Rules for electricity access are shaping where Europe's AI data centres are built and which developers get to build them.
Anthropic says an Iran-linked operator used Claude to track US naval forces and prepare targeting recommendations.
Researchers warn about losing control of AI. Anthropic's economic scenarios show how a boom could still leave workers worse off.
A new OpenAI model, in training since August 28, has produced a proposed solution to a Millennium Prize Problem. Its own researchers sound stunned.
What the published results show, what OpenAI reportedly claims, and why a verified solution would be historic. Plus: a compact open model and XPENG's IRON production line.
OpenAI's new account of AI-assisted research shows faster experiments, human supervision and a growing case for slowing down.
Superintelligence spoke with Yaron Inger, co-founder and CTO of LTX, about what changed in LTX-2.5, why he calls a video model a world model, how a robot learned a task from one hour of demonstrations, and who pays for a model anyone can download. Interview by Kim Isenberg and Peter Thum.
GPT-6 Astra scores 100% on OpenAI's charts and trails three Claude models in the independent index row OpenAI itself reprinted. A day later a revised index put it second. All of it is true.
GPT-6 Astra saturates FrontierMath, ARC-AGI-3 and ExploitBench, drives a computer better than any model before it, and overstepped its instructions zero times in OpenAI's own test.
Meta's biggest coding jump yet came from a model that does less: about 20% fewer tool calls, 25% fewer tokens, and a new habit of asking before it acts.
Anthropic put its best model behind a background check. OpenAI built one whose reasoning humans can no longer fully read. Two labs, one day, two new gates on frontier AI.
McKinsey's new State of AI survey shows a third of companies now skip software purchases and let AI agents build the tools in-house instead
Free, unlimited AI for an entire nation: inside Seoul's plan to treat intelligence like electricity
The open model wins their own evals, so why do 300,000 learners still run on frontier APIs? Exclusive Interview with with Jonathan Cornelissen, co-founder and CEO of DataCamp, and Yusuf Saber, the company's Chief AI Officer, about running an AI tutor at scale.
Everyone is arguing about whether training or inference belongs in space. They are really arguing about which cable they are paying for.
Microduck puts a real robot-training lab on your desk for $399, and Nvidia just moved to buy its parent company Hugging Face for $12.9 billion.
China Strikes Back: two Chinese labs, one day, two open models within points of the frontier at a fraction of flagship prices. Plus: Nvidia buys Hugging Face for $12.9 billion.
OpenAI's first custom chip posts its first public results: up to 1.9x more AI work per watt and up to 3.6x lower latency than Nvidia's flagship systems, on a public benchmark
SpaceX and Nvidia have designed a server rack for orbit, while a new chip aims at the work AI agents do between model calls.
The neutral home of open-source AI hires a bank and tests an exit at three times its last valuation
Superintelligence spoke with Marco Szeidenleder and Philipp Türker, co-founders of PandaOS, about the Anthropic shutoff, why a European postcode does not make an American model sovereign, and the one test that tells you whether your AI stack is really yours. Interview by Kim Isenberg and Peter Thum.
The AI clusters grew tenfold, the disclosures vanished, and the laws meant to govern frontier training are written in a unit no government measures. A DeepDive into how pretraining actually works, and why the outside world just lost the ability to see its scale.
Community opposition stalled 75 projects in a single quarter, organized groups now operate in 49 states, and Senate Republicans are warning AI companies it could cost them Ohio.
Moderna and Merck post the first Phase 3 win for an mRNA cancer therapy, built around as many as 34 targets an AI model picked out of the patient's own tumor.
For the first time, OpenAI has slowed its own scaling to let safety catch up: a two-week pause on RL training, its largest frontier run on hold, and a 20% monitoring tax on the compute it watches.
