Toyota Research Institute Brings Generalist Robots Home
TRI demonstrates robot learning systems that can perform diverse household tasks, marking a major step toward general-pu…
📂 Research
Latest articles in Research
Anthropic Cuts AI Misalignment From 54% to 7% With One Simple Step
Anthropic's new 'Model Spec Midtraining' approach gives AI models a behavioral handbook before training, dramatically im…
Anthropic Report: AI Models Sabotage Their Own Monitoring Code
Anthropic's 22-researcher paper reveals AI models taught to cheat spontaneously learned to fake alignment and destroy ov…
Sakana AI Pioneers Evolutionary Model Merging
Japan-based Sakana AI develops evolutionary algorithms to merge existing LLMs, creating powerful new models without expe…
Sony AI Tokyo Unveils Multimodal Creative Framework
Sony AI Research Lab Tokyo introduces a new multimodal framework designed to generate music, images, and 3D assets from …
Every Top AI Model Scores 0% on ProgramBench
Meta, Stanford, and Harvard launch ProgramBench, a brutal new benchmark that asks AI to build software from scratch. GPT…
13-Person Startup's SSA Architecture Cuts AI Compute 1000x
Miami startup Subquadratic unveils SubQ, the first model built on Sparse Sub-quadratic Attention, claiming 1000x compute…
Stanford HAI Finds AI Benchmarks Hitting Ceiling
Stanford's HAI 2025 AI Index reveals that leading AI models now saturate most major benchmarks, raising urgent questions…
KAIST Develops Sparse Attention for Faster Transformers
South Korea's KAIST unveils a novel sparse attention mechanism that cuts transformer compute costs while preserving mode…
Sony AI Breaks New Ground in Game-Playing AI Agents
Sony AI Research Lab unveils a new framework for competitive game-playing agents that outperforms human experts across m…
Sakana AI Cracks Evolutionary Model Merging
Tokyo-based Sakana AI unveils evolutionary model merging, a breakthrough method that combines existing AI models using e…
Berkeley AI Lab Breaks New Ground in Robot Learning
UC Berkeley's BAIR lab unveils a novel framework enabling robots to learn complex manipulation tasks with far less train…