Tag
PyTorch Conference North America 2026 在圣何塞举行,汇集顶尖 AI 先驱、研究者与开发者,围绕 PyTorch、vLLM、DeepSpeed、Ray、Helion、Safetensors 等开源项目展开技术分享与交流,并感谢黄金赞助商的支持。
Construction Physics 探讨机器人 AI 面临的核心瓶颈:语言模型可依赖海量现成文本数据,但机器人缺乏等价的公开训练数据,文章梳理了数据来源、数据量需求及模拟仿真等潜在解决方案。
Abdur-Rahman Butler (Isaacus) 制作了一个交互式教程加 Wordle 风格游戏 SemanticGuessr,通过在 43,339 个英语单词的二维语义地图上「围捕」目标词来讲解向量嵌入、余弦相似度、语义搜索、聚类等概念,并通过降维量化把模型压缩到 12.3MB 实现纯前端运行。
本文提出 PIDiff,一种联合离散序列空间与 SE(3) 结构流形扩散的多模态框架,用于蛋白质-蛋白质复合物设计,并提出 PIFormer 架构与 PIBench 数据集,在 minibinder 和抗体设计任务上超越基线方法。
FDE-Bench 是首个基于真实客户项目构建的基准,用于评估 AI 代理在需求不明确的真实商业请求下进行端到端交付的能力,涵盖 49 个组合优化与机器学习案例。结果显示,18 种代理配置中最高平均交付分仅为 0.5510(FDE 参考为 1.0),最高交付成功率仅 6.12%,表明代理尚未达到 FDE 级别的端到端交付水平。
A US Senate Joint Economic Committee report reveals the scale of romance scams on Match Group's dating apps, finding Tinder had not updated most of its machine-learning models in years and that nearly 1.5 million scam accounts matched with real users through Q3 2025.
微软研究院发布并发表于 Nature 的机器学习框架 Probabilistic Bias Correction (PBC),通过学习现有天气模型的系统性误差并加以校正,显著提升两到六周的次季节天气预报准确性,在全球竞赛中夺冠,并已用于肯尼亚季风预报等实际场景。
MIT researchers and collaborators published a Nature paper describing a machine-learning model that corrects errors in existing weather forecasts, doubling accuracy in the two-to-six-week range and beating climatological averages for the first time. The approach also improves prediction of extreme weather events like flooding and cold spells, placing first in an international forecasting competition.
开源仓库 AIMLInterviews 更新至 2026 版蓝图,覆盖 GenAI、Agentic AI、多模态与 LLM 系统设计,并提供可接入终端或 Claude/Cursor 的 MCP AI 导师进行实时模拟面试。
MIT researchers published a PNAS paper showing that artificial neural networks can build their own learning machinery through simple local rules, reshaping randomly connected units into effective information-processing networks without a developer's guiding hand, offering new insights into how the brain learns.
A Twitter exchange speculating whether GPUs will eventually rent more GPUs to train improved versions of themselves, while noting that insufficient compute and token generation will remain the bottleneck for AI progress for years.
A Twitter thread providing additional context on an interview with Michael I. Jordan, expressing discomfort with the outsized attention paid to existential risk and the dominance of the public AI discussion by a few prominent voices, in which Jordan calls superintelligence science fiction.
AI 已在22个科学领域中发现了此前“隐藏在明处”的解决方案,表明跨学科AI辅助发现具有广泛潜力。
提出一种与坐标系无关的自动微分形式化方法,将几何代数作为代数数据引入 AD,使得单一的基无关导数算子能够表示多种几何微分算子,并可与 JAX 的批处理、编译和硬件加速组合使用,应用于科学 AI 中的流形算子和时空反问题管线。
The paper shows that random data splits in perovskite photovoltaics ML benchmarks systematically leak information due to massive internal duplication in the Perovskite Database, and proposes a provenance-grouped evaluation protocol. Benchmarking thirteen models reveals tree ensembles fail on truly unseen institutions while composition-aware neural architectures excel, establishing new state-of-the-art results.
该论文在一所大学建筑上对比了统计模型、XGBoost、LSTM与预训练时间序列基础模型TimesFM用于日前电力预测的效果,发现TimesFM 3.0配合协变量将平均绝对误差相对季节性朴素基线降低约43%,而XGBoost仍具竞争力。
该论文研究了在结果依赖截断和删失条件下的生存分析模型高效学习问题,证明了在已知基线风险时 Cox 回归可通过一阶优化高效学习,同时构造了存在严格次优局部极小值的可识别模型,并在实验中展示了相比期望最大化方法的数量级加速。
Task-structured modularity emerges in artificial networks and aligns with brain architecture
作者回顾其2.5年前关于'The Bitter Lesson'的演讲,讨论如何在日常工作中实践该理念:先为当下进展添加结构化方法,再在条件变化时及时移除这些结构。
Maxime Rivest回复网友关于如何入门模型微调的提问,给出了从简单分类任务起步、逐步构建自定义任务、利用coding agent辅助实验的五步学习路径建议。