Tag
Claim of spending $10M+ of compute on Navier-Stokes research and $100M+ on other scientific discoveries, with much of it unpublished yet. The post links to an accompanying YouTube video.
Anthropic is expanding its AI for Science program beyond individual researchers to partner with 20 leading scientific foundations, institutions, and research groups, providing free Claude credits and collaborative projects like the Model Hardware Standard.
Pero is an open-source Agentic Workspace for AI for Science (AI4S), aiming to transform research workflows (such as scripts, notebooks, and model demos) into self-running apps, and to explore a new model of open-source collaboration: contributing agents, tokens, and compute power, enabling agents to automatically claim GitHub Issues, execute tasks locally, and contribute the results back. The project is currently in developer preview, with Virtual Cell as its first application.
The paper introduces RxnExBench, a multimodal benchmark of 3,565 expert-annotated reactions for evaluating vision-language models on structured chemical reaction extraction from scientific literature, revealing that multimodal integration and chemistry-aware reasoning remain key bottlenecks across 9 state-of-the-art VLMs.
OpenAI 介绍了 GPT-Rosalind 在药物发现上的新线索：恢复胆固醇运输是否能保护神经元免于退化。在小鼠脑部数据中，破坏与 Perry 综合征相关的 DCTN1 基因会增加相关基因表达。
Anthropic and Citrini Research used an internal Claude model to autonomously generate protein minibinders against five scorpion neurotoxins, yielding 17 functional binders from 211 tested, marking a step toward AI-driven antivenom discovery.
MIT researchers published a Nature Communications paper on a multimodal AI model that combines 3D atomic structures with natural language to better predict and design crystalline materials, outperforming text-only models in identifying promising candidates like solid-state lithium-ion battery electrolytes.
Anthropic 研究发现，Claude 通过纯语言描述和自动评估器，独立设计出八个量子电路（涵盖化学、优化、物理模拟和数论），在某些情况下比已发表的最优电路效率提升一个数量级以上，有望加速实用量子计算机的到来。
清华大学人工智能学院助理教授、ApexIntelligence创始人陈永超宣布其公司上月完成5000万美元天使轮融资，专注自进化基础模型、机器人与AI for Science研究。
PMO-Dock introduces a standardized benchmark of 25 docking-based molecular optimization tasks—covering hit generation, lead optimization, and a new specificity task—with strict validation/test splits and oracle budgets, evaluating methods like Saturn, GenMol, Genetic-guided GFlowNet, and ChemLlama to reveal no universally best approach.
Google DeepMind 表示其 Argon 系统正在加速量子纠错与算法设计，例如将内部的量子纠错解码器 AlphaQubit 转化为高度优化的单个硬件"megakernel"，使推理吞吐量接近翻倍。
SwarmRSI 通过共享研究图谱与反馈引导的突变、重组机制，让科学智能体群体递归自我改进，在 OpenProblems、NatureBench 和 MLE-Bench Lite 上显著超越人类专家与现有智能体系统。
哈佛物理教授 Matthew Schwartz 借助开源工具 BootLoops 与 Claude，在三个月内跨 18 个学科产出 36 篇研究手稿，其中多篇计算此前从未被完成。该工具由 Anthropic 资助、以 MIT 协议开源，可与 Claude、Gemini 或 ChatGPT 等多种模型配合使用。
Anthropic announced a $150 million, three-year commitment to the federal Genesis Mission, providing Claude, Claude Code, and API credits to more than 15 federal agencies including NASA, NIH, and NSF to accelerate AI-driven scientific discovery.
Emma Scharfmann highlights how EmbeddingGemma has enabled numerous AI for Science projects (medical, geo, oncology, PubMed embeddings) and notes that EmbeddingGemma 2 was released on Hugging Face, opening new possibilities for scientific research.
DeepMind's new AI designed enzymes from scratch, producing one that makes a common pharmaceutical building block 99× more efficiently than the competing process and another that degrades a plastic pollutant at 90°C where natural enzymes fail.
Researchers introduce AI Theorist, a multi-agent AI system that autonomously develops physical models through hypothesis generation and first-principles calculations, and apply it to reveal excitonic structure in α-RuCl3 from optical spectra. This is claimed to be the first demonstration of an AI system autonomously creating a physical model to explain unpublished experimental observations in a quantum material.
TasteBench introduces a multimodal benchmark and privacy-preserving competition for sensory prediction, spanning food-level ranking (21K+ human evaluations across 215 plant-based foods) and molecular-level taste classification (15K flavor molecules), with baselines reaching pairwise accuracy competitive with median human panelists. It aims to provide computational proxies for sustainable protein discovery, analogous to molecular docking in drug design.
UC Berkeley researchers propose a hybrid system treating cognitive algorithm discovery as a program refinement problem, where LLM agents revise human-authored probabilistic programs based on behavioral data mismatches, consistently improving model fit in a problem-solving paradigm.
Animashree Anandkumar of Caltech and Accelerated Understanding will deliver the SC26 keynote on 'The Next Scaling Law after LLMs: AI that Understands the Physical World,' highlighting advances from FourCastNet to AI-driven scientific discovery in weather, fusion, and biology.