In a groundbreaking achievement, scientists have developed an AI model that mimics the mouse brain's visual cortex, enabling virtual neuroscience experiments with the potential to accelerate research by accurately predicting the responses of tens of thousands of neurons to new visual stimuli.
Researchers employ Prolog, a logic programming language, to ingeniously solve a cryptic video game puzzle, showcasing the language's bidirectionality, pattern matching prowess, and ability to define predicates for calculating test scores and deducing the answer key.
BrowseComp, a novel AI benchmark, evaluates web browsing skills by posing 1,266 complex queries requiring synthesis of information across multiple websites, with a deep research model achieving 51.5% accuracy, surpassing other models' performance on this challenging task.
Researchers developed a multimodal AI tool that combines language models and graph-based models to efficiently design new molecules with desired properties from natural language queries, outperforming existing methods and potentially automating the entire molecule design and synthesis process.
Meta's new AI model Maverick tops the LM Arena benchmark, but the optimized version tested raises transparency concerns as it differs from the publicly available model, making real-world performance predictions challenging due to undisclosed customizations.
A study on US manufacturing reveals short-term productivity losses but long-term gains from AI adoption, with older businesses facing initial challenges mitigated by growth strategies and within-firm spillovers, highlighting the transformative potential of AI amid transitional pains.
Scientists developed a starfish-shaped wearable device that leverages AI to precisely track heart health in real time with over 90% accuracy, even during movement, enabling comfortable long-term at-home monitoring with wireless charging and skin-friendly materials.
Transformer models, neural network architectures employing self-attention mechanisms, have revolutionized natural language processing tasks, achieving remarkable performance in machine translation, question answering, and text generation through models like BERT, GPT, and ChatGPT.
Researchers developed an end-to-end neural network trained with reinforcement learning, enabling robots to learn human-like walking in simulation and seamlessly transfer this capability to real hardware without additional tuning, achieving robust and natural locomotion across multiple robots.
OpenAI's groundbreaking o3 model, combining deep learning and program search, shattered records with a remarkable 75.7% score on the ARC-AGI benchmark, showcasing unprecedented task adaptation capabilities, yet falling short of true AGI as tougher challenges loom.