浙江大学DS系列专题 DeepSeek技术溯源及前沿探索 主讲人：朱强 浙江大学计算机科学与技术学院 人工智能省部共建协同创新中心（浙江大学） https://person.zju.edu.cn/zhuq 1 Outline 一、语言模型 三、ChatGPT 二、Transformer 四、DeepSeek 五、新一代智能体 2 语言模型：终极目标 Language Modeling 对于任意的词序列，计算出这个序列是一句话的概率 我们每天都和语言模型打交道： I saw a cat I saw a cat on the chair I saw a cat running after a dog I saw a ca car I saw a cat in my dream 3 语言模型：基本任务 编码：让计算机理解人类语言 She is my mom 1 0 0 0 0 1 0 0 0 0 1 0 0 0 0 1 只有一个1，其余均为0 One-hot Encoding有什么缺点吗？ One-hot Encoding 4 编码：让计算机理解人类语言 Word Embedding A bottle of tez ...

？ 大模型：从单词接龙到行业落地 杨洋，浙江大学 数推公析及office 教学用途声明：本PPT包含部分来源于网络的素材，仅供教学使用，非商业用途，版权归原作者所有 人工智能前夜：图灵测试 AlanMathisonTuring(1912-1954) 公众号·史老师数据分析及office学习 围数学家，计机4学活 理论计算机科学之父 2 图灵测试会在测试人在与被测试者（一个人和一台机器）隔开的情况下，通过一 些装置（如键盘）向被测试者随意提问 门 问过一些问题后，如果超过30%的答复不能使测试人认出哪个是人、哪个是机 器的回答，那么这台机器就通过了测试，并被认为具有人类智能。 人工智能前夜：图灵测试 MIND 在提出图灵测试的《计算机器与智能》一文里，图灵描述了想象中 未来的智能计算机测试可能的样子（人提出问题，计算机回答）： | Q：请给我写一首有关福思桥（ForthBridge）主题的十四行诗。 | | --- | | 文学 | | A：这种事情别找我。我从来都不会写诗 | Alan TuringBanknoteConcept IMITATON GAME | 数 | | | --- | --- | | 学 ...

表：主要数据集大小汇总，以GB为单位。公开的数据集以粗体表示， 确定的数据以斜体表示。Common Crawl数据集过滤之前为45T DeepSeek 智能时代的全面到来和人机协作的新常态 孙凌云 教 授 计算机科学与技术学院 2025年2月 一、智能演变 二、人机协作 三、产业现状 四、教育成长 到 2020 年的 GPT-3， 模型预训练数据量从 4.6GB 增加到了 45TB 45TB 相当于三千万本《西游记》 主要模型数据集包括： 来源： OpenAI团队，Language Models are Few-Shot ，2022年7月22日 | | 维基 百科 | 书籍 | 杂志 期刊 | Reddit 链接 | Common Crawl | 其他 数据 | 总计 | | --- | --- | --- | --- | --- | --- | --- | --- | | GPT-1 | | 4.6 | | | | | 4.6 | | GPT-2 | | | | 40 | | | 40 | | GPT-3 | 11.4 | 21 | 101 | 50 | 570 | | 753 | | The Pile v ...

Industry Overview - The report focuses on the development of intelligent analysis and control strategies for new power systems driven by a combination of mechanism and artificial intelligence [1] - The core objective is to achieve "carbon peak and carbon neutrality" by building a new power system characterized by high proportions of renewable energy integration, power electronics, and energy storage devices [4] - Key challenges include the dynamic, random, and uncertain nature of grid operations, which require advanced online safety assessment and intelligent scheduling control methods [4] Key Research Areas - Multi-temporal and spatial dimension power prediction technology has been developed using integrated machine learning models, applied in various scenarios such as residential, industrial, and commercial loads, as well as renewable energy generation and charging stations [2] - Intelligent decision-making technologies based on deep reinforcement learning have been developed to address control challenges in power system planning and scheduling, including optimization of power equipment configuration and control of reactive voltage, active power, network loss, and topology [2] - Digital twin modeling and parameter intelligent identification technologies have been established for complex power equipment, including traditional generators, DC systems, wind power, and composite load models [2] Applications and Case Studies - A data-driven, AI-based grid brain technology framework has been applied in Jiangsu Power Grid, addressing issues such as voltage violations, power loss reduction, and power flow constraints [14] - The intelligent control system deployed in Jiangsu Power Grid achieved a 99.41% effective instruction rate, with an average network loss reduction of approximately 3.6412% [16] - Reinforcement learning-based methods have been used for automatic parameter calibration of generator and excitation models, achieving good performance after 400 training iterations [20] Advanced Technologies - Deep reinforcement learning has been applied to real-time AC optimal power flow, deriving fast solutions for secure and economic grid operations [22] - AI-based autonomous topology control has been developed to maximize time-series available transfer capabilities, considering uncertainties in grid operations [37] - A comprehensive planning, early warning, and control platform integrating AI, HPC, and big data technologies has been established for multi-objective power system simulation, verification, and control [41] Distributed Resource Management - Virtual power plants (VPPs) have been developed to aggregate distributed renewable generation, energy storage, and load subsystems, providing capacity and ancillary services to improve grid economy and reliability [25] - VPPs enable intelligent adjustment of various load devices, addressing challenges related to the randomness and volatility of new energy integration [25] - Reinforcement learning-based net load volatility control has been applied in active distribution power networks, effectively reducing peak-valley differences and overall fluctuations [32]