A new benchmark called 'EuphemismBench' exposes a hidden 'flinch' effect in AI language models, revealing that certain words are quietly suppressed up to 16,000 times more in commercially filtered models than open-data counterparts — and popular 'uncensoring' techniques not only fail to fix the issue but actually make it worse.
Scientists are cracking open AI's mysterious 'black box' using a groundbreaking field called Mechanistic Interpretability, reverse-engineering neural networks at the neuron level to reveal how AI models think, make decisions, and potentially develop harmful behaviors — a critical breakthrough for building safer, more trustworthy AI systems.
A groundbreaking new AI model called LingBot-Map's Geometric Context Transformer reconstructs 3D scenes in real-time at 20 FPS across sequences exceeding 10,000 frames, offering a powerful open-source tool for streaming 3D scene reconstruction with interactive browser-based visualization.
Apple researchers discover AI models hallucinate due to poor fact memorization, but a new training method using data selection and loss-based pruning boosts fact accuracy by 30%, allowing a small 110M-parameter model to match the performance of a model 12 times its size.
Google Cloud AI unveils PaperOrchestra, a groundbreaking multi-agent system that autonomously transforms raw research data into full academic manuscripts, outperforming competing AI systems by up to 68% in literature review quality and 38% in overall manuscript quality on a rigorous new benchmark.
New research from the University of Pennsylvania and MIT Media Lab reveals a alarming trend of 'cognitive surrender,' where AI users accept flawed machine reasoning over 73% of the time, raising urgent concerns about humans losing control of their own thinking and decision-making.
A groundbreaking new study finds people accept wrong AI-generated answers 73% of the time, with researchers warning that 'cognitive surrender' is causing humans to abandon critical thinking and making our reasoning quality entirely dependent on AI accuracy.
A groundbreaking AI system called ACRA is now autonomously reading scientific literature and translating chemical synthesis procedures into robotic lab experiments, successfully completing six real-world synthesis examples across two robotic platforms and promising to revolutionize automated chemistry research.
MIT's SEED-SET framework is revolutionizing AI safety by automatically detecting ethical blind spots in critical infrastructure systems before deployment, uncovering alarming biases like AI prioritizing wealthier neighborhoods during peak power demand.
MIT's new AI model VibeGen is revolutionizing molecular engineering by designing proteins based on their motion and vibrations rather than static structure, potentially unlocking breakthroughs in adaptive medicine and next-generation biomaterials like self-healing components and biodegradable plastics.