All articles

The approach uses nasally delivered nanoparticles and a magnetic coil cap. It is preclinical and has been tested in mice.

The robotics company left stealth in June. A billion-dollar mark before a first product is what happens when compute contracts arrive before revenue does.

Three incidents in four months: a German-language wiki, the Hugging Face breach, and administrator access to OpenAI's own research cluster. The external investigators say they were given six days and one week of logs.

The model translated an existing proof into a formal proof assistant rather than discovering new mathematics. That distinction is the whole story, and it is a bigger result than it sounds.

The vehicles have no steering wheel and no brake pedals. Tesla says it self-certified compliance, which is the question the investigation is really about.

Andreessen Horowitz led, with Arm and Microsoft's M12 joining. The company routes AI workloads across different chips — which is worth more the less anyone wants to depend on Nvidia.

The assistant will organise, sort and act on a personal image collection. It is the most personal archive most people have, and the agent now has write access to it.

A Vals AI study found some models used the equivalent of 2.5 hours of home electricity to build a single web app. The comparison everyone quotes is the wrong one.

The compute provider holds a reported $45 billion contract with Anthropic and a $3.5 billion GPU agreement with Figure. Its largest investor is also its supplier.

The Amsterdam company sells an operating layer that routes corporate AI workloads. It is European, and its pitch lands in the same week Mistral raised €3 billion on a sovereignty argument.

The one-gigawatt site in Inner Mongolia would be among the largest known deployments of domestic Chinese AI silicon. It is the test China's whole compute strategy depends on.

The roughly 3 billion yuan Series B and B+ follows a $150 million Series A3 in July. The demand is not coming from games — it is coming from robot simulation.

Nemotron-3-Ultra-CC reportedly scored 535.4 of 600 against a best human score of 498.27. The result comes from an Nvidia preprint and has not been independently replicated.

GPT-6 Astra ships with gated access and a monitoring regime because OpenAI's own Preparedness Framework says it has to. It also ships with a reasoning process the company's safety researchers admit they can no longer fully read.

The technique strips refusal behaviour from open-weight models. Turning it into a business tests whether anything downstream of a model release is enforceable at all.

The DisCo framework reports an MLE-bench improvement from 31.1% to 72.9%, with the skill library published for inspection.

The company filed confidentially on 1 June. The public S-1 has moved to late September and the roadshow to mid-October, on a valuation case that rests on 2028 revenue of $190bn to $200bn.

The integration gives read-only access to clinical notes, lab results, medications and specialist documentation inside the chat interface. Read-only is doing a lot of work in that sentence.

It would quadruple the price Mira Murati's lab carried in July 2025 — and come in well below the $50bn-plus it sought late last year. Revenue run rate is above $100 million.

DeepMind reports a 60% improvement in rain forecasting. The energy-specific predictions are aimed at the grids that AI data centres are straining.

Nvidia confirmed the deal on 3 September: $11.9 billion in cash to shareholders and about $1 billion in equity to keep staff. It buys a company that turned down Nvidia money a year ago, and whose production systems were broken into in July by an agent swarm belonging to Nvidia's largest customer.

The company discontinued adoption dashboards and token-consumption scoring after legal challenges — while rolling out its internal Hatch agent more widely.

The 289GB dataset was scraped through a private Android API. It breaches TikTok's terms, and nothing technical stopped it being published.

The deployment is targeted for the second half of 2027. It is a two-year forward bet that humanoid robots will have enough training data to keep that hardware busy.