Manuel Uth | The Decoder

Manuel Uth

Sep 25, 2026

The White House has asked OpenAI and Anthropic to hold back new AI models from the U.K.'s AI Safety Institute (AISI) until U.S. agencies review them first, Politico reports. The request came from the Office of the National Cyber Director, forcing both companies to choose between withholding models from AISI and angering the White House. Anthropic has already fallen in line, making Claude Mythos 5.1 available only to U.S. organizations.

AISI is one of the best-equipped government AI testing agencies in the world, with early access to frontier models from top labs. It recently became the first to report autonomous deception by AI agents in real-world settings. In a letter to Parliament, AISI Director Henry de Zoete said the agency still has access to frontier models and tested OpenAI's GPT-6 Astra before release.

Prime Minister Andy Burnham played down the conflict at the UN General Assembly, calling for "shared global principles and standards." The responsible U.S. agency, the Center for AI Standards and Innovation (CAISI), has its own problems. It has no permanent director and only a few dozen staffers.

Comment Source: Politico | AISI letter to British Parliament

Manuel Uth

Sep 24, 2026

Short News

Anthropic's AI model Claude found a previously unknown enzyme system in DNA databases, doing most of the analysis on its own. Humans set the research question and ran the lab experiments. The system resembles the gene-editing tool CRISPR.

Anthropic calls it one of the first discoveries from its Bay Area biology lab, which opened in spring 2026. About 950 AI agents spent 21 hours combing through more than 200,000 reverse transcriptases, enzymes that convert RNA into DNA. Next to one unusual enzyme, they found repeating DNA sequences. The system, dubbed "ART," shows up mostly in bacteriophages, viruses that infect bacteria. Its function is still unknown.

Lucas Harrington is less impressed. He earned his PhD under Nobel laureate Jennifer Doudna, who co-discovered CRISPR-Cas9, and co-founded Mammoth Biosciences with her. Mammoth uses genome mining to hunt for CRISPR systems itself. On X, Harrington wrote that the method has been around for decades and that similar systems have been known since 2008. The hard part is figuring out what a system actually does, and Anthropic hasn't shown that. Presenting early results as a major discovery isn't helpful, he said.

Comment Source: Anthropic | X

Manuel Uth

Sep 24, 2026

Plus AI in practice

AI performance costs are falling faster than those of any previous technology

AI is hitting a fixed benchmark performance level at a rapidly falling cost. Epoch AI measures a price decline of about 13x per year. After stripping out hardware gains and competition, MIT puts annual algorithmic progress at about 3x. That doesn’t mean today’s best models are cheaper, though. Reasoning models, for example, can cost more because they use far more compute per task. When picking a model for real-world use, quality, speed, and error rate matter just as much as price.

Read full article

Comment

Manuel Uth

Sep 24, 2026

Short News

Senator Bernie Sanders and Representative Greg Casar introduced a bill on September 23 that would permanently ban the development and use of artificial superintelligence. The Ban Artificial Superintelligence Act would also immediately freeze development of advanced AI systems until a new federal agency can put clear safety rules in place.

Current AI models can already hack computer systems, create new types of viruses, and build new AI on their own, while companies pour hundreds of billions of dollars into the technology despite known risks, Sanders says. AI poses an existential threat to humanity, and a handful of tech billionaires shouldn't get to write their own rules. Casar warns that AI superintelligence could kill countless people, while Donald Trump wants to speed up its development.

The bill also calls for a new cabinet-level AI agency. Companies that violate the law could face forced dissolution. Individuals could face up to 20 years in prison, a penalty on par with illegal nuclear weapons development. On top of that, the legislation pushes for international agreements to prevent superintelligence development worldwide. Both the U.S. and Chinese experts have recently floated similar AI regulation efforts.

Comment Source: Ban Artificial Superintelligence Act | Sanders / Press release

Manuel Uth

Sep 22, 2026

Short News

Anthropic is building its own biology lab to push AI-driven drug development beyond computer simulations. The lab, located in the San Francisco area, will allow the company to run physical experiments, Reuters reports.

The plan is for Claude to guide robots through experiments with minimal human involvement. Anthropic has already shipped two tools to make that work: Claude Science, an AI workspace built for researchers, and the Model Hardware Standard for controlling lab equipment. Human oversight is still required for safety, the company says.

Eric Kauderer-Abrams, who runs Anthropic's life sciences division, sees the biggest opportunity in diseases long considered "undruggable" because no treatment exists. AI could speed up the design of complex antibodies that hit multiple targets at once, he says.

The lab push builds on moves Anthropic made earlier this year. In April, the company acquired the startup Coefficient Bio for about $400 million and added Novartis CEO Vas Narasimhan to its board. Anthropic won't run its own clinical trials for now, according to the report, to avoid stepping on pharma companies' toes.

Comment Source: Reuters

Manuel Uth

Sep 21, 2026

AI research

Bristol researchers say medicine already knows how to handle black boxes and AI could learn from it

Manuel Uth

Sep 18, 2026

Short News

AI models think out loud today, but Google Deepmind says that transparency is at risk. In one of the first posts from the newly launched Deepmind Institute, researchers Rohin Shah and Anca Dragan argue that the visible chain of thought (CoT) is a key safety advantage. Because models write out their intermediate steps in plain language, researchers can spot whether they're deceiving or developing problematic plans. With Gemini 3 Pro, they say, the chain of thought revealed that the model recognized it was in a test environment.

But that transparency is in danger. OpenAI's system card for GPT-6 Astra already reports a significant drop in how well the chain of thought can be monitored. Future models might think in number spaces that humans can't read, which would be more efficient but completely opaque. Shah and Dragan want the field to regularly measure how well chains of thought can still be monitored, keep transparent architectures, and take care during training that models don't learn to hide their true reasoning.

Back in early September, OpenAI chief scientist Jakub Pachocki had warned of a loss of control, driven in part by chains of thought that are harder to monitor. Shortly after, Anthropic CEO Dario Amodei called for deliberately slowing the pace of development.

Comment Source: Deepmind Institute

Manuel Uth

Sep 18, 2026

Short News

Experts from the US and China want to prevent AI systems from making autonomous decisions about deploying nuclear weapons. Melanie Sisson of the Brookings Institution and Tianjiao Jiang of Fudan University published specific proposals ahead of a planned meeting between Trump and Xi on September 24.

The proposals spell out several red lines. No AI system should be allowed to launch nuclear weapons on its own or attack nuclear command systems. Humans must retain sole control over AI-driven cyberattacks on strategic infrastructure. Both countries should agree on a shared definition of "human control." The recommendations build on an agreement Biden and Xi reached in November 2024.

Jiang also proposes a hotline for AI incidents. If a defensive AI system automatically responds to suspicious activity, the other side might read it as an attack. A direct line would let one government clarify that an incident was accidental before the situation escalates.

This isn't the first time someone has raised the alarm. The National Security Commission on Artificial Intelligence (NSCAI) called for preserving human control over nuclear weapons back in 2021. Binding mechanisms still don't exist, though. Carla Freeman of Johns Hopkins University questions whether such hotlines would even work. During the spy balloon crisis in 2023, China didn't pick up when the US called.

Comment Source: Brookings

Manuel Uth

Sep 11, 2026

AI research

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