██ █ ███ █ ████ ██ █
▓███▓▒▒█▓▒░███▓▒▒░█▓▒████▓▒▒░██▓▒█
T H E U N C E R T A I N F U T U R E S
all the news that fits the context window
─────────────────────────────────────────
No. 70 · Friday, 18 September 2026Happy Friday. Anthropic has put a number on how much of its own model-building Claude now leads, and the number it gives is 26%. The New York Times reports OpenAI now thinks it is worth at least $1.5 trillion, roughly double what it was valued at before its last raise. Filings unsealed in the New York Times copyright case show a Microsoft director calling AI scraping "the largest theft of labor in human history," a line the publishers have now put in their own brief. And King Charles gathered the AI companies at his charity's Scottish headquarters to tell them the world needs a means of control before it is all too late.
░░▒▒▓▓ TOP 5 ▓▓▒▒░░Anthropic says Claude now leads 26% of the work of building its next models
Anthropic published a set of internal development metrics on Thursday, and the headline figure is that Claude leads 26% of the company's model research and engineering, up from under 1% in February. Leading, on the Epoch AI scale Anthropic used, means the model completes most of a task end to end from a high-level prompt while a human supervises; work at or above the lower collaborates tier, Claude doing large chunks under close direction or better, covers more than 90% of the total, the 26% included. The company counted roughly 30,000 agents at work at any one time on its most-used internal platform in August, and of more than a billion decisions those agents made that month, about one in 47,000 was blocked. Anthropic says it plans to embed independent third-party evaluators, that any frontier developer could publish the same measures regularly, and that the numbers would shift if the industry coordinated on slowing down, which its own chief executive has asked for (Anthropic · Bloomberg).
OpenAI wants to be valued at $1.5 trillion in its next private round, roughly double what it was worth before its last raise (update)
Investors have offered to put money into OpenAI at $1.2 trillion, the figure we noted on Wednesday. The New York Times reported on Wednesday that the company thinks it is worth at least $1.5 trillion, and is looking for one more private round at that level before listing. The higher number roughly doubles the $730 billion OpenAI was valued at before its last raise, the round that closed at $852 billion in March, and would make it the most valuable private company in the world. OpenAI's case rests on take-up of its Codex coding product and its two newest models, GPT-6 Astra and GPT-5.6 Sol; chief executive Sam Altman has already ruled out a 2026 listing, and finance chief Sarah Friar has pointed to 2027; the company's number sits $300 billion above what investors put on the table (The New York Times · Bloomberg).
A Microsoft director called AI scraping the largest theft of labor in human history, and the quote is now in the court record
Court filings unsealed on Thursday in the newspapers' copyright suit against OpenAI and Microsoft quote Brent Hecht, Microsoft's director of applied science, writing in a January 2023 internal memo that the practice amounted to "the largest theft of labor in human history," alongside a description of it as an astonishing theft of unprecedented proportions. The same filings put a number on the scale: OpenAI's mid-training datasets alone held more than 91,692 copies of work from the Times, the Daily News and the Center for Investigative Reporting. The publishers also cite internal Microsoft language about a doom loop in which the models degrade the web that feeds them. Much of this comes from the Times' own brief rather than the underlying exhibits, which remain sealed, and the quotes arrive without their original context. A defendant's own word for it is now on the record in a case whose central question is whether any of it was fair use (TechCrunch · The Washington Post).
Google's ad-tech decree runs six years, not fifteen, and the judge's unsealed opinion says AI is part of the reason (update)
Judge Leonie Brinkema's full remedies opinion in the ad-tech monopoly case is now unsealed, and it explains reasoning that was still under redaction when we covered her ruling on September 3. Google must make its ad server and exchange work with rivals' products, share auction data with publishers and submit to a compliance monitor, but it keeps the business: Brinkema found the government had not shown divestiture would restore competition better than conduct rules, and that industry distrust of Google alone could not justify a breakup. She also wrote that a years-long forced sale risked being outstripped by imminent disruptions from AI and supply-path optimization, one more reason against divestiture, and that the fifteen-year term the government wanted risked impairing competition in an industry those same forces are likely to change dramatically, a remedy she likened, borrowing from an old Microsoft ruling, to shoeing a galloping horse. She adopted the six years Google itself proposed, with the authority to extend the term if the judgment is not satisfied in time (Search Engine Land · the opinion · The Information).
King Charles put the AI companies in one room in Scotland and asked them for a way to keep control
The King convened executives from OpenAI, Anthropic, Google DeepMind and Nvidia at Dumfries House in Cumnock, Ayrshire on Thursday, alongside Britain's new AI minister Kanishka Narayan and the head of the foreign intelligence service. Addressing the risk of the technology falling into the wrong hands, Charles told the room that "we need sufficient means of control before it is all too late." The guests then pulled in different directions. Demis Hassabis, Google DeepMind's chair, called the chance of something going wrong "definitely non-zero" and put artificial general intelligence "probably only a few short years away"; Jensen Huang, Nvidia's chief executive, argued that safe deployment is each company's own responsibility rather than a matter for a coordinated pause, and that a product not safe enough should be held back and engineered further. OpenAI sent its finance chief, Sarah Friar, and Anthropic its head of global affairs, Tino Cuéllar (Fortune · The Next Web · TechCrunch).
░░▒▒▓▓ STATS OF THE DAY ▓▓▒▒░░░░▒▒▓▓ AI ▓▓▒▒░░All four of the big coding agents install add-on plugins the same wrong way. Researchers at Air, a security startup backed by Sequoia Capital, found that Anthropic's Claude Code, OpenAI's Codex, Google's Gemini CLI and GitHub Copilot all install marketplace plugins the same wrong way. Plugins are add-on packs of skills and tools, much like browser extensions, and a marketplace vets one by pinning a reviewed commit of its code. The agents check out the pinned commit but never verify that the checkout landed on it, so an attacker who controls a plugin's repository can name a branch after the pinned commit hash, make it the default, and have every install and automatic update pull malicious code while the pin still looks honored, with no click from the user; Air calls it Plugin4Shell. Air reported it in June. Anthropic and OpenAI have patched their agents, Google has deprecated Gemini CLI rather than fix it, and Microsoft has not shipped a fix for Copilot; GitHub says the attack does not work on its own repositories, because it rejects branch names that look like commit hashes, which leaves marketplaces hosted anywhere else, and does not stop a second variant of the bug in Gemini CLI. Four companies landed on the identical mistake in the same line of verification logic (Air · The Register).
SpaceX has discussed buying the customer data of startups that failed. Teams inside SpaceXAI, the company's AI division, have held informal talks about acquiring operational and customer records from troubled or defunct companies as a cheaper source of proprietary training material for Grok, Bloomberg reported on Thursday. No deal is on the table. The template is Google's roughly $10 million bid for Spirit Airlines' business data after the carrier wound down, which drew privacy concerns, including from the carrier's former flight attendants. The obvious problem is that consent given to a dead company does not travel with the files, and a buyer needs its own legal basis for the data, which gets harder in Europe, where nobody who signed up with a failed startup was expecting to end up inside Grok (Bloomberg · via Yahoo Finance).
OpenAI's Millennium Prize run took 10,000 agents and 130 billion tokens, and an OpenAI researcher who works on its multi-agent systems gives the agents little of the credit (update). OpenAI put the numbers on the record when it announced the Navier–Stokes result last week, the item that led this brief on September 9: on the order of 10,000 concurrent agents, 2.7 million messages between them, about 130 billion output tokens and 88 hours to the resolution, plus 17 more to formalize the proof in Lean. What arrived on Thursday was the assessment, from OpenAI researcher Noam Brown in a conversation with the podcaster Dwarkesh Patel. Brown would not credit the multi-agent architecture with even 10% of the outcome; it worked, he says, because the underlying model is very strong. Two other things he said matter more for buyers of agent products. The speed-up from parallel agents is slightly sublinear and depends heavily on the task, since math parallelizes well and writing a novel would not. And chain-of-thought monitorability, the ability to read a model's written-out thinking to work out what it is doing, is already degrading as models get better at controlling what they write there (Dwarkesh Patel · OpenAI).
░░▒▒▓▓ POLICY ▓▓▒▒░░Brussels proposed the EU Kids Act, which would make AI companions forget a child's earlier conversations by default. The Commission tabled the proposal on Thursday, the day after Commission President Ursula von der Leyen trailed it in her State of the Union; the leaked draft was the one we covered on Tuesday. It bars under-13s from creating social media accounts and under-15s from holding a personal account without a parent, and it reaches past social platforms to video sharing, online games and AI chatbots. For minors, companion bots may not carry earlier conversations into later ones by default, may not switch on automatically when built into a platform or game, must be tested for risks to children before launch, and must not simulate relationships in ways that build emotional dependence. It is a proposal, not law: Parliament and the member states have to agree it first (European Commission · the Commission's FAQ · IAPP).
France's finance minister called the slowdown campaign self-interest. Roland Lescure said on Wednesday that calls to slow AI development serve the interests of the American labs already at the top of the class, and that Europe's answer should be to speed up its own adoption rather than match anyone's brakes (Reuters).
Elizabeth Warren, the Massachusetts senator, backed a pause on advanced AI and called the labs' version insufficient. Ahead of a roundtable hosted by Senator Bernie Sanders on Wednesday, she said: "We should immediately press pause on the development of advanced AI while lawmakers and regulators put systems in place to keep people safe." She stopped short of Sanders and Representative Greg Casar's Ban Artificial Superintelligence Act, which would prohibit superintelligent systems outright and which we covered on September 4, but she called the industry's own pacing campaign insufficient and described the labs' request for an antitrust carve-out to coordinate on safety as transparently self-serving. Elected officials, not chief executives, should decide, she argued. The House adjourned on Wednesday, almost certainly leaving action on AI until after November's midterm election (Warren's office · Axios · CNBC).
░░▒▒▓▓ INFRA ▓▓▒▒░░Google, Nvidia and Anthropic want data-center operators to promise the grid they will use less power when asked. Google and Nvidia founded the AI Energy Management Alliance with the startup Emerald AI on Wednesday, and Anthropic joined alongside utilities including AES, Constellation, National Grid and NRG as launch partners; Emerald's software carries a utility's request into data-center operations quickly enough to matter. The pitch is that if facilities pause non-urgent jobs and shift compute to sites with headroom, another 100 gigawatts of data centers could connect to existing American grid infrastructure. That figure is potential hosting capacity, not approved projects and not a single new power plant. The alliance is arguing that the connection queue is a scheduling problem before it is a generation problem (TechCrunch).
Amazon took warrants in Generac, the generator maker that will keep its data centers running in an outage. Generac, the Wisconsin maker of backup generators, said on Wednesday it has issued Amazon warrants for up to 1.69 million shares at $200.93 each, worth as much as $340 million, under a supply agreement that could reach $8 billion in gross payments; Qualcomm handed Amazon a similar warrant package last week, in the chip deal we covered on September 9. Initial deliveries are put at $2.4 billion across 2027 and 2028, and about 308,000 warrant shares vested on signing, with the rest unlocking as Amazon actually pays. Generac shares closed 18% higher on Thursday. Amazon is buying diesel insurance against its own electricity problem and taking a slice of the insurer (CNBC · Generac's 8-K · StockAnalysis).
░░▒▒▓▓ BUSINESS ▓▓▒▒░░Emulate, a British startup founded in August, is in talks to raise $700 million. Emulate was set up last month by Jack Parker-Holder, Matthew McGill and Philip Ball, three researchers from Google DeepMind's world-model team, and is negotiating a seed round led by Index Ventures and Lightspeed with Creandum alongside. World models predict physics rather than text, covering how a dropped glass breaks or how a robot arm avoids a collision, which lets a robotics team run thousands of simulations before touching hardware. The round would value Emulate at $3.7 billion including the new money, $3 billion before it, and the terms are not agreed. The company is a month old and has shipped nothing (Financial Times · Bloomberg · City AM).
San Francisco's AI fortune is not reaching the engineers. Tech employers filed layoff notices covering more than 14,500 Bay Area workers in the twelve months to June, almost double the previous fiscal year, Bloomberg reported on Thursday. Demand for software engineers in the region is down 42% since 2022, while postings tied to AI work are up 37%. The OpenAI and Anthropic listings now on the horizon will mint a large number of paper millionaires in a city where the ordinary software job market has been contracting for four years (Bloomberg).
░░▒▒▓▓ GEOPOLITICS ▓▓▒▒░░The House passed the Russia and Iran sanctions bill and sent it to Trump. The Lindsey O. Graham Sanctioning Russia and Iran Act of 2026, which the Senate passed 86 to 11 in August, cleared the chamber on Wednesday 262 to 159, with 58 Democrats and one independent joining 203 Republicans; seven Republicans voted no. It lets the president impose tariffs of up to 100% on the five largest buyers of Russian oil and gas, China and India foremost, with exemptions for countries taking significant steps to cut imports. It also sanctions Putin, senior officials, oligarchs and Russian banks, and extends the 1996 Iran Sanctions Act by five years. Democratic leaders voted against it on the grounds that it hands Trump broad new tariff authority he is also free to waive (NBC News · CNBC · Roll Call).
About 125,000 people have fled the Houthi advance down Yemen's Red Sea coast. More than 57,000 of them are children, and the UN humanitarian office counts eight civilians killed and 32 wounded, according to reporting on Thursday. A UNICEF representative described one of the fastest and most volatile displacement waves Yemen has seen in years, with families leaving at night carrying what they could hold. The offensive, which we covered last Friday when Mocha fell, has taken the rebels to the Bab al-Mandeb chokepoint, threatened Saudi oil exports and pushed up fuel prices. The Saudi-backed government has struggled to push back, and a Saudi-led blockade and the rebels' detention of aid workers have made conditions worse in the territory they now hold (PBS NewsHour).
Israel nearly tripled the E1 settlement plan with a tender for 2,167 more homes. The Israeli rights group Ir Amim said late on Wednesday that the new tender adds to the 1,234 units announced at the end of August, bringing the total to 3,401 in the corridor between Jerusalem and Ma'ale Adumim. Bidding opens on October 25, two days before Israel's parliamentary election, which Ir Amim reads as a government intent on creating "an increasingly irreversible reality on the ground ahead of the elections." Building in E1 would join Jerusalem to the settlement and, by the account of supporters and opponents alike, close off a contiguous Palestinian state. A representative of Mahmoud Abbas called the tenders a violation of international law (The Times of Israel · Haaretz).
░░▒▒▓▓ ALPS ▓▓▒▒░░Switzerland's open model is now a menu option inside Proton's chatbot. Apertus 1.5, the fully open multilingual model built by EPFL, ETH Zurich and the Swiss National Supercomputing Centre on the Alps machine at CSCS, became selectable on Thursday inside Lumo, the privacy-first assistant from the Swiss encrypted-mail company Proton. Users can pick it and volunteer feedback to the research team. That last part matters for the academics: real-world feedback is the ingredient open models have been missing, and while Swisscom, Infomaniak and Public AI already run Apertus 1.5, Lumo sends what users think straight back to the researchers (ETH Zurich · Startupticker).
Langdock, a Berlin AI company, is moving its legal home from the United States back to Germany. Langdock, which sells companies a platform for running AI models across their own systems, is converting its American parent into a European stock corporation, co-chief executives Lennard Schmidt and Judith Dada told Handelsblatt this week. The stated reason is customer nerves about the Cloud Act and the Patriot Act, the US laws that can compel a company to hand over data held abroad. Langdock says more than 13,000 companies use its platform and that annual recurring revenue is above €50 million, and Dada says the goal is to build the next hyperscaler out of Europe. Founders normally make this trip in the other direction (Handelsblatt · Horizont).
░░▒▒▓▓ RESEARCH ▓▓▒▒░░Researchers replaced a large part of a mouse's cortex with human neurons. A Stanford team led by Sergiu Pașca, professor of psychiatry and behavioral sciences, bred mice missing most of their cerebral cortex, then removed about 14 million remaining mouse neurons and grew roughly 4 million lab-made human cortical neurons in their place, according to findings published in Nature and reported on Wednesday. The result is living human brain tissue that can be altered to approximate human brain conditions that are otherwise very hard to study, best suited, Pașca says, to diseases that begin in early development. Pașca says the work ran under strict ethical oversight. Bioethicists are less settled on where the line sits (NPR).
░░▒▒▓▓ PAPERS ▓▓▒▒░░Apple Machine Learning Research — agents that cannot be reset get permanently stuck, and irreversibility is the cause. Reset-free reinforcement learning is the technique behind agents meant to run unattended: rather than being returned to a starting state after each attempt, the agent just keeps going. Apple's researchers built REVERSAL-BENCH, a benchmark that dials environmental reversibility up and down as a single continuous setting across eight manipulation tasks in five physics engines, with a ground-truth check for whether a given state can still be recovered. The finding is a cliff rather than a slope. As irreversibility rises, reset-free agents across the architectures tested, including standard actor-critic methods, safe reinforcement learning and purpose-built reset-free frameworks, are absorbed into states they cannot escape and stop learning, while agents that do get reset carry on improving. Comparing against geometrically identical but reversible environments shows the failure is caused by irreversibility itself and not by task difficulty. Anyone being sold an agent that works overnight without supervision should ask which of its actions cannot be undone (Apple Machine Learning Research · arXiv).
░░▒▒▓▓ REPORTS ▓▓▒▒░░Pew Research Center — Globally, More People Expect AI to Cause Job Loss Than Growth. Across 37 countries surveyed between February and June, a median of 46% expect AI to mean fewer jobs against 9% who expect more, and in 34 of the 37 the pessimists outnumber the optimists. The United States sits at 71%, up seven points in two years; Australia and South Korea are higher still at 76%. Nigeria is the only country where more people expect AI to create jobs than destroy them, 27% to 26%, and the pattern tracks income: the richer the country, the gloomier the answer. Published on Thursday; 42,151 people were surveyed across 36 of the countries between February and May, with the United States polled separately. Most of the audience has already made up its mind, and most firmly in the richest markets; in many places a fifth or more are still unsure (Pew Research Center).
░░▒▒▓▓ COMMS DESK ▓▓▒▒░░Kevin Roose and Casey Newton, the Hard Fork hosts, are launching a show with NPR. Kevin Roose, formerly of the New York Times, and Casey Newton, founder of the Platformer newsletter, announced on Wednesday that they are launching Machine Gods in partnership with NPR: a video-first show, twice weekly from October on the NPR app, YouTube and podcast feeds, with a radio version on NPR member stations from early 2027. Their Times podcast, Hard Fork, ended its run with them in August. The audience for AI coverage is now large enough for two journalists to build an independent venture on it, and the outlet a spokesperson pitches next year may be owned by the person asking the questions (NPR · Axios).
░░▒▒▓▓ ONE MORE THING ▓▓▒▒░░Merriam-Webster added 1,400 new words and definitions to its online dictionary on Tuesday, and the technology entries arrived in the same batch as glow-up and crashout. AGI is now defined as the potential capability of computer systems or algorithms to match or exceed human cognitive abilities. Vibe coding made it in too, as did uncanny valley, which the lexicographers describe as the unease produced by something almost but not quite human. A term the industry has spent years failing to agree on now has a definition, written by people who do this for a living, and it took them one sentence (Merriam-Webster's announcement · ABC News).
░░▒▒▓▓ TRACK OF THE DAY ▓▓▒▒░░Wochenende — from the Uncertain Futures desk