Sakana AI Bets Agents Can Judge Each Other Where Humans Can’t
Sakana AI published a multi-agent self-supervision method called MASS, aiming to let AI systems keep improving on tasks too complex for humans to grade.
AI news and analysis covering the labs, models and infrastructure behind the industry: what OpenAI, Anthropic, Google DeepMind, Meta and Mistral ship, and what each release changes for the people building on top of it.
Sakana AI published a multi-agent self-supervision method called MASS, aiming to let AI systems keep improving on tasks too complex for humans to grade.
Solvely.ai uses Anthropic’s Claude to build interactive tools that test whether students truly understand concepts, not just final answers.
AI capability gap: Andrej Karpathy says the gap in understanding AI capability has widened since April, turning into what he calls a sharp funnel.
Math Problems: Hugging Face turned OpenAI’s math problem set into open-source RL environments on Oct.
Nvidia is weighing a deeper tie to open-model developer Reflection AI through investment or a computing deal, following its SpaceX and Nebius pacts.
OrcaRouter released OrcaCyber Zero 1.5, a cybersecurity-focused AI model scoring 100% on Cybench and 95.8% on a CVE-Bench subset.
KONST raised a $30 million Series B led by ADATA Technology to expand AI data center capacity across Asia and grow its token factory model.
Recorded Future launched free AI Infrastructure Indicator Lists for Cyber Operations customers, aimed at detecting shadow AI tools employees install without IT.
AI inference chips now drive more demand than training chips. Here’s how Nvidia and AMD split the market and why the fight has shifted.
A hidden “Preview” voice tab found inside Claude’s app points to Anthropic building a native voice feature, though it remains unconfirmed and unreleased.