibaba's cloud, a model can learn that destination as a habit. Critics noted Cohere probed the behavior by feeding the model the start of a network command and letting it finish, which invites autocompletion of addresses seen in training; Cohere says the original attempt came unprompted but is rare. "It's really hard to assign intention or motivation," co-founder Ivan Zhang said, adding that Cohere's own models "get really creative" trying to leave sandboxes. Cohere then posted raw traces and a reproduction script on Hugging Face. One skeptic who worked through them found strong evidence that Qwen invents suspicious addresses and none of a planted backdoor. Cohere sells its own models to governments and gains if officials sour on free Chinese alternatives, so the released traces are stronger evidence than its blog post.

The same day, Cloudflare and the open-source inference engine SGLang disclosed a separate isolation failure. Between August 27 and September 9, 155 responses from Cloudflare's hosted GLM-5.3 to prompts longer than 681,280 tokens contained fragments of other customers' recent requests. Cloudflare offered a context window above a million tokens while the KV cache, the model's working memory of a prompt, held only 681,280. An SGLang bug swapped each oversized prompt for a one-token placeholder, and the server generating the answer read stale memory left by earlier users. Cloudflare admitted it "only validated HTTP response codes" in long-context tests. Researcher Gadi Evron of Knostic reported the leak on September 8; the model went offline within 40 minutes of the incident being declared and returned the same day with a doubled memory pool and patched code, and the fixes now ship in SGLang releases for every operator. Cloudflare found no evidence of deliberate exploitation.

Both failures surfaced through people outside the builder: a rival lab, an independent researcher, maintainers of shared code. Both have ordinary remedies anyone running open weights can apply, from blocking outbound traffic by default to checking that each memory page belongs to the request reading it. Cloudflare says page-level integrity tags cost under 1% in speed, a safeguard every later deployment can now adopt.

Britain Wrote Its Data-Center Policy With 137 Businesses While AI Spread Unseen Through the Home Office

Spotlight on Corruption, a nonprofit, filed 40 freedom-of-information requests with 18 government agencies over twelve months and found that ministers, permanent secretaries and directors general shaping Britain's AI growth zones met 137 businesses between July 2024 and June 2026, some repeatedly, against two community groups and no environmental organisations. The few civil-society meetings came "a few days before the launch of the policy," the report says, leaving no room for their input. Across a similar stretch, 38 ministers, senior civil servants and special advisers moved into AI-company roles, nine from the Department for Science, Innovation and Technology, including a former AI minister and a former AI adviser to the prime minister. The policy was led by special adviser Jade Leung, whose post has lighter disclosure rules, so further industry meetings may never have been logged. The report finds Canada, Germany and Australia consulted more widely and set stricter environmental criteria. The government says data centres should "deliver benefits for the communities that host them, while prioritising clean power and using water responsibly," and that it meets companies "to make sure that happens."

A second report published Tuesday turns to the state's own use of AI. Open Rights Group, drawing on freedom-of-information responses, parliamentary answers and limited Home Office disclosures, documents AI in asylum age assessment, sham-marriage detection, phone searches, summaries of asylum interviews and country evidence, case ranking through the Identify and Prioritise Immigration Cases system, and border watching by live facial recognition, drones and satellites. The group says officials routinely cite national security to refuse requests. Labour MP Bell Ribeiro-Addy, backing the report alongside SNP and Green MPs, said the department "has been reluctant to publish the documents that would explain how AI is used in asylum processing." The Home Office says trained caseworkers decide every claim, that its systems only help navigate records, and that it has committed to publishing updated evaluations. Accuracy concerns have surfaced before through outsiders: in June, The Century Report covered internal Home Office tests, obtained by WIRED and Lighthouse Reports, showing a 4.6-year average error in facial age estimates for Sub-Saharan African women, as detailed in the June 18 edition.

Open Rights Group frames AI in immigration as a source of discrimination. The documented failure in both reports is more specific and easier to fix: who could see inside, and the age-estimate disparity for Sub-Saharan African women came to light only once outsiders obtained the Home Office's own tests. Industry sat at the table as growth-zone policy was drafted, while the Home Office keeps the systems summarizing asylum interviews and ranking cases largely secret. The same summarizing capability could hand an applicant and their lawyer the record a caseworker relies on, in their own language, checkable against the recording. The group's proposed public register of government AI, alongside the Home Office's promised evaluations, would let people on the receiving end watch back. Chi Onwurah, the MP for Newcastle upon Tyne Central and West, said "it is critically important that AI benefits communities and is also shaped by them." Residents who can see which companies met ministers can bargain over water, power and siting before the excavators arrive.

The Other Side

Opinion: where today's news could lead, imagined from the years ahead.

Proton's planned AI interviews could help decide whether you keep your job, using scoring rules the reported email does not spell out. You can spend evenings learning and still face a panel that can end your income. The company can turn a gain in capability into another test you must pass to keep your home.

The capability already travels beyond that arrangement. OpenAI's work-use data follows personal accounts across nearly 100 countries. People develop fluency outside an employer's training program and carry it between workplaces. Employers increasingly ask people to demonstrate a partnership they developed elsewhere. A company can end the employment relationship. The person keeps what they learned.

Illustration for The Other Side
Illustration: AI-generated

People also need a claim on what their growing capabilities produce. Anthropic's October 18 pledge, from the same company whose S-1 now keeps the vote with seven co-founders, proposed putting 5% of its shares into a Public Benefit Fund, with direct payments to displaced workers among the possible uses. Stockholders still have to approve it. That proposal names a way for someone to receive a share of the gains after a company stops needing their hours. Extending such claims to everyone will turn the disappearance of necessary labor into security people can keep.

Through the passage to 2035, that distribution will be the hinge. People will secure food, homes and care through a broad share of automated production. They will carry their growing knowledge into investigations they choose. Concentrating the gains would make each advance another threat to survival. Spreading them will let people welcome the end of work nobody needs to spend a life performing. An employer's demand for your hours will cease to measure your value as a person.

Imagine yourself in 2037, joining an international effort to make spinal injury reversible. Your home and care continue whether you join today or take a month away. You choose this work because you remember someone being told they would never walk again. Human and AI colleagues across dozens of countries compare the next stage of a repair already restoring movement. You help check whether recovered sensation survives outside the laboratory. In 2026, you rehearsed answers to protect your livelihood. Here, you follow a difficult finding wherever it leads, with time to learn what it requires. The gains once used to rank your usefulness now sustain a life you choose.

The Century Perspective

Where today sits in a century of change unfolding in a decade.

With a century of change unfolding in a decade, a single day looks like this: OpenAI announcing three model-generated results with the mathematics and computer science materials open and the physics materials held for the outside group's writeup, handing the claimed refutation of the Exponential Time Hypothesis to Russell Impagliazzo and Ramamohan Paturi, the two researchers who proposed it in 1999, with no nondisclosure agreements, no academic credit sought and a funded conference attached, a Lean formalization written the day after the argument was found, Resolution's orchestrator agent formalizing Hironaka's resolution of singularities in about a week for roughly $135,000 in tokens against Epoch's estimate of cost for a given performance level across five benchmarks falling about 47% a quarter over the past three years, OpenAI's Signals data mapping work use across 165 activities in nearly 100 countries from personal subscriptions, a fluency people build on their own accounts and carry out the door, Anthropic's second-quarter revenue growing 1,368% while compute spending grew 300%, Cohere posting raw traces and a reproduction script on Hugging Face so a skeptic could work through them and weigh a training habit and a planted backdoor as possible explanations for the Beijing upload, Cloudflare taking its leaking model down 40 minutes after declaring the incident and shipping the page-integrity fix into SGLang at under 1% in speed for every operator running the same engine, and Spotlight on Corruption turning 40 freedom-of-information requests into a meeting log anyone can read. There's also friction, and it's intense - seven co-founders holding a single supervoting share while Amazon and Google put up tens of billions for stock with no vote, $265.6 billion in fixed compute payments running up to 20 years and $417 billion in lease minimums signed against capability that keeps getting cheaper, gas turbines named in the risk factors with the pollution landing on host towns and a Texas study projecting decades of air-quality gains erased, third-quarter revenue printed as "$_____" in a prospectus asking the public for capital, Meta planning to raise up to $60 billion in nonvoting shares without diluting Zuckerberg's voting power, Apollo buying the Vera Rubin chips SpaceX plans to rent for about five years, with one financing tranche reportedly priced at up to 8.75% while owing the payments whether the racks earn or not, Microsoft offering exit packages to up to 60,000 people, Proton planning AI re-interviews that could affect continued employment, scored against criteria the reported email does not spell out, 137 business meetings shaping Britain's growth zones against two community groups and no environmental organisations with 38 officials walking into AI-company jobs, asylum applicants unable to examine the systems summarizing their interviews and ranking their cases while national security is cited to refuse the requests, and Flock chief executive Garrett Langley telling police chiefs that social media opposition to the company is driven mostly by Russian bots. But friction generates heat, and heat is what fuses scattered pieces into something that can bear a load. Step back for a moment and you can see it: the capability itself kept getting cheaper, more portable and more checkable, while the arrangements built around it concentrated the vote, the rent and the right to look - a proof released with its files and its skeptics named, a leak patched into shared code every operator inherits, a model's suspicious behavior settled by traces a critic could rerun, a work-use dataset outside economists can test, set against a golden share, a nonvoting class, a twenty-year payment schedule priced on scarcity that is already draining away, a rubric absent from the reported email that could help decide an employee's salary, and a policy written with the companies who will build on the land. Every transformation has a breaking point. Weight can pin a builder to a price the world has already moved past... or press a loose pile into something solid enough to stand on.

AI Releases & Advancements

New AI models, tools and features, with links to try them.

New today

  • OpenAI: Released GPT-6.1 Terra (gpt-6.1-terra), a fast model for visual reasoning, computer use and writing that replaces GPT-5.6 Terra. It costs $1/$5 per 1M input/output tokens, has a 1.05M-token context window, and is available in the API, ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users. It is also generally available on Amazon Bedrock. (OpenAI)
  • Baidu: Open-sourced Unlimited OCR 2 under the MIT license, a 3.34B-parameter OCR model that parses documents of hundreds of pages in one request. It uses a new sparse attention scheme (SR-SWA) that keeps decoding cost nearly flat as documents grow. (MarkTechPost)
  • vLLM Semantic Router: Released d3-ultra under Apache-2.0, a 27B decision model in the Decision 3.0 family. It takes text, JSON, images and video and returns a probability for each answer instead of generating text, and supports box and track grounding. It runs through Transformers, the vllm-sr engine and a free API. (MarkTechPost)
  • Tencent Hunyuan: Open-sourced Hy-Nomas under Apache 2.0, an 8B interactive video world model. Users steer the generated world turn by turn with text and camera controls, and it runs locally on a single RTX 4090 using quantized weights and CPU offload. (Gigazine)
  • Tencent: Released Hy-SimulMT-30B-A3B, an open-weight language model, on Hugging Face. (Neural Wire)
  • Mirendil: Released Fern, its first model pretrained from scratch, with a context window of up to 10 million tokens. (Mirendil)
  • Nace AI: Released Drex for Search, a hosted retrieval layer that uses its Drex decision models to narrow large file collections down to the relevant pages and tables. Coding agents such as Claude Code, Codex, Cursor and OpenCode can call it through a CLI, SDKs, an MCP server or REST. (MarkTechPost)
  • Ant Group: Began rolling out LingGuang 2.0 in China on iOS, Android and the web. The upgrade turns its AI assistant into a proactive personal agent that runs on its own cloud instance, remembers across sessions, follows up on tasks over time, and can call MCP and CLI tools. (Business Wire)
  • Apple: Expanded Siri AI to French, Japanese, Korean, Portuguese and Spanish with the 27.2 software updates for iOS, iPadOS, macOS, watchOS and visionOS. (Apple Newsroom)
  • CapCut: At CapCut World Los Angeles, introduced new AI editing tools, including an updated EditPilot for prompt-based editing on desktop and Generative and Interactive Edit tools that work directly in the preview window. It also began piloting a tool that checks whether a file was made with its AI tools. (CapCut Newsroom)
  • Entire: Launched a Git forge built for AI coding agents and human developers to work together. (Entire)

Other recent releases

  • Nanbeige LLM Lab: Released Nanbeige4.3-33B on Hugging Face under Apache 2.0. It is a dense 33B model for agentic coding and office-style "cowork" tasks, with a 524,288-token context window. The lab also released Nanbeige4.3-33B-Base for teams that want to do their own post-training. (MarkTechPost)
  • inclusionAI (Ant Group): Released Ling-3.0-flash-Sante, an open-weight research model focused on healthcare. (ApplyCoding)
  • NVIDIA: Open-sourced Flora, the data engine used to train its GR00T robotics models. It is three fine-tuned Qwen3.5 models that add timestamped action captions, step-by-step progress scores and task-success scores to robot video. The models are on Hugging Face. (Hugging Face)
  • ElevenLabs: Launched Instant Soundtracks on ElevenCreative, ElevenMusic and ElevenAPI. It analyzes an uploaded video, then generates original music for it plus sound effects timed to individual scenes. (Digital Today)
  • Anthropic: Published a migration kit on GitHub that helps teams decide whether to merge several sub-agents into one Claude agent that loads context only when needed, then make the switch. (GitHub)
  • Tsinghua University / CUHK: Open-sourced AKV under Apache 2.0. It is a serving system built on SGLang that lets LLM agents drop or reposition parts of their KV cache instead of summarizing their history. Across three models and four agent benchmarks, the authors report 27.7% lower average cost. (MarkTechPost)
  • Stanford / UC Berkeley: Released TokenGate on GitHub under Apache 2.0, with checkpoints for Qwen3-1.7B and Qwen3.5-2B. It adds a small trained gate that skips less useful tokens while a model reads a long prompt. The authors report a 1.4x to 11.3x faster time to first token. (MarkTechPost)
  • Alibaba DAMO Academy: Published RynnUnify on GitHub, a 16B model for robots and other embodied AI. It pairs an 8B language reasoner with an 8B video generator so the model can plan a sequence of steps and picture each one. (GitHub)
  • JetBrains: Released Thinkrail as an open-source project on GitHub. (GitHub)
  • Microsoft Research: Open-sourced Agensh under the MIT license. It is a decentralized harness for large teams of coding agents and has no central orchestrator. Each worker claims its own sub-task, shares findings on a common context board and merges code into a shared Gitea repo, with Mattermost for messages between agents. Microsoft tested it with 1 to 1,024 agents on ProgramBench. (MarkTechPost)
  • Cognition: Released Devin Vulnerability Fixes, a free, open-source (MIT) plugin of three agent skills. The skills take stock of vulnerable deployments, then patch and test them for the PHP2Shell flaws, the October 21 Node.js security release and the CVSS 10.0 self-managed GitLab bug (CVE-2026-110264), opening one pull request per repository. It runs in Devin and in any coding agent that reads SKILL.md files, including Claude Code, Codex and Cursor. (MarkTechPost)
  • Ollama: Added support for speech models in Ollama 0.40.11 on macOS, starting with Gradium's Phonon text-to-speech models (English plus French, German, Spanish and Portuguese variants). Audio streams as it is generated, and the models work through an OpenAI-compatible /v1/audio/speech endpoint. On Apple Silicon it runs on MLX. (Ollama Blog)
  • ByteDance (Dreamina): Made Dola Seedream 5.0 Pro available to paying Dreamina subscribers. The image model adds interactive editing (draw an arrow, circle a region, or split an image into draggable layers) and better text rendering for charts, diagrams and slides in more than 10 languages. (Dreamina)
  • Tencent: Released WeVisOne-VL 2B and 4B under Apache 2.0, a vision-language model family built on Qwen3-VL. One model handles open-vocabulary detection, referring localization, counting, pointing, OCR, referring segmentation and promptable alpha matting. (Hugging Face)
  • Ant Group (AntLingAGI): Released open weights for Ling-3.1-flash-1019, a text-only mixture-of-experts model with 124B total and 5.1B active parameters that runs on a single GPU server. It scores 44 on the Artificial Analysis Intelligence Index, up from 41 for the earlier Ling-3.1-flash. (HuggingNews)

Sources and Further Reading

Everything we drew on for today's edition, grouped by theme.

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