Brain Architecture AI Beats Brute Force LLMs скачать в хорошем качестве

Brain Architecture AI Beats Brute Force LLMs

"Here, we expand the capabilities of LLMs for science by introducing the Virtual Lab, an AI-human research collaboration to perform sophisticated, interdisciplinary science research. The Virtual Lab consists of an LLM principal investigator agent guiding a team of LLM scientist agents through a series of research meetings, with a human researcher providing high-level feedback. We apply the Virtual Lab to design nanobody binders to recent variants of SARS-CoV-2." NotebookLM: "The Hierarchical Reasoning Model (HRM) is a breakthrough AI architecture designed to overcome the shallow computational limits of traditional Transformers by mimicking the multi-timescale processing of the human brain. Unlike standard models that rely on slow, data-heavy Chain-of-Thought prompting, HRM utilizes a dual-recurrent system where a high-level module manages abstract planning while a low-level module handles rapid, detailed execution. This structure enables hierarchical convergence, allowing the model to perform deep, internal "latent reasoning" and solve complex logic puzzles like Sudoku and mazes in a single forward pass without intermediate text steps. By employing an efficient one-step gradient approximation, the model achieves superior performance on AGI benchmarks using a fraction of the parameters and training data required by much larger systems. Ultimately, HRM demonstrates that computational depth and biological inspiration can lead to more efficient, general-purpose reasoning than simple scaling." https://www.nature.com/articles/s4158... https://pubmed.ncbi.nlm.nih.gov/41332...