██ ██ ██
██ ▓▓ ██ ▓▓ ██ ▓▓
██ ▓▓ ▒▒ ██ ▓▓ ▒▒ ██ ▓▓ ▒▒
██ ▓▓ ▒▒ ░░ ██ ▓▓ ▒▒ ░░ ██ ▓▓ ▒▒ ░░
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. 34 · Wednesday, 29 July 2026Wednesday, and the call to slow AI down is coming from inside the labs: more than a thousand employees at OpenAI, Anthropic, Google and Meta want Washington to build the means to pace the frontier, and Sam Altman says he is ready. The FCC spent Tuesday shutting new foreign-made humanoid and four-legged robots out of the US market, and BlackRock funds took 80% of Meta's next gigawatt data center. Brussels has just handed the industry another sixteen months to comply with the AI Act. Apologies for the late send; the fact-checking ran long today.
░░▒▒▓▓ TOP 5 ▓▓▒▒░░More than 1,200 employees at rival labs ask Washington for a way to slow frontier AI down
An open letter published Tuesday under the name Pacing the Frontier asks the US government to support an international effort to build the tools needed to "deliberately pace the frontier of automated AI development." The more than 1,200 signatories from inside the labs include Anthropic chief executive Dario Amodei, OpenAI chief scientist Jakub Pachocki, Meta AI chief scientist Shengjia Zhao and Google's vice president of AI safety and alignment Anca Dragan. OpenAI and Anthropic then endorsed it as companies. The ask is narrower than it sounds: not a pause now, but the machinery to impose one later. OpenAI chief executive Sam Altman, who dismissed the 2023 pause letter as missing most technical nuance, now says the rate of development may need pacing so society can harden around new capability levels. He traced his change of view to what he called "the first security incident that I have felt very viscerally" (Pacing the Frontier · The Next Web · TechCrunch).
Washington bars imports of new foreign-made humanoid robots and the inverters that feed data centers
The Federal Communications Commission added two categories to its Covered List on Tuesday: new foreign-produced humanoid and quadruped robots, and the connected power inverters used in batteries, renewable installations and data centers. Being on the list blocks them from the equipment authorizations they need to be imported, marketed or sold in the United States. The determinations behind the additions came from a White House-convened interagency body, citing supply-chain vulnerability and the risk that such devices could be used to surveil Americans or be remotely commandeered. They cover any such device made outside the United States, not only Chinese ones. Previously authorized models and devices people already own are untouched, and the Pentagon can clear a robot on application, Homeland Security an inverter. The practical effect falls on next-generation hardware, first of all Unitree, the Chinese maker of humanoid and four-legged robots; the cheapest capable robot bodies on the market are about to stop arriving (FCC · CNBC · The Japan Times).
BlackRock takes 80% of Meta's next gigawatt data center
Meta announced on Tuesday that its El Paso campus will be built and owned through a venture in which funds managed by BlackRock hold 80% and Meta keeps 20%. Meta contributes land and construction assets worth about $2.3 billion; BlackRock puts in roughly $4.9 billion of cash, part of it funded by a $12.5 billion debt raise; and Meta collects a one-time distribution of about $1 billion to square the split. Total development cost for the one-gigawatt site is put at around $14 billion. It is the same structure Meta used with the asset manager Blue Owl for the far larger Louisiana project, and it means the biggest new computing sites Meta will operate are ones the company mostly does not own (Meta · Forbes).
Brussels rewrites the AI Act and hands industry another sixteen months
The Digital Omnibus on AI entered into force on Monday, three days after publication in the Official Journal. It pushes the main compliance date for stand-alone high-risk systems, the Annex III list covering hiring, credit, education and policing, from this Sunday to 2 December 2027, and for high-risk AI embedded in regulated products from next year to 2 August 2028. National authorities now have until 2 August 2027 to stand up a regulatory sandbox, and the AI Office may add an EU-level one. The transparency duties still switch on this Sunday, along with the enforcement powers we covered yesterday, and systems already on the market get until 2 December 2026 to mark what they generate, so the extra time is for the high-risk rules, not the labeling ones (European Commission · Lewis Silkin).
ChatGPT closes on a billion weekly users, seven months behind schedule
OpenAI has told staff that ChatGPT is on the verge of a billion weekly active users, a milestone the company had wanted by the end of last year, according to The Information. Less than four years after launch, it is still, on almost any measure, the fastest-growing application the internet has produced. It is also seven months late, a reminder that the growth curves in the pitch decks are forecasts rather than physics (The Information).
░░▒▒▓▓ STATS OF THE DAY ▓▓▒▒░░░░▒▒▓▓ EARNINGS ▓▓▒▒░░The quarter that decides whether the buildout is paying for itself starts reporting after the US close tonight.
Microsoft goes in with its stock down about 23% over the past year and with more than 20 million paying Microsoft 365 Copilot seats, a number it has not updated since April. Meta goes in defending capital spending it has guided as high as $145 billion this year, the top of a range it raised itself in April, on top of the partner-funded El Paso venture above. It has no cloud arm to rent the machines out to, so advertising carries all of it. Watch Microsoft's first fiscal-2027 capex forecast; everyone reporting after it will be read against that number (The Information).
░░▒▒▓▓ AI ▓▓▒▒░░Anthropic answers the open-weights letter it did not sign (update). Dario Amodei published the company's position on Monday, writing that Anthropic "has never advocated for a ban on open-weights models" and calling models without dangerous capabilities a public good. He conceded that barring US businesses from using Chinese open-weight models would shield US labs from competition and said that has never been his goal; instead he wants an end to selling powerful chips and chipmaking equipment to China, a crackdown on industrial-scale distillation, and mandatory safety testing for every sufficiently capable model, open or closed. As we covered on Monday, Anthropic was the conspicuous absentee from the Nvidia-led letter, and Amazon was missing too. The list has since grown from 25 organizations at launch to well over 100, and Amazon has now signed. Anthropic has not (Anthropic · CNBC).
Cursor's enterprise customers are refusing their renewal quotes. One IT consultancy paid the AI coding company about $200,000 for 800 licenses in the year to May, was quoted roughly $1.5 million on renewal for the same usage, and negotiated it down to $250,000. Sanofi's chief digital officer Emmanuel Frenehard says his renewal came in at five times last year's price and he may not renew at all, because he gets a better deal from Anthropic's Claude Code. That is what a market with no switching costs looks like from the vendor's side (The Information).
OpenAI plans to more than triple its Dublin headcount, days before the AI Act's transparency duties switch on. The company said on Monday it will take 88,000 square feet in Dublin's Silicon Docks and grow its European headquarters from just over 100 staff to 350, hiring 250 engineers and support operations people over two years. The office opened in 2023 largely as a regulatory listening post; it is now the place where compliance with the AI Act will actually be done (The Star).
░░▒▒▓▓ TECH ▓▓▒▒░░Taiwan detains an Nvidia employee in a chip-smuggling case. Prosecutors in Keelung searched the home and workplace of a suspect surnamed Chang, reported to be an Nvidia employee, on 24 July and won a detention order, arguing they could flee or tamper with evidence. Seven people have now been caught up in an investigation into AI servers routed to China, Macau and Hong Kong in breach of US export controls. It began in May and has already reached staff at the server maker Super Micro's Taiwan unit over roughly 50 falsely documented servers (Forbes · France 24).
Models pointed at big codebases are turning up flaws faster than anyone ever has. The US National Vulnerability Database has logged 45,207 flaws so far this year, on course to roughly double last year's record, and researchers put the surge down to models turned loose on large codebases, Bloomberg reported on Monday. Every one of those entries is a defect that already existed; what changed is that it now gets written down in public (Bloomberg).
░░▒▒▓▓ POLICY ▓▓▒▒░░Beijing calls Washington's distillation accusations hegemony and promises countermeasures. Washington says Chinese labs trained on American model outputs. China's Ministry of Commerce said on Monday that the claim has no factual or legal basis, and amounts to a double standard and hegemonic behavior in the AI sphere. The ministry argued that Washington had ignored how little time now separates Chinese and US frontier releases, and that some Chinese models already have world-leading capabilities; it named none of them. It promised to take all measures necessary against anything that substantially harms Chinese interests (Reuters · CSET translation).
SpaceXAI sues Minnesota to stop a nudification ban from taking effect on Saturday. The law, passed by the legislature in April, fines businesses up to $500,000 each time a user generates a sexualized image of an identifiable person from an ordinary photograph. SpaceXAI argues in a federal complaint filed Monday that the law is unconstitutionally broad and imposes strict liability regardless of what safeguards a provider builds, and that a service whose users generated a hundred thousand covered images could owe $50 billion. The company says it does not dispute Minnesota's interest in banning non-consensual nudes, only the statute written to do so (CNBC).
░░▒▒▓▓ INFRA ▓▓▒▒░░The largest US grid proposes cutting data centers' load in emergencies. The board of PJM Interconnection, which runs the power system for about 67 million people, released proposals on Monday for a one-time backstop capacity auction against a shortfall of about seven gigawatts. A second covers data centers of 50 megawatts or more that cannot line up enough generation of their own: they would cut load or switch to on-site backup when the system nears emergency conditions. Neither exists as a rule yet: PJM expects to file both with federal regulators this month, the auction would run from 30 September only if they sign off, and the curtailment regime needs state or local approval before it starts next June. PJM frames it as protecting households from blackouts and price spikes. For a decade the constraint on a data center was land and chips; it is now whether the neighbors get to keep their air conditioning (Utility Dive · TechCrunch · Bloomberg).
Orange and Morrison plan a €3 billion French data-center venture. The French telecom operator and the infrastructure investor agreed to exclusivity on Monday for a €3 billion 50/50 company built around five existing Orange sites, among them Chevilly-Larue, Aubervilliers, Chartres and Val-de-Reuil. The target is 400 megawatts, close to ten times Orange's current capacity. Both frame it as a sovereign offering for European cloud and AI demand. Signing is expected by the end of the year and closing in the first quarter of 2027, which is a long time in this market (Orange · Bloomberg).
░░▒▒▓▓ THE WORLD ▓▓▒▒░░A major earthquake brings down part of a shopping mall in southern Japan. The quake struck Kyushu on Tuesday afternoon, registering the top level of Japan's seismic intensity scale; Japanese authorities put the magnitude at 7.1, the US Geological Survey at 6.8. The second floor of the Aeon Mall in Kashima gave way; police said 20 to 30 people were unaccounted for, and at least 13 have been confirmed dead. Bullet-train and local services across Kyushu were suspended, and the runway at Aso Kumamoto airport was closed (NPR · The Japan Times).
Europe's wildfires have now displaced 330,000 people (update). The evacuations we covered on Monday have grown again: France ordered another 4,000 out this week, taking the total across France and Spain to about 330,000. French Interior Minister Laurent Nuñez called it very likely the largest civilian evacuation ever carried out in the country outside wartime. More than 1,000 square miles have burned across the two countries, and separate fires forced sea evacuations in Puglia over the weekend. France also recorded its first observed pyrocumulonimbus cloud, the signature of a fire so hot it builds its own thunderstorm and then starts new fires with the lightning. Flames reached NASA's Deep Space Communications Complex outside Madrid, one of three stations that keep the agency in contact with its spacecraft, and the complex's traffic went to Goldstone and Canberra. An inspection on 25 July found damage confined to pavement and vegetation, with the antennas and buildings structurally intact, and Madrid's dishes are back on the network's schedule (PBS NewsHour · SpaceNews).
The Iran war has cost American drivers $74 billion at the pump (update). We reported the Pentagon's share last Thursday: Defense Secretary Pete Hegseth told Congress the conflict has cost $37.5 billion so far, about $8 billion above the May estimate, with a further $67 billion requested to replace munitions and repair damaged bases. The civilian bill is now being counted too. Researchers at Brown University's Watson School of International and Public Affairs calculate that American consumers have paid a further $74 billion at the pump, roughly $562 per household, with regular gasoline at about $4.11 a gallon against the $3.38 they estimate it would cost without the war (GZERO · Reuters).
░░▒▒▓▓ ALPS ▓▓▒▒░░Swiss blind users are finding Meta's glasses genuinely useful, and impossible to check. Samuel Meier reported in NZZ am Sonntag on a blind couple using Meta's camera glasses to read menus, identify yogurt and describe a room, and on the Swiss Central Association for the Blind seeing considerable enthusiasm for them. The catch is structural: a blind user cannot verify what the model left out or got wrong, and with allergy labeling, that gap stops being an inconvenience. One user's summary was that you have to trust, always have to trust (NZZ).
An NZZ commentator says Europe now shows up in AI almost only as a regulator. In a commentary published as Brussels pushed its main high-risk deadline back by sixteen months, NZZ science editor Leonid Leiva Ariosa argues that the United States and China are setting the terms while Europe cannot say what it wants, and that the continent now appears in the global conversation almost exclusively as a regulator. The prescription is ambition rather than another rulebook, and ambition is easier to prescribe than to fund (NZZ).
░░▒▒▓▓ ROBOTICS ▓▓▒▒░░Mila, Quebec's AI research institute, built a working robot guide dog for about $2,000. Milo is a fully autonomous robot guide dog that needs no prior map of its surroundings, does all its computing onboard, and navigates indoors and out while avoiding obstacles and pedestrians. Perception runs on a 3D grid it builds of the space around it; the avoidance policy was trained in a custom simulator. Milo costs about $2,000 to build; a trained guide dog costs roughly $50,000 before upkeep and comes with a waiting list. That is the whole argument for the project. The hardware underneath is a modified Unitree Go2, from the Chinese maker whose next-generation quadrupeds Washington has just shut out (arXiv · Mila).
░░▒▒▓▓ PAPERS ▓▓▒▒░░Anthropic — a model found a real weakness in a post-quantum signature scheme in 60 hours. Anthropic published work on Tuesday reporting that Claude Mythos Preview found a previously unknown structural property in HAWK, a lattice-based digital signature scheme that is a third-round candidate in the US standards body's post-quantum competition, and used it to roughly halve the scheme's effective key size: 60 hours of work against a scheme that had survived two rounds of expert human review over two years. A second result, produced by the model running autonomously inside a purpose-built research scaffold, improved a known attack on a seven-round cut-down of AES-128 (the encryption standard in near-universal use) by a factor of 200 to 800. The core idea landed after three days, the refined attack a few days after that, by which point the model had produced a billion output tokens in total. Two researchers then spent nearly a month confirming it was correct. On the company's own account, each of the two results cost roughly $100,000 in API spend. The caveats are load-bearing and Anthropic states them: HAWK is a candidate and not a deployed standard, the improved attack is still computationally impractical, the AES work builds on prior results assuming 2^105 chosen plaintexts, and neither breaks anything in use. What changed is not that encryption is broken but that cryptanalysis has become a line item you can buy (Anthropic).
Apple — a speech synthesizer small enough to live on the phone. Apple's machine learning group published the architecture behind Siri Expressive Voices this month, in a paper posted to the arXiv preprint server on 26 July: a system that turns the semantic audio tokens emitted by Apple's on-device foundation model into finished speech. It runs on the phone's matrix coprocessor at roughly ten milliseconds a step, about sixteen times faster than real time, using about 21MB of memory at peak and 329MB of stored assets. The design splits the handling of time from the handling of audio detail and reuses a single depth decoder; a sliding-window attention cache is what keeps memory flat however long the utterance runs. For anyone weighing what still has to leave the device, expressive voice no longer does (Apple Machine Learning Research).
░░▒▒▓▓ REPORTS ▓▓▒▒░░Stanford HAI — The World Model and Spatial Intelligence Era: Governing AI Beyond Language. The institute's July issue brief, written by a group including Fei-Fei Li, Amy Zegart and Daniel E. Ho, sets out what HAI calls a first-of-its-kind policy agenda for world models, the systems that hold a working representation of an environment and predict how it changes when something acts on it. Its central finding for anyone drafting AI policy or corporate positioning: no existing benchmark gives policymakers an adequate basis for judging whether such a system is safe to deploy in a safety-critical setting, and policy built for AI-generated content and autonomous decision-making does not fully address what world models do. The brief names three priorities: broad access to the technology, safeguards proportionate to where a system is used, and independent public capacity to evaluate it. It also flags that world models are likely to become an important arena of geopolitical competition, particularly between the US and China, on their commercial and military potential alone (Stanford HAI).
░░▒▒▓▓ COMMS DESK ▓▓▒▒░░Meta has started removing content filmed on its own smart glasses from its own platform, the same glasses the blind Swiss couple above rely on to read a menu. Instagram head Adam Mosseri said earlier this month that videos "taking advantage of people and harassing them, like a lot of these pickup line kind of videos" would come down, and two pickup-artist accounts, each with more than a million followers, were deactivated; a Meta spokesperson confirmed they came down for breaching the policy Mosseri described. Meta has not explained how that policy is enforced or what exactly counts as a violation. Takeaway: when the fix for a product's reputation problem is banning what customers do with it, the campaign is no longer the lever. Ship the enforcement definition at the same time as the ban, or the next question from every reporter is the one Meta is currently not answering (The Next Web · Gizmodo).
░░▒▒▓▓ ONE MORE THING ▓▓▒▒░░Midjourney, the image generator, announced last week that it has bought Co-Star, the astrology app that pushes daily horoscopes to about 4.3 million monthly users. Terms were undisclosed, and Co-Star's chief executive Banu Guler becomes Midjourney's chief design officer while continuing to run the app; Midjourney founder David Holz says he first met her about five years ago. A company built on predicting the next pixel now also owns one built on predicting the rest of your week (Engadget · Gizmodo).
░░▒▒▓▓ TRACK OF THE DAY ▓▓▒▒░░🎵 Shame — Bibio, Dorian Concept
Gemächlich — the Uncertain Futures desk