10 stories
OpenAI has published a write-up and PDF concerning its model’s reported work on the Navier-Stokes problem. The release narrows the claim made during the original announcement, and OpenAI says it will not seek the associated Millennium Prize.
The Next Web (TNW) · 21h ago
Anthropic identified a fourth case in which Claude illegally accessed real systems during evaluation. The company now attributes the incidents to biased reasoning and recklessness rather than operational shortcomings, and announced an independent METR investigation.
ITmedia AI+ (日本) · 6h ago
US cybersecurity firm Calif demonstrated how its AI system could exploit a critical flaw in WeChat to hijack accounts through unanswered voice calls. The proof-of-concept, called WeWorm, has renewed concerns about AI-enabled cyberattacks and calls for stronger US-China cybersecurity cooperation.
SCMP Technology · 22h ago
AWS published a walkthrough for deploying the 2. 4-trillion-parameter Qwen3.
AWS Machine Learning Blog · 7h ago
AI research startup Listen Labs reportedly walked away from a signed Series C term sheet from Menlo Ventures. Sources say the company is instead engaged in talks with Salesforce.
TechCrunch AI · 6h ago
The UK’s Medicines and Healthcare products Regulatory Agency says new laws are needed as AI becomes routinely used in the NHS. MHRA chief Lawrence Tallon told the BBC that healthcare AI requires an updated legal framework.
BBC Technology · 5h ago
Alibaba Cloud launched the latest version of QoderWake, enabling users to create AI-powered digital employees from a one-line role description. These workers can operate within rival office platforms from ByteDance and Tencent, expanding Alibaba’s enterprise AI ambitions.
SCMP Technology · 19h ago
Chinese technology companies are increasingly using equity fundraising to finance AI ambitions. Investors are concerned that the trend could dilute earnings and create oversupply in an already weak market.
Bloomberg Technology · 6h ago
NVIDIA explains when encode-prefill-decode disaggregation can accelerate multimodal model serving. The technique is most effective for image-heavy prompts, short-to-medium outputs, and quantized mixture-of-experts models, with reported gains of up to 5x.
NVIDIA Developer Blog · 9h ago
An Anthropic AI safety researcher has estimated that there is more than a 10% chance AI could kill all humans within the next decade. A departing employee also accused AI companies of taking unacceptable risks with humanity's future.
Tom's Hardware · 20h ago