Meta releases Code World Model, a groundbreaking 32-billion-parameter AI that learns how code actually executes rather than just syntax, achieving a 65.8% pass rate on coding benchmarks by developing internal world models that enable self-verification and step-by-step simulation.
Major AI systems from Google DeepMind, Microsoft, and Meta generate millions of promising new materials, but scientists discover 80-84% contain fatal flaws that make them impossible to synthesize in reality.
Meta's V-JEPA AI learns physics from videos like a human baby, achieving 98% accuracy on reality tests by discovering object permanence and showing measurable surprise at impossible events, though its memory currently lasts only seconds.
Hugging Face co-founder Thomas Wolf claims current AI models are too agreeable and consensus-driven to achieve Nobel Prize-level scientific breakthroughs, arguing that while AI excels as research assistants, transformative discoveries require the contrarian thinking that AI systems fundamentally lack.
Microsoft researchers discover AI can redesign deadly toxins to bypass DNA screening software and create undetectable biological weapons, prompting emergency patches to security systems as experts warn of an escalating arms race between AI-enabled bioterrorism and defense capabilities.
New research reveals that switching from CSV to Markdown-KV format dramatically improves AI language model accuracy by 37%, achieving 60.7% success rate compared to CSV's 44.3% when processing employee data, though the superior format requires nearly three times more computational tokens.
RNA structure prediction remains far behind protein prediction as AlphaFold3 struggles with RNA's extreme flexibility, with only 7,759 RNA structures mapped versus 216,212 proteins, though most RNA therapeutics like mRNA vaccines rely on secondary structure which scientists can now predict with high accuracy.
Scientists develop groundbreaking 'manifold Muon' algorithm that prevents neural networks from breaking down during training by constraining weight matrices to maintain unit condition numbers, achieving higher accuracy than existing methods in early tests.
A Microsoft and University of York study finds that AI in-context learning technically qualifies as 'true' learning, but is dangerously brittle — with LLMs prioritizing statistical patterns over meaning so heavily that nonsense prompts perform nearly as well as meaningful ones, raising urgent concerns about real-world reliability.
Cutting-edge AI models now pass the grueling CFA Level III exam in minutes—a financial certification that demands 1,000 hours of human study—with researchers testing 23 language models and finding that advanced systems like o4-mini and Gemini 2.5 Pro successfully tackle complex analytical reasoning that stumped AI just two years ago.