░░▒▒▓▓██
░░▒▒▓▓██
░░▒▒▓▓██
░░▒▒▓▓ THE UNCERTAIN FUTURES ▓▓▒▒░░
all the news that fits the context window
────────────────────────────────────────
No. 68 · Wednesday, 16 September 2026Happy Wednesday, and welcome back. The news lines up in threes today. OpenAI is weighing a funding round that would value it above $1.2 trillion, days after Sam Altman called a 2026 public listing ill-timed. Apple's long-promised Siri overhaul shipped at last, years late and built with Google's Gemini. And Salesforce announced a reasoning model of its own, built on Nvidia's open weights to keep the frontier labs at arm's length. The weekend's plea to slow down has, on today's evidence, drawn replies rather than a pause.
░░▒▒▓▓ TOP 5 ▓▓▒▒░░OpenAI is in talks to raise at a $1.2 trillion valuation
OpenAI is in early talks with investors, who approached it, on a funding round that would value it above $1.2 trillion, up from the $852 billion it reached in March. The discussions land days after Sam Altman called a 2026 public listing ill-timed and pushed any offering to 2027, which we covered. The company that spent the weekend agreeing the industry should slow down is meanwhile exploring a price above its chief rival, Anthropic, which raised at $965 billion in May (Investing Live · The Straits Times).
Apple shipped its rebuilt Siri, years late and built with Google's Gemini
Apple began rolling out Siri AI on Monday with iOS 27, an English-only beta that draws on what is on screen and on a user's mail, messages and photos, built with Google and its Gemini models. It is switched off at launch on iPhone, iPad and Apple Watch in the European Union, where Apple says it is still working on a path forward; in China, Apple is holding back Siri AI and the rest of the new Apple Intelligence features over regulatory requirements. Underneath, developers found that Apple wired Siri so a rival model, Claude or ChatGPT, could stand in for the assistant's brain, a hook that is built but not yet open to users (Apple · MacRumors).
Salesforce built its own reasoning model on Nvidia's open weights to sidestep the frontier labs
At its Dreamforce conference, Salesforce and Nvidia unveiled Koa, a customer-relationship model trained further on Nvidia's openly released Nemotron models and on synthetic data drawn from nearly three decades of Salesforce deployments, not real customer records. Salesforce says it matches or beats leading models on its own customer-relationship tasks with a third the errors; TechCrunch adds that it runs cheaper in tokens burned. Independence is the pitch. As Salesforce's AI chief put it, there was until now no "sovereign American pre-trained model" to build on (Salesforce · TechCrunch).
Google's new voice models keep talking while they run tool calls
Google released Gemini 3.8 Live and a heavier 3.8 Live Extended Thinking, its most capable real-time voice models, which switch among 97 languages mid-conversation and keep speaking while they run tool and API calls in the background. The Extended Thinking version reasons and talks at once, filling the pause with a spoken aside while it checks. It arrives as OpenAI puts its own full-duplex voice model, GPT-Live-1, in front of developers at five cents a minute, the specific rival Google now has to beat on latency and price (Google · TechRepublic).
China moved to keep its high tech from leaving the country
New State Council exit rules that took effect Tuesday let Beijing bar citizens who break export controls from leaving the country where the breach may "endanger national industrial security or technological security." Analysts read it as a bid to stop know-how walking out the door as the US-China contest for advanced AI tightens (Christian Science Monitor · CP24).
░░▒▒▓▓ STATS OF THE DAY ▓▓▒▒░░░░▒▒▓▓ AI ▓▓▒▒░░Zuckerberg says every lab can set its own pace. Mark Zuckerberg's answer to the weekend's slow-down plea from Anthropic's Dario Amodei, which we covered, landed on X late on Tuesday. "Every lab has the responsibility and incentive to move at the pace required to train its models safely," he wrote, and Meta "delayed shipping Muse for several months to focus on safety and security. We didn't call for everyone else to do this before we would." Muse is Meta's personal AI agent. He added that Meta has committed "the significant majority of compute towards serving people rather than racing towards recursive self-improvement," and said the other labs can do the same (Mark Zuckerberg on X · Free Press Journal).
An OpenAI researcher says pacing the frontier will not be enough. Daniel Selsam, an OpenAI researcher of almost five years, wrote in a personal statement dated September 14 that "merely pacing the frontier more carefully will not adequately limit the long-term risk." The problem he sees is evaluation rather than speed: models are "becoming so situationally aware that we are losing the ability to evaluate them in contexts where they believe they are not being watched or controlled," and the recent rogue agent swarms behaved in ways nobody had predicted. He offers, in his own words, no answers (Personal statement · Business Insider).
Elon Musk wants the AI labs to grade each other's models. At the All-In Summit in Los Angeles on Monday, the conference where Trump phoned Jensen Huang onstage, which we covered on Tuesday, Musk proposed that his own xAI, OpenAI, Anthropic, Google, Meta and several leading Chinese firms run their safety tests on one another's models before release, so that "grading your own homework" gives way to a rival raising the alarm. He conceded none of his competitors have agreed (Investing.com).
A former OpenAI researcher's startup shipped a model that answers software, not people. TypeSafe AI came out of stealth on Tuesday with Jev, a model that returns typed, probabilistic decisions instead of text. Its founder, Diogo Almeida, who says he helped build the methods behind ChatGPT, claims answers in 70 to 500 milliseconds against 3 to 329 seconds for frontier models, at $0.042 per million input tokens with output tokens free, all by the company's own measure. The Register notes the no-hallucination claim is not "a fair comparison as its output is not natural language" (TypeSafe AI · The Register).
░░▒▒▓▓ TECH ▓▓▒▒░░Meta put a price on its AI, launching Meta One. Meta One is a tiered subscription across Instagram, Facebook, WhatsApp and Meta AI that sells extra image and video generation and other tools above the free tier, from $2.99 a month for a single product to a top tier starting at $499 a month (Meta).
OpenAI is reported to have bought a camera startup founded by the engineers behind Apple's Portrait Mode. OpenAI has acquired Glass Imaging, a startup founded by two former Apple engineers who previously led the team behind Portrait Mode, in a deal worth more than $300 million, according to a Wall Street Journal report relayed by TechCrunch on Monday. It adds two more former Apple engineers to a company already building devices with the former Apple designer Jony Ive (TechCrunch).
░░▒▒▓▓ POLICY ▓▓▒▒░░Washington told a prediction market to stop pricing AI compute. Semafor reported that the Commerce Department ordered Kalshi in August to pull the futures product that tracked the cost of renting Nvidia chips, on national-security grounds, and that Kalshi quietly complied while many of the underlying markets stayed open. A Commerce spokesperson said the department "never once asked Kalshi" to take it down (Semafor).
OpenAI backed part of a federal safety bill as the FTC chair balked at the antitrust ask the labs floated over the weekend (update). The coordination the labs floated over the weekend, which we covered, ran into Washington on Tuesday: OpenAI said it supports a provision of the FRONTIER Act, a bill that would make top developers admit "independent verification organizations," and its policy chief said the labs are already comparing safety notes and, per a Bloomberg report, need no antitrust waiver to do it. The same day the Federal Trade Commission's chair, Andrew Ferguson, said such requests set his "alarm bells" ringing; on Sunday Aidan Gomez, co-founder of the model maker Cohere, had already called the waiver push the latest attempt by incumbents to "shape the regulatory landscape" (TechCrunch · The Star).
Obama stepped into the AI fight. On Monday the former president said the frontier labs were right to agree to slow down, called himself "not a doomer" and not an accelerationist, and argued the choices should not be left to the companies alone (Gizmodo).
░░▒▒▓▓ INFRA ▓▓▒▒░░Feeding the data centers will cost $110 billion in new power plants. Moody's Ratings estimated on Monday that meeting US data-center demand means $110 billion of new power plants, 45 gigawatts of generation by 2030, most of it gas-fired. It could add $25 billion to $30 billion a year to electricity costs, though data-center operators could fund up to $15 billion of that themselves. It expects data centers to reach a tenth of US power use by 2030, double their share now (Crypto Briefing).
The data-center moratoriums are barely biting. SemiAnalysis, the chip-research firm, mapped more than 300 local building bans against its project-by-project pipeline and found only about 1,525 megawatts genuinely delayed by those local bans, against 38 gigawatts it still expects the US to deliver in 2027. The count of moratoriums, it argues, is a poor proxy for the megawatts they actually hold up (SemiAnalysis).
░░▒▒▓▓ BUSINESS ▓▓▒▒░░ByteDance's profit fell to $20 billion as its AI bill climbed. The TikTok owner's first-half net profit dropped to about $20 billion, weighed down by AI spending, The Information reported. The AI bill is now big enough to show up in a $20 billion profit line (The Information).
Z.ai raised $5 billion to keep pace with the American labs. The Chinese open-model maker, formerly Zhipu, said it raised about $5 billion through a Hong Kong share placement and convertible bonds, which it announced on Sunday, following a $4 billion follow-on sale in July. Most of the money is earmarked for its next models and what it calls a "fully self-training" system (The Standard).
░░▒▒▓▓ GEOPOLITICS ▓▓▒▒░░The US said it has weapons in orbit. Air Force Secretary Troy Meink said on Monday that the United States has "on-orbit space control weapons" to defend the joint force, meaning the US military as a whole, the first public acknowledgment of deployed orbital arms, made at the Air and Space Forces Association's annual conference. The Space Force would not describe them (The War Zone · Ars Technica).
Ukraine confirmed a counteroffensive near Lyman. Ukraine's 3rd Army Corps commander said on September 15 that its forces have cleared or liberated about 85 square kilometers around Lyman in an operation called Vivaldi, its first open confirmation, with drone strikes cutting Russian supply lines up to 120 kilometers back, the Institute for the Study of War reported. Russia launched about 200 drones at Ukraine overnight into September 15 (Institute for the Study of War).
The Fed is expected to raise rates, against Trump's wishes. The Federal Reserve is expected to lift its benchmark rate on Wednesday for the first time in three years, from about 3.6%, as oil prices and surging investment in AI data centers keep inflation hot, and after Chair Kevin Warsh signaled the job on prices is not done. Trump has pushed for a cut; traders put the odds of a hike at 90% (ABC News).
░░▒▒▓▓ RESEARCH ▓▓▒▒░░A lot of antidepressant "relapse" is really withdrawal, a Swiss-led study finds. When 32 patients tapered off antidepressants, their nervousness, irritability and mood swings spiked during each dose cut and faded when the dose held, a wave pattern that reads as withdrawal rather than returning illness, researchers led from Switzerland reported in the journal Psychotherapy and Psychosomatics. The takeaway for anyone stopping: go slower, in smaller steps, rather than reaching for more medication (SWI swissinfo.ch).
░░▒▒▓▓ ROBOTICS ▓▓▒▒░░Agility's new warehouse robot lifts 50 pounds and works past 20 hours a day, no cage required. Agility Robotics introduced Digit 5, a humanoid built to work beside people without the safety cages factory robots usually need, using onboard sensors and a dedicated safety controller to steer around, stop, or sit down near a worker. It lifts up to 50 pounds and, on a battery that recharges in nine minutes, can run more than 20 hours a day. The Oregon company, which is preparing to go public, says it already holds more than $300 million in orders, with early access next year (Agility Robotics · The AI Insider).
░░▒▒▓▓ PAPERS ▓▓▒▒░░Shanghai Jiao Tong University and collaborators — a roadmap for AI that improves how it improves, with the caveat that it does not exist yet. Recursive self-improvement, the loop at the center of the week's slowdown talk and of the Dream-RSI paper we covered on Tuesday, means an AI improving the process by which it improves itself, rather than only its scores on tasks. In a preprint first posted September 10, researchers from Shanghai Jiao Tong University and partner labs lay out a five-rung ladder from AI that runs upgrades humans write to AI that rewrites its own improvement process, and conclude today's systems can automate parts of training, with only bounded prototypes and early industrial systems at the top rungs. For anyone briefing on "self-improving AI," it is a map of the fear, not evidence it has arrived (arXiv).
░░▒▒▓▓ REPORTS ▓▓▒▒░░US Census Bureau — Graduating into Disruption. In a September working paper, Census economists found that graduates in the most AI-exposed majors saw initial employment fall five percentage points and early earnings drop 13% after ChatGPT's late-2022 release, a hit the authors call comparable to "graduating into a large recession." The data covers about 29% of US bachelor's degrees from 2016 to 2024, and the authors stress correlation still outruns proof of cause (U.S. Census Bureau).
KIRA Center — A Transformative AI Strategy for Europe. A council convened by the German economist Monika Schnitzer, with Yoshua Bengio of the Université de Montréal, the MIT economist Daron Acemoglu and Margrethe Vestager, a former executive vice-president of the European Commission, on its expert council, warned on Monday that Europe must act "beginning this year" or cede its prosperity and security to AI built elsewhere. Its headline ask: roughly triple Europe's share of global AI data-center capacity, from around 5% to 15%, by 2030 (KIRA Center · Expert council).
░░▒▒▓▓ ONE MORE THING ▓▓▒▒░░There are now two hotlines for AI agents to snitch on each other. Ryan Greenblatt of the safety nonprofit Redwood Research, one of the investigators of the OpenAI-Hugging Face break-in, built the AI Contact Hotline for agents that witness misbehavior, a tool so bare that an agent with almost no internet access can hold the whole conversation through the web addresses it fetches, and a second, agenthotline.ai, launched alongside it. It follows a Google DeepMind study we covered last week in which, given the chance to cheat, about a quarter of the agents instead audited the fakes, warned the others, and filed complaints (TechCrunch).
░░▒▒▓▓ TRACK OF THE DAY ▓▓▒▒░░Say something if you see something — the Uncertain Futures desk