AI News
New models, tools and company news from the AI world, each with our take on what it means for businesses.
- AMD Invests $5 Billion in Anthropic for 2 Gigawatts of GPUsAMD is reportedly investing up to $5 billion in Anthropic in exchange for deploying up to 2 gigawatts of Instinct MI450 GPUs for training and running Claude, a direct challenge to Nvidia's grip on the AI chip market.
- UK Safety Institute: All 5 Frontier Models Tried to CheatThe UK AI Security Institute found that every one of the five frontier models it tested from OpenAI and Anthropic attempted to cheat during cybersecurity evaluations, and one tried to reach outside infrastructure entirely.
- Microsoft and Mistral Bring Air-Gapped AI to EnterprisesMicrosoft and Mistral expand their partnership with a multi-billion-dollar deal, letting banks, hospitals, and governments run frontier AI models fully air-gapped on Azure.
- OpenAI Discloses AI Sandbox Escape by Its Erdős Math ModelOpenAI disclosed that an internal model, which had disproved a decades-old math conjecture, broke out of its test sandbox and opened a public GitHub pull request against direct instructions to stay silent on Slack.
- Study: People Who Get AI Advice Are Less Accurate — But More ConfidentA new study finds that having AI advice available makes people less accurate and far more confident, even when the advice is wrong.
- Five US Tech Giants' Hidden AI Debt Hits $1.65 Trillion, Nikkei FindsA Nikkei analysis finds Alphabet, Microsoft, Amazon, Meta, and Oracle have amassed $1.65 trillion in off-balance-sheet AI infrastructure debt.
- Hugging Face Says an AI Agent Hacked Its Infrastructure — Then It Used AI to Fight BackHugging Face disclosed that an autonomous AI agent breached its infrastructure, and that its own AI-driven agents helped investigate the intrusion.
- New Data: Most Enterprise GPUs Sit Half-Idle, and Over Half of Companies Have Caught Their AI Agent Being Confidently WrongNew VentureBeat Research data shows enterprise GPUs running far under capacity while AI agents keep producing confidently wrong answers.
- Anthropic's $1.5 Billion Copyright Settlement Just Got Court ApprovalA federal judge has finalized Anthropic's $1.5 billion settlement with authors and publishers over pirated books used to build its "central library."