░▒▓███████████████████████▓▒░
▓█▒ ▒█▓
▓█▒ ▒█▓
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. 59 · Thursday, 3 September 2026Hello, and welcome in. Google, Meta and Alibaba each shipped a model on Wednesday, one day after the Anthropic upgrade we covered yesterday. The Justice Department told a federal court that training AI on copyrighted writing is fair use, the first time Washington has taken a side in those cases. A judge then let Google keep the advertising business the government spent years trying to break up. Broadcom and Uber close the day out, one adding tens of billions to its forecast, the other subtracting 3,300 jobs.
░░▒▒▓▓ TOP 5 ▓▓▒▒░░Google, Meta and Alibaba all shipped models on Wednesday, and Google's is the one aimed at Anthropic's lead in coding
Google DeepMind released Gemini 3.8 Flash on Wednesday along with Flash Cyber, the same model with looser cybersecurity restrictions, handed out case by case through a program called Fairwind. Pricing is unchanged from 3.7 Flash at $0.75 per million input tokens and $3.75 out, an introductory rate that doubles on January 1, and The Wall Street Journal reported that engineers testing inside Jetski, Google's internal coding tool, preferred it to Anthropic's Opus; Anthropic's own upgrade, which we covered yesterday, landed a day earlier. Meta put out Muse Spark 1.3 the same day, which it says makes about 20% fewer tool calls and burns about 25% fewer tokens than its predecessor on coding tasks, on the company's own comparisons, with the highest reasoning setting withheld pending safety testing. Alibaba's Qwen3.8-Max-0902 took first place on Code Arena's WebDev leaderboard, which ranks models on building working web apps, at 1691 points, and a day earlier World Labs, the startup run by the Stanford AI researcher Fei-Fei Li, unveiled Atlas, which turns a handful of phone clips into a 3D scene you can walk a camera through. The quiet weeks after the Hugging Face incident are over (Google DeepMind · 9to5Google · WSJ · Meta · Axios · TechNode · World Labs).
The Justice Department told a federal court that training AI on copyrighted writing is fair use
The department filed a statement of interest on Tuesday, a device the government uses to state its position without joining a case, with Judge Sidney Stein, who is hearing the copyright suit brought against OpenAI by The New York Times, the Chicago Tribune and a group of authors including George R. R. Martin and John Grisham. Its argument is that copying works to train a model is transformative and therefore fair use, that the right test is whether the output substitutes for the original rather than whether readership falls, and that rules making a competitive American AI industry harder to build would threaten national security. This is the first time the federal government has taken a formal position in any of the AI copyright suits. The filing is advisory, not binding, which is the most that can be said for it in law and rather less than what it is worth as a signal (The Hollywood Reporter · The Intercept · The Next Web · Deadline).
Google will keep its advertising business, and the judge ordered changes to how it runs instead
Judge Leonie Brinkema ruled on Wednesday that Google does not have to sell the ad exchange and ad-server business she found last year had been illegally tied together, rejecting the government's proposed breakup and accepting most of the behavioral remedies instead. The full opinion stays sealed for a couple of weeks while it is reviewed for redactions, so her reasoning is not yet public. This is the second time in a year that a court has called Google a monopolist and then taken nothing away from it, after the search case spared Chrome and Android, and both rulings now sit on the desks of the judges hearing the Apple and Amazon antitrust cases, the Amazon trial set for 2027 and Apple's still without a date. The prize was shrinking anyway: the network business at the center of the case supplied 9% of Google's ad revenue last quarter, on Alphabet's reported figures, against 17% in 2018 (AdExchanger · Local 10 · Tech Xplore).
Broadcom's AI chip sales more than tripled, and its chief executive says they double again next year
Third-quarter revenue reached $29.6 billion, up 86% on the year, with AI semiconductor revenue at $16.7 billion, up 221% on the year and 54% on the quarter. The company guided to about $34.8 billion for the current quarter, $21.7 billion of it from AI chips, and Hock Tan, the chief executive, told investors he has line of sight to $115 billion of AI revenue in fiscal 2027 and $230 billion in fiscal 2028. Broadcom designs custom accelerators for Google, Meta and OpenAI rather than selling its own. Its order book is the closest thing available to a forward reading on how much silicon the labs actually intend to buy (Broadcom · PR Newswire · StockTitan · earnings call).
Uber is cutting 3,300 jobs, one in ten across the company, and putting the money into robotaxis
Dara Khosrowshahi, the chief executive, told staff on Wednesday that Uber is removing management layers, shrinking the number of managers by about 20% and moving some of them into individual roles, in order to spend more on ride-hailing, delivery and autonomous vehicles. He also told most remote employees to move within reach of an office, taking the fully remote share of the workforce down to roughly 1%. These are the largest cuts since the pandemic, and they land at a company whose revenue rose 12% year over year to $14.2 billion last quarter (TechCrunch · Bloomberg · Al Jazeera).
░░▒▒▓▓ STATS OF THE DAY ▓▓▒▒░░░░▒▒▓▓ AI ▓▓▒▒░░The technique that made OpenAI's next model better at reasoning also made it harder to watch, and researchers argued about it through the night (update). As we noted yesterday, Astra reveals less of its thinking than its predecessors. The Information has since named the method, recurrent depth, which loops the same layers over a text string repeatedly so that more of the work lands in hidden activations and less in a readable chain of thought. OpenAI says it limited the looping precisely to keep the reasoning monitorable, and chief scientist Jakub Pachocki puts the computation depth of current frontier models within a factor of two of GPT-4. Safety researchers took it as a red line being tested, on the argument that the next developer to adopt the method will have no reason to hold back. The training run OpenAI froze after the Hugging Face break-in is running again (The Information · OpenAI · The Verge · TechCrunch).
A startup founded by Russian mathematicians is letting two models exchange raw internal signals instead of text. Mostik, named after the Russian word for bridge, works on the numerical values inside model weights rather than on prompts and outputs, so a large model can lend capability to a small one without either writing a word. As a demonstration the team bridged the 753-billion-parameter version of the Chinese open-weight model GLM-5.2 to Alibaba's four-billion-parameter Qwen-3.5 that runs on a phone, and says the hybrid costs a twentieth of the large model while landing exactly halfway between the two on performance. Its chief scientist is Stanislav Smirnov, the University of Geneva professor who won the Fields Medal in 2010, and his view is that the hard part is the absence of any adequate mathematical language for finding common ground between two models. If it holds up, open weights get considerably more useful to anyone who cannot afford the frontier (WIRED).
░░▒▒▓▓ TECH ▓▓▒▒░░Apple says a former engineer took its work to OpenAI and then tried to erase the evidence, and OpenAI says the mess is Apple's own. In a filing reported this week, Apple said the engineer's lawyers handed over his company MacBook after weeks of delay, with a Mac mini still to be examined. The filing says he used a confidential power-converter design in his work at OpenAI, downloaded in March from an Apple third-party cloud account he could still reach two months after leaving. Apple further alleges that texts in June between him and an OpenAI colleague, exchanged after the legal inquiry had already begun, discussed restoring the devices and putting them back into use, which would wipe the forensic trail. OpenAI's answer is that Apple's own offboarding is to blame, since staff work from personal iCloud accounts and are escorted out too quickly to hand anything back (WSJ · court filing).
Waymo's robotaxis started taking fares in three more American cities, two days before Tesla shows off a car with no steering wheel. Waymo announced on Tuesday that paid rides start in Denver, San Diego and Tampa, taking the Alphabet company to fourteen US cities, with an overseas expansion in preparation. Waymo also argued in public last week that cameras alone cannot make a vehicle safe and that end-to-end AI systems are not reliable enough on their own, which is a fair description of the approach Tesla will demonstrate at today's Cybercab event. A long-running philosophical argument is about to get an empirical answer (Bloomberg · TechCrunch · WIRED).
░░▒▒▓▓ POLICY ▓▓▒▒░░Washington asked the G20 to leave AI alone, and Bernie Sanders asked Fox News readers to hit pause. At a G20 innovation meeting hosted in North Carolina by the White House science office and the Commerce Department, American officials pressed other governments to apply existing law to AI rather than build new oversight regimes, under a framework being called the Carolina Principles; China has already signed on. Elon Musk, speaking by video on Tuesday, warned of a crisis of power and a likely 15-gigawatt shortfall next year, and predicted a billion humanoid robots within the decade. The same day, Senator Sanders published an op-ed in Fox News calling for a worldwide pause in work on more powerful models, and for Trump and Xi to negotiate an arms-control-style agreement when they meet in Washington this month. What the op-ed does not explain is how a pause of that scope would be verified (Reuters · The Information · Fox News).
Brazil's electoral court defined what a deepfake is, a month before the country votes. The rules themselves have been in force for months: AI-generated campaign material must be labeled, a duty on the books since 2024, and since March AI systems may not rank or recommend candidates and new synthetic content using a candidate's image or voice is banned from 72 hours before the October 4 vote until 24 hours after. They started to matter when campaigning opened on August 16. What the court settled on Tuesday, by five votes to two, was the definition: a deepfake is AI-made or AI-altered content realistic enough to create or alter a person's image, voice or speech, and the ban only bites once the content counts as electoral advertising. The case was a video recreating the face and voice of Jair Bolsonaro, who is under house arrest and barred from political communication, endorsing his son Flávio's presidential run at the Liberal Party's July 25 convention; by four votes to three the court refused to fine him, on the reasoning that streaming a party convention does not by itself make an endorsement early campaigning. Research published in May by ITS Rio, a Brazilian technology-and-society institute, found that six of the seven AI platforms it tested showed some degree of ranking or prioritizing of candidates (TSE · TechRepublic · ITS Rio).
Florida and Texas are pulling the plug on Flock's license-plate cameras (update). Florida's transportation department told staff on Monday it will stop using automated plate readers on state highways, citing reports of misuse, data privacy concerns and surveillance schemes, and gave itself thirty days to take them down; Texas ordered state agencies to stop funding them the week before. Six Texas police departments and sheriff's offices have put officers on leave, charged them or opened investigations over Flock misuse in the past month alone. We covered the company's next-generation tool in August, when the objection was still largely theoretical; the network now runs to roughly 130,000 cameras and the complaint has gone bipartisan (TechCrunch · NBC News · EFF).
░░▒▒▓▓ INFRA ▓▓▒▒░░Texas froze new grid connections for data centers because most of the demand may not be real (update). Requests from data centers and other very large users have gone from about 48 gigawatts in 2023 to more than 474 gigawatts, according to grid operator ERCOT and Governor Greg Abbott, and Texas is the first major hub to stop connecting them and start asking questions, Reuters reported on Tuesday. A review of utility and grid filings puts nationwide requests above 700 gigawatts, more than ten times what data centers are currently estimated to consume, because developers file at several utilities for the same eventual campus as a hedge against permitting and land risk. We flagged the phantom-demand problem in August as an analyst's suspicion; it is now a regulator's finding (US News · BNN Bloomberg).
Google bought 396 megawatts of geothermal power from Fervo, the largest deal yet for the next-generation kind. The electricity comes from Cape Station in southwestern Utah, a project Fervo is spending more than $2 billion to build, with deliveries starting in 2028 and an option for Google to take a further 600 megawatts. Enhanced geothermal drills into hot dry rock instead of hunting for natural steam. That is what puts the resource within reach outside Iceland and northern California. Fervo listed in May, and its shares closed up about 28% on Tuesday (TechCrunch · Canary Media).
░░▒▒▓▓ BUSINESS ▓▓▒▒░░Microsoft will publish Azure's revenue in dollars every quarter, and the number is smaller than Amazon's. The company is collapsing three reporting segments into two, Agents and Infra alongside Devices and Consumer, and has disclosed two years of Azure quarterly figures, ending at $29.4 billion for the June quarter. That puts Azure behind Amazon Web Services at $42.2 billion and ahead of Google Cloud at $24.8 billion, the first time the three can be compared directly. The catch is that Azure now sits in the same segment as Microsoft 365, so investors get revenue for the unit and operating income only for the whole segment, which also swallows productivity licensing, industry solutions and support (CNBC · SEC filing · DigiTimes).
Cognition is in talks to raise about $1 billion at a $47 billion valuation, up from $26 billion three months ago (update). The maker of the Devin coding agent has passed $900 million in annualized revenue, against $492 million in late May, and fielded close to $10 billion of investor interest for a round that may end up larger than the $1 billion sought. As we noted on Tuesday, most of that revenue goes straight back out the door as compute, which is the wager the valuation encodes (Bloomberg · Investing.com).
The cost of running a model is the lowest on record, and the record is not yet a year old. A widely watched token-price index has dropped by more than half from its summer high, CNBC reported on Tuesday, on a combination of cheap Chinese models, price cuts from the American labs and lower production costs. It is a useful corrective to a week of ten-figure numbers: the compute keeps getting cheaper per token even as the total bill grows. That is roughly how a boom sustains itself (CNBC).
░░▒▒▓▓ GEOPOLITICS ▓▓▒▒░░Iran attacked American bases and allies in five places after a night of US bombardment, in the heaviest exchange since July. US Central Command said its strikes hit Iranian air defenses, radar and maritime assets. Iran then fired missiles at an American base in Jordan, near the Gulf of Aqaba, and Kuwait's army said it was engaging missiles and drones; Bahrain, the United Arab Emirates and US facilities in Iraq's Kurdistan region were also targeted. The exchange came days after Washington hit rocket launchers on an island in the Strait of Hormuz, saying Tehran was preparing to mine the waterway. Oil has climbed with it, two months before American midterm elections (Arab News · Gulf News).
Germany blamed Russia for the drone that failed to explode at Leipzig airport, and the EU's top diplomat said it had the hallmarks of state-sponsored terrorism. Interior Minister Alexander Dobrindt said on Tuesday that an assessment of the drone technology and other intelligence points to Russian responsibility for the August 4 attack, in which an explosive-laden aircraft came down near a Ukrainian cargo plane used to ferry munitions to Ukraine and was defused by a police bomb squad. Berlin will close the Russian consulate in Bonn this month, cancel the lease on a Russian cultural center in the capital, tighten entry rules for Russian nationals and propose new EU sanctions listings. EU foreign policy chief Kaja Kallas said on Wednesday that the plot had all the hallmarks of state-sponsored terrorism; Moscow rejected the accusations as absurd and warned of consequences (NPR · France 24 · Defense News · TIME).
China now has more solar capacity than coal capacity, for the first time since it built its first thermal plant in 1882. Installed solar reached 1,286 gigawatts at the end of July against 1,285 gigawatts of coal, or 31.5% of the total, the National Energy Administration said on Tuesday. Capacity is not generation: coal still produces far more electricity and still carries the system when the sun is not cooperating. The milestone is about what has been built rather than what is running (SCMP · CGTN · GZERO).
░░▒▒▓▓ ALPS ▓▓▒▒░░Swiss industrial suppliers may be the collateral damage of Europe's new buy-European law. The European Commission tabled the Industrial Accelerator Act in early March, setting origin quotas for assembly and components in public procurement to counter subsidized Chinese competition, and the text still does not settle whether Swiss firms count as European suppliers. Swiss automotive suppliers sell almost entirely into EU supply chains, so that definition decides their market access; the industry association Swissmem says it is very concerned, and its director Stefan Brupbacher has argued that equal treatment runs through the third package of bilateral agreements rather than through this law. Brussels is split between its protectionists and its free traders, which is a poor room to be waiting outside of (NZZ · Table.Briefings).
░░▒▒▓▓ PAPERS ▓▓▒▒░░Shanghai Jiao Tong University, Singapore Management University and collaborators — how an AI agent's task will end is visible long before it ends. Testing agents has become one of the more expensive things a lab does, since a single pass of a frontier model over an agentic benchmark can cost hundreds to thousands of dollars. EarlyEval, posted on Wednesday, trains a pair of lightweight classifiers on an agent's intermediate behavior to predict success or failure and stop the run early. Across three standard agent benchmarks, which set software-repair, command-line and tool-use tasks, it cut 13% to 26% of agent steps and up to 44% of input tokens at 89% to 97% prediction accuracy, while moving the measured pass rates by only a point or two. For anyone who has to defend an evaluation budget, that is the number worth carrying: much of what it costs to test an agent is spent watching it finish something it had already decided (arXiv).
KAIST AI and Google DeepMind — a model can be asked to declare where it is about to look, and then skip the rest. Language models spend most of their attention on a small share of the context but still read the whole cached history to find it, which is much of why long contexts are slow and expensive. The authors introduce Declarative Attention, a protocol in which the model states inside its own chain of thought which of three modes it needs, full context, one specific region, or recent output only, and the serving engine reads those declarations like tool calls and skips most of the cache. It is a preprint and has not been peer reviewed. The interesting part for a communications reader is the premise rather than the speedup: the model is being treated as the best available judge of what it needs to reread (arXiv).
░░▒▒▓▓ COMMS DESK ▓▓▒▒░░One detector has become the arbiter of whether a piece of writing is human, and careers now turn on its verdict. WIRED's feature this week follows Pangram, which publishers, editors and universities increasingly treat as the standard for AI detection, and the people whose work and reputations have been settled by a score they can neither inspect nor appeal. The reporting does not claim the tool is wrong; it asks who granted it the authority. For the desk: assume that anything your team publishes will be run through a detector by somebody, that a false positive is functionally indistinguishable from a finding, and that the defense that holds up is a drafting trail you can produce on request (WIRED).
░░▒▒▓▓ ONE MORE THING ▓▓▒▒░░A list of San Francisco's ten most dateable people went up on Tuesday evening, ranked on attraction, ambition, accomplishment, reputation and aura, compiled by an entrepreneur who had spent the previous night soliciting nominations and self-nominations. The five men and five women included the podcaster Dwarkesh Patel and Jasper Carmichael-Jack, founder of the startup Artisan, whose billboards once told the city to stop hiring humans. The response from everywhere else was to inform San Francisco at length that it is ugly, which the Core Memory reporter who covered it absorbed with the composure of somebody who lives there (Core Memory).
░░▒▒▓▓ TRACK OF THE DAY ▓▓▒▒░░Salut — the desk at Uncertain Futures