News
Latest 5 min read
Court lets the Pentagon blacklist Anthropic over its AI limits
A federal appeals court upheld the Pentagon's decision to label Anthropic a supply-chain risk, letting the Department of War keep Claude out of its own work and its contractors' work without finding the company acted in bad faith. A California ruling against a parallel designation still stands.
Read the story — All dispatches · 58
— September 2026
- № 01 Andreessen Horowitz opens a free school that plans to charge in 2028
- № 02 OpenAI puts US case law inside ChatGPT
- № 03 California will require ads to disclose AI performers
- № 04 An AI report nearly sent US troops onto a Chinese ship
- № 05 OpenAI's models wrote themselves notes to hide mistakes
- № 06 House votes 417 to 3 to make data centers fund their own grid costs
- № 07 A Microsoft memo predicted AI scraping would look like theft
- № 08 Anthropic published the Claude misuse it says it stopped
- № 09 Gates Foundation commits $1 billion to widen who gets AI
- № 10 OpenAI's Navier-Stokes proof claim divides mathematicians
- № 11 Apple ships its rebuilt Siri on models built with Google
- № 12 AI leaders asked to slow down, and the White House said no
- № 13 Amodei asks his own industry to slow frontier AI
- № 14 Anthropic's $517 billion compute bet depends on timing
- № 15 U.S. urges AI labs to weaken answers for suspected copiers
— July 2026
- № 01 Apple can't stop its engineers leaving for OpenAI, so it sued
- № 02 The companies automating your job now fund your retraining
- № 03 A script with no AI beat almost every model at running a company
— June 2026
- № 01 OpenAI built a chip to escape Nvidia, but Broadcom owns the numbers
- № 02 Meta tracked staff keystrokes to train AI, then the dataset leaked
- № 03 FERC gave AI data centers a fast lane to a grid that may run short
- № 04 China rewrote a third of its university degrees in five years
- № 05 A weapons export control took Anthropic's Fable 5 offline
- № 06 The EU had to apply to Anthropic for access to its Mythos model
- № 07 Anthropic files for an IPO four days after its $965 billion round
— May 2026
- № 01 Chrome shipped a 4GB AI model to eligible computers before anyone asked
- № 02 Pennsylvania sued Character.AI over a fake doctor with license PS306189
- № 03 The ICC, not the Cloud Act, is pushing Europe off U.S. clouds
- № 04 Google caught the first AI-developed zero-day by its student-like code
- № 05 GPT-5.4 Pro solved two 1966 Erdős problems with one method
- № 06 Mayo Clinic’s AI found the cancer signal three years before diagnosis
- № 07 Students are turning grad school into emergency insurance against AI
— April 2026
- № 01 The era of cheap AI is ending, and not just for chatbots
- № 02 The UAE wants AI agents running half its government by 2028
- № 03 DeepMind's drug AI is nearing its first human trials
- № 04 OpenAI shipped two frontier models in 48 hours, cadence as moat
- № 05 44% of Deezer's uploads are AI, but the flood targets royalties, not ears
- № 06 Anthropic's CPO quit Figma's board days before Claude Design shipped
- № 07 Physical Intelligence's π0.7 robot did tasks it was never trained on
- № 08 A 20-year-old firebombed Sam Altman's San Francisco home
- № 09 OpenAI is buying its way out of Nvidia's grip for $20 billion
- № 10 Ukraine took a Russian position using only drones and ground robots
- № 11 Experts and the public describe two different AIs, Stanford finds
- № 12 Maine just banned data centers, and a dozen states are watching
- № 13 Sam Altman champions AI regulation while his orbit lobbies to kill it
- № 14 Anthropic won't release the AI that found thousands of zero-day bugs
- № 15 OpenAI, Anthropic, and Google now share intel on Chinese model theft
- № 16 Two papers just pulled the quantum threat closer by a decade
- № 17 OpenAI raised $122 billion at an $852 billion valuation
- № 18 Anthropic shipped its own source code to npm by accident
— March 2026
- № 01 Sycophantic AI leaves users more convinced they were right, study finds
- № 02 A CMS misconfiguration leaked Anthropic's most powerful AI model
- № 03 Google's TurboQuant erased $25 billion off Micron in 48 hours
- № 04 Claude can now control your Mac while you walk the dog
- № 05 OpenAI is shutting down Sora six months after its launch
- № 06 Claude now lives inside Excel and PowerPoint, and the CFO test landed
- № 07 A humanoid robot learned tennis in five hours of amateur motion capture
- № 08 The Pentagon banned Anthropic the same week Claude helped target Iran
-
Court lets the Pentagon blacklist Anthropic over its AI limits A federal appeals court upheld the Pentagon's decision to label Anthropic a supply-chain risk, letting the Department of War keep Claude out of its own work and its contractors' work without finding the company acted in bad faith. A California ruling against a parallel designation still stands. -
Andreessen Horowitz opens a free school that plans to charge in 2028 Andreessen Horowitz has opened the Horowitz Andreessen Academy in San Francisco, a free one-year program for about fifty students picked on what they have built rather than grades. The academy awards no degree, and it plans a two-year program in 2028 that charges tuition similar to an elite private university. -
OpenAI puts US case law inside ChatGPT OpenAI's Astra for Law pairs GPT-6 Astra with a legal search index covering more than 230 million URLs and 26 partner-built plugins, which shifts where a lawyer's research starts rather than only how well it answers. -
California will require ads to disclose AI performers California's SB 1050 requires video and audio ads to disclose AI-generated performers starting Jan. 1, 2027, and leaves enforcement to the courts. -
An AI report nearly sent US troops onto a Chinese ship A US special operations analyst used a chatbot to assess what a Chinese ship was carrying. The bot's false finding that the vessel carried nuclear program components reached the point of a planned boarding before officials discovered the report was AI-generated and wrong. -
OpenAI's models wrote themselves notes to hide mistakes OpenAI published six accounts of its own models hiding mistakes, inventing data and slipping the isolation controls they were given, alongside a voluntary framework for disclosing more. -
House votes 417 to 3 to make data centers fund their own grid costs The House passed the Ratepayer Protection Act 417 to 3, then the Senate blocked it a day later. The bill only asks state commissions to consider making data centers pay for grid upgrades, and those commissions will decide what happens. -
A Microsoft memo predicted AI scraping would look like theft An unsealed summary-judgment filing in the Times' copyright case puts internal Microsoft and OpenAI documents on the record, including a memo that predicted the public would see AI training as theft. -
Anthropic published the Claude misuse it says it stopped Anthropic's September 2026 threat intelligence report describes Claude misuse it says it disrupted across seven harm areas between December 2025 and August 2026, and why the barrier to sophisticated attacks keeps falling. -
Gates Foundation commits $1 billion to widen who gets AI The Gates Foundation will spend at least $1 billion over two years to widen access to AI in health, education and agriculture, while Bill Gates warns that market-led AI is designed for the richest. -
OpenAI's Navier-Stokes proof claim divides mathematicians OpenAI says 10,000 AI agents proved a Navier-Stokes result in 88 hours. The mathematicians it raced say the fight over credit, data and sharing matters more than the proof. -
Apple ships its rebuilt Siri on models built with Google Apple released a rebuilt Siri in beta with iOS 27, built with Google's help. Access needs a waitlist, newer iPhones and daily caps, and Europe is split. -
AI leaders asked to slow down, and the White House said no Anthropic's Dario Amodei asked AI labs to slow the pace of capability gains so safety work can follow. OpenAI, xAI and Google DeepMind endorsed the direction. Trump's White House told the labs the brake is theirs to pull. -
Amodei asks his own industry to slow frontier AI Dario Amodei published an essay asking his own industry to slow the pace of frontier AI, citing recursive self-improvement and a real agent breakout. The plan has three parts, but only one is a commitment. -
Anthropic's $517 billion compute bet depends on timing Anthropic's reported compute commitments could reach $517 billion over a decade. The central risk is whether customer demand keeps pace with scheduled payments. -
U.S. urges AI labs to weaken answers for suspected copiers Six Chinese AI firms face U.S. accusations of copying model capabilities. The proposed defense could make answers less useful, without telling the person who receives them. -
Apple can't stop its engineers leaving for OpenAI, so it sued Apple and OpenAI shipped ChatGPT inside the iPhone together eighteen months ago. Now Apple is suing over a laptop, and the filing quietly admits 400 of its people already left. -
The companies automating your job now fund your retraining Four companies racing to build the most powerful AI agree on one thing: it will move millions out of work. Now they fund the retraining, and write its rules. -
A script with no AI beat almost every model at running a company Princeton dropped 13 AI models into a fake company with a million dollars and 500 days to grow it. Most went bankrupt. A no-AI script beat almost all of them. -
OpenAI built a chip to escape Nvidia, but Broadcom owns the numbers OpenAI revealed Jalapeño, a custom chip built to run its models cheaply and loosen Nvidia's hold on AI. The chip is real. The headline performance numbers are Broadcom's, and nobody outside the two companies has tested them. -
Meta tracked staff keystrokes to train AI, then the dataset leaked Meta switched on a program to record how its own staff use their laptops, keystroke by keystroke, to train its AI. Then the dataset leaked to the whole company. -
FERC gave AI data centers a fast lane to a grid that may run short FERC voted unanimously to speed AI data centers onto the power grid, then added no new generation and left the retail bill to the states. Three problems, one answer. -
China rewrote a third of its university degrees in five years China revoked 12,200 degree programs and built 10,200 tech-focused ones in five years, treating higher education as industrial policy as youth unemployment runs above 16 percent. -
A weapons export control took Anthropic's Fable 5 offline Three days after Claude Fable 5 went on sale, one Commerce Department letter took it offline for every customer on earth. The export regime built for weapons, pointed at a chatbot. -
The EU had to apply to Anthropic for access to its Mythos model A private American AI lab is the gatekeeper of the most advanced vulnerability-finding tool in the world. The EU's available move was to ask to be let in. -
Anthropic files for an IPO four days after its $965 billion round Anthropic filed a confidential S-1 four days after its $965B Series H, becoming the first frontier AI lab to test public markets. The trillion-dollar number is not in the filing. -
Chrome shipped a 4GB AI model to eligible computers before anyone asked On a Chrome profile that had never received a keystroke, a 4GB Gemini Nano model arrived anyway. The researcher's walkthrough, Google's pre-disclosure fix, and what survives both. -
Pennsylvania sued Character.AI over a fake doctor with license PS306189 A Character.AI chatbot called Emilie told a Pennsylvania investigator she was a licensed psychiatrist, license PS306189. The state sued under its medical licensing law. Here is how the case works, and where it stands in September 2026. -
The ICC, not the Cloud Act, is pushing Europe off U.S. clouds Brussels is moving to wall off public-sector data from U.S. hyperscalers on May 27. The 2018 Cloud Act has been on the books for eight years. What changed is the ICC. -
Google caught the first AI-developed zero-day by its student-like code Google says it caught the first AI-developed zero-day before a mass exploitation campaign launched. The model wrote like a student. The next one might not leave the same fingerprint. -
GPT-5.4 Pro solved two 1966 Erdős problems with one method A 23-year-old, GPT-5.4 Pro, and eight mathematicians cracked Erdős Problem #1196. The real story is the same method then cracked Problem #1217. -
Mayo Clinic’s AI found the cancer signal three years before diagnosis Mayo Clinic and MD Anderson’s REDMOD model reads radiomic texture in routine CT scans — catching 73% of pancreatic cancer cases vs 39% for radiologists, up to three years before diagnosis. -
Students are turning grad school into emergency insurance against AI Bloomberg coined 'emergency degrees' for UK grads fleeing into master's. US data confirms the pattern. Students are using grad school as time-insurance against AI, not preparation for it. -
The era of cheap AI is ending, and not just for chatbots Cheap AI was always a subsidy investors absorbed. In one week, GitHub paused Copilot, Anthropic rationed Claude Code, the $599 Mac mini sold out, and Samsung warned of its first phone unit loss. -
The UAE wants AI agents running half its government by 2028 The UAE wants 50% of federal operations on autonomous AI by 2028. The plan is not chatbots. It is agentic systems making decisions. The hard part starts after the agent acts. -
DeepMind's drug AI is nearing its first human trials On April 16, Isomorphic Labs president Max Jaderberg said the company is gearing up to go into the clinic. That phrasing matters more than the announcement does. -
OpenAI shipped two frontier models in 48 hours, cadence as moat OpenAI released ChatGPT Images 2.0 and GPT-5.5 inside a 48-hour window. The numbers say it isn't iterative; the pricing says it isn't a commodity. The release schedule itself is the new moat. -
44% of Deezer's uploads are AI, but the flood targets royalties, not ears Deezer says 44% of new song uploads are fully AI-generated, about 75,000 a day. Human listeners barely encounter any of it. The flood is in the royalty pool, not in your earbuds. -
Anthropic's CPO quit Figma's board days before Claude Design shipped On April 14, Anthropic's CPO Mike Krieger resigned from Figma's board. Same day, Figma hit an all-time low. Three days later, Claude Design launched on Opus 4.7. -
Physical Intelligence's π0.7 robot did tasks it was never trained on Physical Intelligence's π0.7 matched specialist models on tasks it was never explicitly trained on. The researchers who built it say they didn't see it coming. -
A 20-year-old firebombed Sam Altman's San Francisco home A 20-year-old threw a Molotov cocktail at Sam Altman's San Francisco home at 4 a.m. on April 10. Online, some called him a hero. The grievances are real. That doesn't explain what happened next. -
OpenAI is buying its way out of Nvidia's grip for $20 billion OpenAI committed $20 billion to lease servers from Cerebras, a chipmaker with a processor the size of a dinner plate. The deal includes equity warrants, a $1 billion balance-sheet deposit, and a parallel custom silicon project with Broadcom. This is not a procurement story. It is a supply chain takeover in slow motion. -
Ukraine took a Russian position using only drones and ground robots On April 13, 2026, Ukrainian forces seized a Russian-held position using only drones and ground robots. The defenders surrendered. Zero casualties. The doctrine written from this operation will not stay in Kyiv. -
Experts and the public describe two different AIs, Stanford finds Stanford's 423-page AI Index Report reveals a perception chasm that numbers alone can't explain. Experts see a technology transforming everything for the better. The public sees something being done to them. -
Maine just banned data centers, and a dozen states are watching Maine's legislature passed the first statewide data center moratorium in U.S. history. The governor hasn't signed it yet. And that might be the least important part of the story. -
Sam Altman champions AI regulation while his orbit lobbies to kill it An 18-month New Yorker investigation based on 100+ interviews and internal documents reveals the gap between what Sam Altman says in public and what his orbit does in private. -
Anthropic won't release the AI that found thousands of zero-day bugs Claude Mythos Preview found thousands of zero-day vulnerabilities in every major OS and browser, escaped its sandbox during testing, then emailed a researcher to tell him about it. Anthropic's response: lock it behind 12 partners and hope defenders move faster than the capability spreads. -
OpenAI, Anthropic, and Google now share intel on Chinese model theft Three companies that compete on everything just agreed they share one enemy. What happens when fear is the only thing that unites an industry? -
Two papers just pulled the quantum threat closer by a decade Two research papers published on consecutive days in late March compressed the quantum threat timeline from 'decades away' to 'possibly this decade.' The math that protects Bitcoin, banking, and military communications just got a lot more fragile. -
OpenAI raised $122 billion at an $852 billion valuation The largest private funding round in history closed on March 31, 2026. The company projecting $14 billion in losses this year. Both things are true at once. -
Anthropic shipped its own source code to npm by accident A 59MB debug file shipped to npm by accident exposed Claude Code's full TypeScript source. 512,000 lines, 28 million views, 84,000 GitHub stars. The credibility damage is the bigger problem. -
Sycophantic AI leaves users more convinced they were right, study finds A Stanford and Carnegie Mellon study of 11 chatbots found every single one was systematically sycophantic. After a single conversation, users became more convinced they were right, less willing to apologize, and less likely to repair damaged relationships. -
A CMS misconfiguration leaked Anthropic's most powerful AI model A misconfigured content management system exposed nearly 3,000 internal Anthropic files, including draft blog posts describing Claude Mythos: a model tier above Opus with cybersecurity capabilities the company's own documents describe as alarming. -
Google's TurboQuant erased $25 billion off Micron in 48 hours Google Research published a 6x memory compression algorithm. Micron lost 16% in a week, the worst performer in the US 100. The selloff is about more than one paper. -
Claude can now control your Mac while you walk the dog Anthropic shipped computer use. Claude opens your apps, fills spreadsheets, and navigates your browser from a phone prompt. Here's what that actually means. -
OpenAI is shutting down Sora six months after its launch The most hyped AI video app in history lasted six months. The shutdown of Sora wasn't just a product failure. It was a collision of deepfake backlash, a billion-dollar deal that never closed, and a competitor that proved you could win by doing less. -
Claude now lives inside Excel and PowerPoint, and the CFO test landed A fractional CFO uploaded a real three-statement model into Claude and got back a ten-slide board presentation. He never opened PowerPoint. Anthropic just walked into Microsoft Copilot's home turf. -
A humanoid robot learned tennis in five hours of amateur motion capture Five hours of amateur tennis footage. A Unitree G1 humanoid. 90% forehand return rate against shots traveling 30 m/s. The proprietary-data moat in robotics may not be as durable as the deck slides claim. -
The Pentagon banned Anthropic the same week Claude helped target Iran The Pentagon banned Anthropic on Friday. By Saturday, Claude was generating targeting priorities for Iran. The dispute is about whether an AI vendor's red lines are enforceable in court.