{ "name": "必得 AI 精选", "updated_at": "2026-11-04", "publisher": "必得工坊", "items": [ { "id": "neural-networks", "title": "神经网络是什么？从一张手写数字看懂 AI", "short_title": "神经网络是什么", "topic": "basic", "category": "基础入门", "level": "入门", "bvid": "BV1bx411M7Zx", "youtube_id": "aircAruvnKk", "original_title": "But what is a Neural Network?", "bilibili_title": "【官方双语】深度学习之神经网络的结构 Part 1 ver 2.0", "source_url": "https://www.3blue1brown.com/lessons/neural-networks/", "bilibili_date": "2017-10-19", "lesson_date": "2017-10-05", "duration": null, "creator": "3Blue1Brown · Grant Sanderson", "summary": "从手写数字识别入手，先理解输入、权重和输出，再看复杂模型。", "takeaways": [ "像素亮度变成输入数字，网络逐层计算。", "权重和偏置共同影响神经元的激活值。", "整个网络是一个可调参数的函数，学习就是调整这些参数。" ], "why": "先建立一张看得懂的结构图，后面再遇到训练、大模型和注意力，就有可以放置新知识的位置。", "prerequisites": "不需要编程经验；知道加法和乘法即可开始。", "question": "网络识别错了一个数字，应该改输入、参数，还是输出标签？分别在什么情况下合理？", "exercise": "在纸上画一个“两输入、一输出”的神经元，自选权重和偏置，手算三组输入的结果，再改变一个权重比较差异。", "prompt": "请用两输入、一个输出的简化例子，带我手算加权和与偏置。每次先出题，等我回答再讲解。说明这个例子省略了哪些真实神经网络细节，最后检查我能否区分输入、参数和输出。", "tags": [ "神经网络", "深度学习", "权重", "入门" ], "theme": "network", "duration_seconds": 1153, "duration_iso": "PT19M13S", "duration_label": "19:13", "url": "https://bdskb.com/ai/watch/neural-networks/", "bilibili_url": "https://www.bilibili.com/video/BV1bx411M7Zx/", "notes_url": "https://bdskb.com/ai/watch/neural-networks/notes.txt", "notes_basis": "原作者配套图文；非视频完整转录", "practice_author": "必得工坊", "verified_at": "2026-11-04" }, { "id": "gradient-descent", "title": "梯度下降：神经网络到底怎样学习", "short_title": "模型怎样学习", "topic": "basic", "category": "基础入门", "level": "入门进阶", "bvid": "BV1Ux411j7ri", "youtube_id": "IHZwWFHWa-w", "original_title": "Gradient descent, how neural networks learn", "bilibili_title": "【官方双语】深度学习之梯度下降法 Part 2 ver 0.9 beta", "source_url": "https://www.3blue1brown.com/lessons/gradient-descent/", "bilibili_date": "2017-11-08", "lesson_date": "2017-10-16", "duration": null, "creator": "3Blue1Brown · Grant Sanderson", "summary": "用“沿山坡下行”的直觉，理解训练为什么要计算误差和调整参数。", "takeaways": [ "损失函数衡量预测与目标的差距。", "负梯度提供局部下降方向，学习率控制步长。", "训练集表现变好后，仍要用未见数据检查泛化。" ], "why": "理解训练是在怎样改参数，能帮助你区分“模型回答一次”和“重新训练模型”，也能看懂学习率、损失这些常见词。", "prerequisites": "建议先看神经网络入门；知道函数和斜率会更容易。", "question": "训练时的误差一直下降，就能断定模型在新数据上更好了吗？还缺哪一步检查？", "exercise": "对 f(x)=(x−3)²，从 x=0 出发，分别用学习率 0.1 和 1.1，按 x下一步=x−学习率×2(x−3) 更新五步，观察收敛与震荡。", "prompt": "请带我做 f(x)=(x-3)^2 的梯度下降练习。从 x=0 开始，分别使用学习率0.1和1.1，列出五次更新及函数值。先让我猜变化趋势，再核对计算，并解释这个一维例子与训练大型神经网络的差别。", "tags": [ "训练", "梯度下降", "损失函数", "学习率" ], "theme": "descent", "duration_seconds": 1261, "duration_iso": "PT21M1S", "duration_label": "21:01", "url": "https://bdskb.com/ai/watch/gradient-descent/", "bilibili_url": "https://www.bilibili.com/video/BV1Ux411j7ri/", "notes_url": "https://bdskb.com/ai/watch/gradient-descent/notes.txt", "notes_basis": "原作者配套图文；非视频完整转录", "practice_author": "必得工坊", "verified_at": "2026-11-04" }, { "id": "transformers", "title": "Transformer 是什么？看懂 GPT 生成文字的流程", "short_title": "看懂 Transformer", "topic": "llm", "category": "大模型原理", "level": "进阶", "bvid": "BV13z421U7cs", "youtube_id": "wjZofJX0v4M", "original_title": "Transformers, the tech behind LLMs", "bilibili_title": "【官方双语】GPT是什么？直观解释Transformer | 深度学习第5章", "source_url": "https://www.3blue1brown.com/lessons/gpt/", "bilibili_date": "2024-04-30", "lesson_date": "2024-04-01", "duration": null, "creator": "3Blue1Brown · Grant Sanderson", "summary": "先建立整条计算流程，再区分词元、向量、注意力与输出概率。", "takeaways": [ "文本切成 token 后映射为向量。", "注意力与前馈层反复变换表示，融入上下文。", "输出概率分布，经采样并重复计算生成后续文字。" ], "why": "把大模型拆成能逐步理解的环节，之后学习提示词、上下文和生成参数时，就能知道这些词大致作用在哪一层。", "prerequisites": "知道神经网络基本结构；向量与矩阵的直觉有帮助。", "question": "一次生成过程中，模型是在更新训练参数，还是在用已有参数计算下一个 token？", "exercise": "自己画一张“文本→token→向量→网络→概率→下一 token”流程图，每个箭头写一句作用；把“训练”和“生成”分别放在图外解释。", "prompt": "请检查我画的“文本→token→向量→网络→概率→下一token”流程。先让我解释每个箭头，再指出错误。用一小段中文举例，但不要把示意切分冒充某个真实分词器结果；如果使用真实分词器，要注明名称和实际输出。", "tags": [ "Transformer", "GPT", "token", "大模型" ], "theme": "transformer", "duration_seconds": 1634, "duration_iso": "PT27M14S", "duration_label": "27:14", "url": "https://bdskb.com/ai/watch/transformers/", "bilibili_url": "https://www.bilibili.com/video/BV13z421U7cs/", "notes_url": "https://bdskb.com/ai/watch/transformers/notes.txt", "notes_basis": "原作者配套图文；非视频完整转录", "practice_author": "必得工坊", "verified_at": "2026-11-04" }, { "id": "attention", "title": "注意力机制：AI 怎样根据上下文理解一个词", "short_title": "注意力怎样工作", "topic": "llm", "category": "大模型原理", "level": "进阶", "bvid": "BV1TZ421j7Ke", "youtube_id": "eMlx5fFNoYc", "original_title": "Attention in transformers, step-by-step", "bilibili_title": "【官方双语】直观解释注意力机制，Transformer的核心 | 【深度学习第6章】", "source_url": "https://www.3blue1brown.com/lessons/attention/", "bilibili_date": null, "lesson_date": "2024-04-07", "duration": null, "creator": "3Blue1Brown · Grant Sanderson", "summary": "看清 Q、K、V 各做什么，理解词元如何从其他位置读取信息。", "takeaways": [ "同一词元的含义需要由上下文补充。", "查询与键计算相关程度，经归一化得到权重。", "值向量按权重汇总；多头可学习不同关联。" ], "why": "把“关注哪些上下文”变成一个可解释的计算过程，适合接着 Transformer 总览往下看。", "prerequisites": "建议先看 Transformer 总览；理解向量和加权平均。", "question": "示意图中的连线为什么不能直接当成模型给出某个答案的完整解释？", "exercise": "写两句含“苹果”的话，一句描述水果、一句描述公司；圈出帮助区分含义的上下文，再手绘一张示意权重表，并注明它不是模型真实注意力。", "prompt": "请用“苹果”表示水果和公司两个例子，带我理解上下文、query、key、value。用很小的手工数字演示一次加权汇总，核对权重总和；必须说明数字是教学示意，不是对某个真实模型注意力的测量。", "tags": [ "注意力机制", "QKV", "Transformer", "上下文" ], "theme": "attention", "duration_seconds": 1570, "duration_iso": "PT26M10S", "duration_label": "26:10", "url": "https://bdskb.com/ai/watch/attention/", "bilibili_url": "https://www.bilibili.com/video/BV1TZ421j7Ke/", "notes_url": "https://bdskb.com/ai/watch/attention/notes.txt", "notes_basis": "原作者配套图文；非视频完整转录", "practice_author": "必得工坊", "verified_at": "2026-11-04" } ] }