2025年初，国内外AI企业接连推出炙手可热的推理模型，人工智能技术迈入加速发展的新阶段。AI产业成为推动全球经济社会发展的重要引擎，催生大量新产业、新模式及新业态。然而，在AI大模型应用的新纪元下，数据安全、隐私保护、算法偏见等问题依然突出，在就业影响、伦理安全和国家治理等方面也引发了人们的广泛讨论。 一 时政热词 1、Large Language Model（LLM）大语言模型 While early language models could only process text, contemporary large language models now perform highly diverse tasks on different types of data. For instance, LLMs can understand many languages, generate computer code, solve math problems, or answer questions about images and audio. 早期语言模型仅能处理文本数据，而当前的大语言模型（Large Language Models, LLMs）已能在多模态数据上执行高度多样化的任务。例如，LLMs 能够理解多种自然语言、生成代码、解决数学问题，甚至可基于图像和音频内容进行问答。 （英文摘自2025.02.15 MIT News网站） 图片来源：央视网 What has surprised many people is how quickly DeepSeek appeared on the scene with such a competitive large language model – the company was only founded by Liang Wenfeng in 2023, who is now being hailed in China as something of an “AI hero”. 令许多人惊讶的是，DeepSeek公司在如此短的时间内就凭借极具竞争力的大语言模型崭露头角——该公司2023才由梁文峰创立，他在中国被誉为“AI英雄”。 （英文摘自2025.02.01 BBC News网站） 2、Mixture of Experts（MoE）混合专家 字节跳动的低成本优势得益于其采用了“混合专家”（MoE）框架，该框架在中国多家人工智能公司中得到广泛应用，其中就包括DeepSeek模型在内。 （英文摘自2025.02.14 Reuters网站） The key to DeepSeek’s frugal success? A method called “mixture of experts.” Traditional AI models try to learn everything in one giant neural network. That’s like stuffing all knowledge into a single brain—inefficient and power-hungry. DeepSeek成功的关键是什么？是一种名为“混合专家”的方法。传统的人工智能模型试图将所有信息整合到一个巨大的神经网络中。这就像将所有知识塞进一个大脑——效率低下且消耗大量能源。 （英文摘自2025.02.16 The Times of India网站） 图片来源：CGTN 3、Open Source 开源 (Paris AI Action Summit’s) official themes include AI for public interest, global AI governance, the future of work, and innovation and culture. There will also be a focus on sustainability. Open-source AI technology will be a key theme, following DeepSeek’s success in this way of working. （巴黎人工智能行动峰会）官方议题包括公共利益人工智能、人工智能全球治理、未来工作及创新与文化等。可持续发展也将成为讨论重点。DeepSeek大获成功之后，开源人工智能技术将成为重要议题。 （英文摘自2025.02.10 Euronews网站） Chinese artificial intelligence (AI) startup DeepSeek launched its Open Source Week initiative and released its first code repository, FlashMLA, on Monday to share their “small but sincere progress with full transparency.” 周一，中国人工智能初创公司DeepSeek启动了“开源周”计划，并发布其首个代码库FlashMLA，“尽管进展不大，但态度真诚且公开透明”。 （英文摘自2025.02.26 China Daily网站） 图片来源：DeepSeek 4、Embodied AI 具身智能 The HKIC will host the first International Conference on Embodied AI Robot to pool together top notch technology enterprises, academic institutions and investors to showcase the latest R&D outcomes and application scenarios, thereby boosting Hong Kong's global influence on technology areas. 香港集思会（HKIC）将举办首届“国际具身智能机器人大会”，汇聚顶尖科技企业、学术机构和投资者，展示最新研发成果和应用场景，以提升香港在全球科技领域的影响力。 （英文摘自2025.02.26 Bastille Post网站） Meta for years has been funding research into “embodied AI,” hoping to develop AI assistants that can look at, listen to and navigate the 3D physical world around them. 多年来，Meta一直资助“具身智能”的研究，希望开发能够观察、聆听并能够在周围3D物理环境中自主导航的AI助手。 （英文摘自2025.02.14 Reuters网站） 5、Reinforcement Learning（RL）强化学习 Large-scale reinforcement learning (RL) training of language models on reasoning tasks has become a promising technique for mastering complex problem-solving skills. 推理任务中语言模型的大规模强化学习（RL）训练已经成为一项解决复杂问题、具有前景广阔的技术。 （英文摘自2025.02.24 MarkTechPost网站） The partners said they will establish a shared reinforcement-learning training pipeline for Boston Dynamics’ Atlas robot to build dynamic and generalizable mobile manipulation behavior. 合作方表示，他们将为波士顿动力公司的Atlas机器人建立共享的强化学习训练流程，来赋予其动态性强、泛化能力高的移动操控能力。 （英文摘自2025.02.05 The Robot Report网站） 图片来源：Boston Dynamics 6、Neural Network 神经网络 The AI neural network will aid biomolecular research by predicting protein function, identifying novel molecules, assessing gene mutations and generating biological sequences. 该人工智能神经网络将通过预测蛋白质功能、识别新型分子、评估基因突变以及生成生物序列来推动生物分子研究。 （英文摘自2025.02.25 MobiHealthNews网站） According to DeepLearning.AI, neural networks have been pivotal in advancing AI from early brain-inspired models to modern transformers, impacting AI’s biggest breakthroughs. 据深度学习（DeepLearning.AI）指出，神经网络在推动人工智能从早期脑启发模型演进至现代Transformer架构的过程中起到关键作用，影响了人工智能若干重大突破的实现。 （英文摘自2025.02.04 Blockchain News网站） 图片来源：TechTarget 7、Group Relative Policy Optimization（GRPO）群体相对策略优化 Developed using an innovative technique called Group Relative Policy Optimization (GRPO) and a multi-stage training approach, DeepSeek R1 sets new benchmarks for AI models in mathematics, coding, and general reasoning. DeepSeek R1采用一种名为“群体相对策略优化”（GRPO）的创新技术以及多阶段训练策略，为人工智能模型在数学、编程和通用推理领域树立了新的性能标杆。 （英文摘自2025.01.28 Medium网站） Group Relative Policy Optimization (GRPO) is a variant of PPO designed to optimize large-scale models efficiently without requiring a critic model. 群体相对策略优化（GRPO）是近端策略优化（PPO）的变体算法，其设计目标旨在无需依赖评判模型的情况下，高效优化大规模模型。 （英文摘自2025.02.12 Kili Technology网站） 二 文化速递 1、The SMC Shanghai Foundation Model Innovation Center “模速空间” The SMC Shanghai Foundation Model Innovation Center is catalyzing the rapid advancement of the large artificial intelligence model industry fully leveraging Shanghai’s AI prowess. Officially unveiled in 2023, SMC stands as one of China’s first dedicated incubators and accelerators for large-scale AI models. It aims to fortify Shanghai's pursuit of establishing a world-class, globally competitive AI industry cluster. “模速空间”充分发挥上海在人工智能领域的优势，推动人工智能大模型产业的快速发展。该中心于2023年正式揭牌，是中国首个大模型人工智能专业孵化和加速载体。SMC致力于助力上海在全球竞争中构建世界一流人工智能产业集群的愿景。 2、AI for Science 科学智能 AI for Science, often abbreviated as AI4S, represents a transformative paradigm in scientific research and discovery. It leverages advanced artificial intelligence (AI) techniques to promote and transform traditional scientific methods. By integrating machine learning, deep learning, natural language processing, and other technologies, AI4S enables researchers to analyze vast amounts of data more efficiently, uncover hidden correlations, and formulate novel hypotheses. 科学智能（AI for Science，简称AI4S）是科学研究与发现领域的变革性范式。它应用先进的人工智能技术来促进和变革传统科学研究方法。通过整合机器学习、深度学习、自然语言处理等技术，AI4S使研究人员能够更高效地分析大量数据，发现隐藏的相关性，并提出新的假设。 撰稿：赵滢 核稿：文晶、钱嘉童
