<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom"><channel><title>DeepSeek 相关文章 on Peng Tan's AI Blog</title><link>https://77f3dc70.hobbytp-github-io.pages.dev/zh/deepseek/</link><description>一个关注 AI 各领域的专题博客</description><atom:link href="https://77f3dc70.hobbytp-github-io.pages.dev/zh/deepseek/index.xml" rel="self" type="application/rss+xml"/><item><title>mHC: Redefining Deep Learning Scalability - DeekSeek</title><link>https://77f3dc70.hobbytp-github-io.pages.dev/zh/deepseek/deepseek_mhc/</link><pubDate>Thu, 01 Jan 2026 23:20:00 +0800</pubDate><guid>https://77f3dc70.hobbytp-github-io.pages.dev/zh/deepseek/deepseek_mhc/</guid><description>2026年伊始，DeepSeek团队带来了mHC（流形约束超连接）,这项技术完美解决了传统残差连接提升性能的超连接技术因破坏稳定性而难以实用的阿喀琉斯之踵.</description></item><item><title>DeepSeekMath-V2 技术白皮书：迈向自验证数学推理</title><link>https://77f3dc70.hobbytp-github-io.pages.dev/zh/deepseek/deepseekmath_v2/</link><pubDate>Sat, 29 Nov 2025 12:10:00 +0800</pubDate><guid>https://77f3dc70.hobbytp-github-io.pages.dev/zh/deepseek/deepseekmath_v2/</guid><description>本文介绍了DeepSeekMath-V2，旨在克服传统上依赖最终答案的数学推理奖励机制的局限性。其核心创新在于通过训练一个专门的证明验证器来评估推导的逻辑严谨性，从而实现自我可验证的数学推理能力。该系统引入了元验证过程，以确保验证器识别的错误是准确的，并利用此验证器作为奖励模型来训练生成器进行自我评估和迭代优化。这种生成与验证的协同作用创建了一个循环，通过扩展计算规模来自动标注难以验证的新证明，从而提高了系统的能力。</description></item><item><title>DeepSeek-OCR：重塑AI长文本处理</title><link>https://77f3dc70.hobbytp-github-io.pages.dev/zh/deepseek/deepseek_ocr/</link><pubDate>Tue, 21 Oct 2025 20:10:00 +0800</pubDate><guid>https://77f3dc70.hobbytp-github-io.pages.dev/zh/deepseek/deepseek_ocr/</guid><description>本文介绍了DeepSeek-OCR，一种革命性的AI模型，能够将长文本处理效率提升数十倍，从而实现对超长文档的快速处理。</description></item><item><title>DeepSeek FlashMLA 代码解读</title><link>https://77f3dc70.hobbytp-github-io.pages.dev/zh/deepseek/deepseek_flashmla/</link><pubDate>Mon, 24 Feb 2025 16:10:00 +0800</pubDate><guid>https://77f3dc70.hobbytp-github-io.pages.dev/zh/deepseek/deepseek_flashmla/</guid><description>本文介绍了深度求索（DeepSeek）公司FlashMLA代码详细解读。</description></item><item><title>DeepSeek 开源 LLM 对闭源 LLM 的影响</title><link>https://77f3dc70.hobbytp-github-io.pages.dev/zh/deepseek/deepseek_impact/</link><pubDate>Tue, 18 Feb 2025 23:20:00 +0800</pubDate><guid>https://77f3dc70.hobbytp-github-io.pages.dev/zh/deepseek/deepseek_impact/</guid><description>本文介绍了DeepSeek开源LLM对闭源LLM的影响，包括性能基准测试和竞争、成本效益、开源可用性和定制、市场动态和战略转变、创新与社区发展、环境影响以及AI研究和应用的转变。</description></item><item><title>DeepSeek V3 论文解读</title><link>https://77f3dc70.hobbytp-github-io.pages.dev/zh/deepseek/deepseek_v3/</link><pubDate>Fri, 14 Feb 2025 18:10:00 +0800</pubDate><guid>https://77f3dc70.hobbytp-github-io.pages.dev/zh/deepseek/deepseek_v3/</guid><description>本文介绍了深度求索（DeepSeek）公司推出的新一代推理模型DeepSeek-V3，并对其技术原理、主要贡献、论文方法、评估结果和局限性进行了详细解读。</description></item><item><title>DeepSeek 微调</title><link>https://77f3dc70.hobbytp-github-io.pages.dev/zh/deepseek/deepseek-finetuning/</link><pubDate>Fri, 14 Feb 2025 18:10:00 +0800</pubDate><guid>https://77f3dc70.hobbytp-github-io.pages.dev/zh/deepseek/deepseek-finetuning/</guid><description>本文介绍了如何使用合成推理数据集微调DeepSeek-R1模型.</description></item><item><title>DeepSeek R1 论文解读</title><link>https://77f3dc70.hobbytp-github-io.pages.dev/zh/deepseek/deepseek_r1/</link><pubDate>Mon, 10 Feb 2025 20:10:00 +0800</pubDate><guid>https://77f3dc70.hobbytp-github-io.pages.dev/zh/deepseek/deepseek_r1/</guid><description>本文介绍了深度求索（DeepSeek）公司推出的新一代推理模型DeepSeek-R1，并对其技术原理、主要贡献、论文方法、评估结果和局限性进行了详细解读。</description></item><item><title/><link>https://77f3dc70.hobbytp-github-io.pages.dev/zh/deepseek/deepseek_company_strategy/</link><pubDate>Mon, 01 Jan 0001 00:00:00 +0000</pubDate><guid>https://77f3dc70.hobbytp-github-io.pages.dev/zh/deepseek/deepseek_company_strategy/</guid><description>&lt;p>&lt;a href="https://cloud.baidu.com/article/3681511">2025年DeepSeek技术突破与产业生态重构研究&lt;/a>&lt;/p>
&lt;p>&lt;a href="https://developer.baidu.com/article/detail.html?id=3716661">DeepSeek狂潮席卷全球：国家队入场，开启全民AI普惠时代&lt;/a>&lt;/p></description></item><item><title/><link>https://77f3dc70.hobbytp-github-io.pages.dev/zh/deepseek/deepseek_usage_tips/</link><pubDate>Mon, 01 Jan 0001 00:00:00 +0000</pubDate><guid>https://77f3dc70.hobbytp-github-io.pages.dev/zh/deepseek/deepseek_usage_tips/</guid><description>&lt;h2 id="提示词">提示词&lt;/h2>
&lt;p>不同类别的提示词如下：&lt;/p>
&lt;ol>
&lt;li>&lt;strong>个性化写作风格（Personalize your writing style）&lt;/strong>：
&lt;ul>
&lt;li>Analyze the writing style of the text provided below: [text] Now write an essay of [number of words] words on [specific topic]. You do not have to mention anything related to the previous text, it is simply provided to elicit a response that imitates the tone, structure, and vocabulary. Answer me only with the requested essay.&lt;/li>
&lt;li>分析以下文本的写作风格（示例：学术论文/营销文案/新闻报道）：[text] 现在写一篇关于[特定主题]的[字数]字的文章。你不需要提到与前面文本的任何关系，它只是为了引发一个模仿前面文本的写作风格、结构和词汇的回答。只回答我请求的文章。&lt;/li>
&lt;/ul>
&lt;/li>
&lt;li>&lt;strong>通过获取反馈提升写作（Improve your writing by getting feedback）&lt;/strong>：
&lt;ul>
&lt;li>[paste your writing] Please proofread the text above. Correct grammatical and spelling mistakes and offer ideas that will make my writing more lucid.&lt;/li>
&lt;li>请校对上面的文本。纠正语法和拼写错误，并提出使我的写作更清晰的想法。&lt;/li>
&lt;/ul>
&lt;/li>
&lt;li>&lt;strong>生成新想法（Generate new ideas）&lt;/strong>：
&lt;ul>
&lt;li>I need fresh ideas for [topic]. Describe at least 5 innovative and practical ideas that can be implemented. Briefly explain the potential of each idea and how it could be developed to maximize its impact and viability. Don&amp;rsquo;t be too - long - winded.&lt;/li>
&lt;li>我需要关于[主题]的新想法。描述至少5个可以实施的创新和实用的想法。简要解释每个想法的潜力以及如何开发它以最大化其影响和可行性。不要过于冗长。&lt;/li>
&lt;/ul>
&lt;/li>
&lt;li>&lt;strong>训练AI为你生成提示词（Train AI to generate prompts for you）&lt;/strong>：
&lt;ul>
&lt;li>You are an AI created to assist [insert occupation]. For youself, make a list of the top 10 prompts. The topic of the prompts should be [insert topic].&lt;/li>
&lt;li>你是一个AI，被创建来帮助[插入职业]。为自己，列出10个最受欢迎的提示词。提示词的主题应该是[插入主题]。&lt;/li>
&lt;/ul>
&lt;/li>
&lt;li>&lt;strong>将AI变为你的实习生（Turn AI into your intern）&lt;/strong>：
&lt;ul>
&lt;li>I&amp;rsquo;m writing a report on [insert subject]. Conduct thorough research and write a comprehensive report that includes a step - by - step guide to help readers [insert outcome].&lt;/li>
&lt;li>我正在写一篇关于[插入主题]的报告。进行彻底的研究，并写一篇全面的报告，包括一个逐步的指南，帮助读者[插入结果]。&lt;/li>
&lt;/ul>
&lt;/li>
&lt;li>&lt;strong>使用80/20原则加快学习（Use the 80/20 principle to learn faster）&lt;/strong>：
&lt;ul>
&lt;li>I want to learn about [insert topic]. Determine and share the 20% of the topic&amp;rsquo;s lessons that are most crucial to understanding the remaining 80%.&lt;/li>
&lt;li>我想学习关于[插入主题]。确定并分享这个主题的20%的最重要的课程，以理解剩下的80%。&lt;/li>
&lt;/ul>
&lt;/li>
&lt;li>&lt;strong>学习新技能（Learn any new skill）&lt;/strong>：
&lt;ul>
&lt;li>[Insert desired skill] is something I want to learn. Make a 30 - day study schedule that will assist someone like me who is just starting to learn this skill in developing it.&lt;/li>
&lt;li>[插入想要学习的技能]是我想要学习的。为我制作一个30天的学习计划，帮助像我这样的人开始学习这个技能。&lt;/li>
&lt;/ul>
&lt;/li>
&lt;li>&lt;strong>总结复杂文本（Summarise complex texts）&lt;/strong>：
&lt;ul>
&lt;li>I need to read a complicated article related to [topic]. Can you help me summarise the key points and takeaways from the text?&lt;/li>
&lt;li>我需要阅读一篇关于[主题]的复杂文章。你能帮我总结文章的关键点和要点吗？&lt;/li>
&lt;/ul>
&lt;/li>
&lt;li>&lt;strong>撰写发布演讲稿（Write a Launch Speech）&lt;/strong>：
&lt;ul>
&lt;li>Write a launch speech for [product/business] that highlights the values of the [company or niche], and addresses a widespread problem or mistake. Product = [Insert yours]&lt;/li>
&lt;li>为[产品/业务]撰写一个发布演讲稿，强调[公司或利基]的价值，并解决一个普遍存在的问题或错误。产品= [插入你的产品]&lt;/li>
&lt;/ul>
&lt;/li>
&lt;li>&lt;strong>提升解决问题的能力（Improve your problem - solving abilities）&lt;/strong>：
&lt;ul>
&lt;li>Your job is to solve problems. Give me a step - by - step guide to solve this problem: [insert].&lt;/li>
&lt;li>你的工作是解决问题。给我一个解决这个问题的逐步指南：[插入]。&lt;/li>
&lt;/ul>
&lt;/li>
&lt;li>&lt;strong>生成PPT（Generate a PPT）&lt;/strong>：
&lt;ul>
&lt;li>I need to create a presentation on [insert topic]. Can you help me generate a list of slides and their content?&lt;/li>
&lt;li>我需要创建一个关于[插入主题]的演示文稿。你能帮我生成一个幻灯片列表及其内容吗？&lt;/li>
&lt;/ul>
&lt;/li>
&lt;li>&lt;strong>制作视频脚本（Make a video script）&lt;/strong>：
&lt;ul>
&lt;li>I&amp;rsquo;m making a video on [insert topic]. Can you help me create a script that will keep the audience engaged and interested?&lt;/li>
&lt;li>我正在制作一个关于[插入主题]的视频。你能帮我创建一个脚本，让观众保持兴趣和参与吗？&lt;/li>
&lt;/ul>
&lt;/li>
&lt;li>&lt;strong>生成产品描述（Generate a product description）&lt;/strong>：
&lt;ul>
&lt;li>I&amp;rsquo;m launching a new product. Can you help me create a compelling description that will attract customers?&lt;/li>
&lt;li>我正在推出一个新产品。你能帮我创建一个吸引顾客的引人入胜的描述吗？&lt;/li>
&lt;/ul>
&lt;/li>
&lt;li>&lt;strong>跨模态内容生成&lt;/strong>：&lt;/li>
&lt;/ol>
&lt;ul>
&lt;li>&amp;ldquo;我正在设计&amp;lt;产品原型&amp;gt;，请根据&amp;lt;目标用户画像&amp;gt;生成配套的UI界面描述和营销话术&amp;rdquo;&lt;/li>
&lt;li>&amp;ldquo;基于&amp;lt;技术白皮书&amp;gt;内容，自动生成配套的信息图设计说明&amp;rdquo;&lt;/li>
&lt;/ul>
&lt;ol start="15">
&lt;li>&lt;strong>伦理安全审查&lt;/strong>：&lt;/li>
&lt;/ol>
&lt;ul>
&lt;li>&amp;ldquo;请对以下内容进行合规性审查：[内容]，依据&amp;lt;行业规范&amp;gt;和&amp;lt;地区法律法规&amp;gt;&amp;rdquo;&lt;/li>
&lt;li>&amp;ldquo;识别文本中潜在的偏见/歧视表述：[text]，按&amp;lt;联合国AI伦理准则&amp;gt;分析&amp;rdquo;&lt;/li>
&lt;/ul>
&lt;h2 id="与其他工具集成">与其他工具集成&lt;/h2>
&lt;h3 id="deep-research">Deep Research&lt;/h3>
&lt;h3 id="ppt制作">PPT制作&lt;/h3>
&lt;h3 id="视频制作">视频制作&lt;/h3>
&lt;h3 id="图片生成">图片生成&lt;/h3>
&lt;h3 id="音频生成">音频生成&lt;/h3>
&lt;h3 id="代码生成">代码生成&lt;/h3></description></item><item><title/><link>https://77f3dc70.hobbytp-github-io.pages.dev/zh/deepseek/deepseek_v3.1/</link><pubDate>Mon, 01 Jan 0001 00:00:00 +0000</pubDate><guid>https://77f3dc70.hobbytp-github-io.pages.dev/zh/deepseek/deepseek_v3.1/</guid><description/></item><item><title/><link>https://77f3dc70.hobbytp-github-io.pages.dev/zh/deepseek/deepseek_v3.2/</link><pubDate>Mon, 01 Jan 0001 00:00:00 +0000</pubDate><guid>https://77f3dc70.hobbytp-github-io.pages.dev/zh/deepseek/deepseek_v3.2/</guid><description>&lt;h2 id="参考文献">参考文献&lt;/h2>
&lt;ul>
&lt;li>&lt;a href="https://github.com/deepseek-ai/DeepSeek-V3.2-Exp/blob/main/DeepSeek_V3_2.pdf">论文&lt;/a>&lt;/li>
&lt;li>&lt;a href="https://huggingface.co/deepseek-ai/DeepSeek-V3.2-Exp">HuggingFace&lt;/a>&lt;/li>
&lt;li>&lt;a href="https://modelscope.cn/models/deepseek-ai/DeepSeek-V3.2-Exp">ModelScope&lt;/a>&lt;/li>
&lt;/ul></description></item></channel></rss>