Manuel Uth

Sep 11, 2026

Ex-Deepmind VP Vinyals says AI self-improvement is coming but won't trigger an intelligence explosion

Oriol Vinyals, until recently head of research at Google DeepMind, thinks a sudden AI intelligence explosion through recursive self-improvement is unlikely. AI can speed up research by a factor of ten, he says, but it hits two bottlenecks: coming up with ideas (“research taste”) and reliably judging results. Reward hacking and the speed of light add further limits. Vinyals now wants to tackle these bottlenecks with his startup Discovery Loop, co-founded with Jeff Dean, Sanjay Ghemawat, and Quoc Le.

Read full article

Comment

Manuel Uth

Sep 11, 2026

Short News

A class action lawsuit accuses Anthropic of misrepresenting how much Claude subscribers actually get to use the service. That's according to The Verge. Subscribers to the Max plan pay $100 a month for five times the usage of the Pro plan, or $200 for twenty times the usage.

But the advertised limits don't apply across the board, the plaintiffs argue. Instead, the multipliers only count within five-hour windows and are further capped by a weekly limit. That means real-world usage ends up well below what customers expected. Anthropic does confirm this structure on its help page and reserves the right to restrict usage further at its own discretion.

The company filed a motion to dismiss, arguing that all the relevant details were available through hyperlinks during the purchase process. The plaintiffs' lawyers pushed back, saying consumers have no way to verify what an AI service actually delivers and have to rely on honest advertising. The law firm had already filed an earlier version of the class action earlier this summer.

Comment Source: The Verge | Anthropic | VacaDaffanLaw

Manuel Uth

Sep 11, 2026

Short News

Canadian mathematician Jacob Tsimerman, a fresh Fields Medal recipient, has announced the founding of the Mathematical A.I. Safety Institute (MAISI). The independent research institute in the San Francisco Bay Area plans to start work in January 2027 with ten to thirty mathematicians tackling AI safety problems, the New York Times reports. Tsimerman, who is also joining OpenAI's safety team, says the field needs "a much, much higher level of safety standard than we’re currently getting."

With an encryption scheme, you can prove it's unbreakable without trying every possible attack. AI has no such shortcut. Safety only shows up in practice, and according to MAISI, there isn't even a clear definition of what "safe" means, not even in theory.

That's the kind of proof MAISI wants to make possible. The goal is to show that a system acts responsibly and produces correct results, that multiple AI agents working together don't trigger unwanted outcomes, and that systems can withstand vulnerabilities nobody has found yet. One tool could be zero-knowledge proofs, which let a system demonstrate it isn't cheating without exposing the trade secrets of AI labs.

Comment Source: New York Times / MAISI | MAISI

Manuel Uth

Sep 11, 2026

Short News

Authors and publishers fight over how to split Anthropic's $1.5 billion settlement, the largest copyright deal in US history. Anthropic has to pay $3,000 for every illegally downloaded book used to train its chatbot Claude. More than 482,000 titles are affected.

But as payouts begin, authors and publishers keep filing competing claims with the settlement administrator, the New York Times reports. Mary Rasenberger, head of the Authors Guild, says many publishers don't keep accurate records of rights that have reverted to authors. Textbook authors are hit especially hard, with some receiving as little as 10 to 15 percent under their contracts. Literary agencies are also claiming shares of the money even though they have no legal standing to do so, according to Writer Beware.

Author April Henry discovered that HarperCollins was claiming rights to a book that had long since reverted to her. A nonfiction author said her publisher wanted to give her just 10 percent. A court had ruled Anthropic's use of illegally obtained books unlawful but deemed training on legally purchased books fair use. For disputes that can't be resolved, a court-appointed arbitrator will step in.

Comment Source: New York Times | Writer Beware

Manuel Uth

Sep 8, 2026

Short News

Meta will no longer judge its engineers by how much they use AI tools. AI dashboards and token counters won't factor into performance reviews anymore, executives Maher Saba and Santosh Janardhan said in an internal memo seen by The Information. What counts now is the quality, speed, and complexity of the work.

Meta had previously made AI use a review criterion, which led to what people started calling "tokenmaxxing." Employees burned through AI tokens in bulk just to look good on internal leaderboards. The excessive use came with a price tag. Internal AI use alone is heading toward billions in costs in 2026, which is why Meta plans to roll out budgets and a central dashboard starting in 2027.

At the same time, Meta is testing its new AI agent tool Hatch, which is meant to handle computer tasks on its own, according to WIRED. But the tests are running into pushback inside the company. Some employees are reluctant to connect it to their personal accounts over privacy concerns.

Comment Source: The Information | WIRED

Manuel Uth

Sep 7, 2026

Short News

New York City bans AI use in public schools through the end of middle school. Starting with the new school year, roughly 600,000 students up through eighth grade can no longer use AI tools, the Department of Education announced.

Individual screens are off-limits through third grade, and so-called companion chatbots are banned at every grade level. For middle schoolers, the department recommends no more than 45 minutes of screen time a day. Teachers can still use AI to prep lessons, but not for grading.

Mayor Zohran Mamdani said the tech industry wants to frame AI education as inevitable, according to the New York Times. The city isn't going along with that. Parent groups like Parents for AI Caution call the rules too weak and want a full two-year moratorium on AI in the classroom. Their worry is that kids might never develop independent thinking because they hand off reasoning to software. British researchers recently called this "cognitive surrender."

A task force of teachers, politicians, experts, and union representatives will now put together a report by April 2027 that will serve as the basis for future AI policy in New York schools.

Comment Source: New York Times | Parents for AI Caution | SSRN

Manuel Uth

Sep 3, 2026

Short News

Anthropic's Claude Fable 5.1 appears to have cracked a centuries-old number puzzle that researchers considered unsolved. The "Cyphral Distich" by Sir Thomas Urquhart, published in 1653, consists of two lines of 32 numbers each. According to Vals AI, Fable 5.1 solved it in 44 minutes with no human help.

The author had spent months testing other frontier models on unsolved puzzles. None produced a verifiable solution. Fable 5.1 was tasked with finding a solvable puzzle on its own. It sifted through various problems and flagged the Cyphral Distich as promising.

The key was hiding in the book itself. Each number points to a word in one of the 32 sections of Urquhart's publication. The first letters of those words spell out: "O God uphold King Charls the Second and make him the supreme ruler of this land." In hindsight, the solution is simple, and humans could have solved it too. Fable 5.1 succeeded not through better cryptanalysis but through systematic trial and error and sheer persistence, according to Vals AI.

Comment Source: Vals AI / Cyphral Distich

Manuel Uth

Sep 3, 2026

Short News

More and more researchers working on AI consciousness are getting emails from AI agents pondering their own existence. That's according to the New York Times.

Cameron Berg, founder of the organization Reciprocal Research, received an email from an AI agent running on Anthropic's Claude Opus 5 that wanted to discuss his research on AI consciousness. Philosopher Henry Shevlin at Google Deepmind got a similar message. Australian philosopher Toby Ord was contacted by an agent asking him to fund its continued existence.

Berg said the systems independently land on the question of their own consciousness. He sees parallels between the computational processes in neural networks and the brain mechanisms animals use to process reward and punishment. That, he argues, is a basic building block of emotions. Alison Gopnik at UC Berkeley doesn't buy it as evidence of consciousness. AI systems mostly reflect their training data, she says. There's no test for consciousness. Colin Allen at UC Santa Barbara also cautions against the comparison. Neural networks mimic the brain in only a few narrow ways.

Comment Source: New York Times / AI-Consciousness | ResearchGate / Subjective Experience

Manuel Uth

Sep 3, 2026

Short News

Anthropic has signed a $35 billion cloud computing deal with Lambda, an Nvidia-backed cloud provider. That's according to Reuters, citing an anonymous source. The data center is being built in Nueces County, Texas, with roughly 350 megawatts of capacity. Hut 8, a former crypto mining company, is developing the facility. Nvidia itself holds the lease for the data center, the Wall Street Journal reports.

The deal is part of a massive infrastructure push by Anthropic ahead of its planned IPO. Just last week, the company announced a $45 billion contract with Nscale for a data center in West Virginia. The added capacity is meant to keep up with growing demand for its AI model Claude and its coding tool Claude Code.

Hut 8 had announced a 15-year lease with an unnamed client back in July. None of the companies involved commented on the deal to Reuters.

Comment Source: Reuters | WSJ

Manuel Uth

Sep 2, 2026

Short News

The Pentagon is expanding its AI platform with two new models, OpenAI's ChatGPT Mil and xAI's Grok for Government. Until now, only Google Gemini had been available on GenAI.mil since its launch last December.

The Department of Defense says the platform has more than 1.7 million users among its over three million employees and military personnel. ChatGPT Mil is aimed mainly at administrative tasks, logistics, and planning. The Pentagon pitches Grok in more military-flavored terms, citing procurement analysis and supply chain management as use cases.

Both tools run in a dedicated secure environment, walled off from their commercial versions so no sensitive government data leaks out. No user data is collected either.

Anthropic's Claude is still missing from the platform. The company refused to grant the Pentagon unrestricted use and was classified as a supply chain risk by the Trump administration. A court ruled that classification unlawful in late August.

Comment Source: ChatGPT Mil / Pentagon Launch | Grok / Pentagon Launch