跳到正文
小杜戴着墨镜、穿着左胸印有小狮子标志的黑色连帽卫衣,双臂交叉,看向镜头

小杜算法攻城狮/ algorithm engineer

我写算法,也写把算法送上线的代码。每天和数据结构、模型、线上系统打交道——这里记下踩过的坑、读过的论文、做过的项目。

最近在写LLM 预训练数据处理流水线

01 / 最近在写

新鲜出炉

AI 资讯、算法拆解和上手教程。新写的在最上面。

全部 276 篇文章
  1. 2026/08/29AI算法与模型约 6 分钟

    LLM 预训练数据处理流水线

    大模型训练前,原始数据要经过一整套处理才能"喂"给模型:从网页抓取、正文提取、语言识别,到质量过滤、去重、内容安全审查、版权核对,再到数据配比、分词器训练和序列打包,每一步都会影响最终模型的能力和行为。本文系统梳理这十个关键环节,讲清楚每一步在做什么、为什么必须做、工业界都有哪些常见方法,帮你理解一个大模型真正是"吃"什么样的数据长大的。
  2. 2026/04/26AI教程与指南约 1 分钟

    99%使用"豆包"的人会忽略掉的安全设置:

    豆包隐私安全设置教程:按照步骤进入“设置 → 隐私与权限 → 帮助模型改进效果”,将相关选项全部关闭,避免个人数据被用于模型训练,保护你的隐私安全。
  3. 2026/04/23AI资讯约 8 分钟

    AI 要释放商业价值,需要强有力的数据织网

    文章认为,企业 AI 的主要瓶颈已不再是模型性能或算力,而是数据质量与业务语境。只有把分散在应用、云和业务流程中的语义、规则与知识连接起来,AI 才更可能做出可信且有业务价值的决策。
  4. 2026/04/23AI资讯约 4 分钟

    特朗普“从处决中救下”的伊朗女性,既真实存在,也经过了 AI 处理

    特朗普称自己促成8名因抗议政权而被判处死刑的伊朗女性获释,但相关说法很快遭伊朗官媒否认。争议中的拼贴图至少经过 AI 修改,不过其中至少6名女性被证实真实存在。
  5. 2026/04/23AI资讯约 4 分钟

    特斯拉营收再次增长,为更多 AI 与机器人布局做准备

    特斯拉公布 2026 年第一季度财报:营收 224 亿美元、净利润 4.77 亿美元,分别同比增长 16% 和 17%,但仍低于华尔街预期。公司同时推进 Optimus 工厂与 Dojo 3 项目。
  6. 2026/04/23AI资讯约 2 分钟

    Anthropic 调查 Mythos 疑似被未授权接触事件

    Anthropic 表示,正在调查一份关于未授权用户通过第三方供应商环境接触 Mythos 的报告。尽管涉事人员据称尚未运行网络攻击提示,但事件再次暴露了高风险模型在封闭测试阶段的治理难题。
小杜戴着透明框眼镜,在绿色代码雨前对着笔记本电脑写代码

02 / AI 课程

按这条路线学 AI

从数学直觉出发,到算法选型、神经网络,再到落地和变现。5 门课,一条线。

5门课30节

全部课程
STEP 01 · 会选算法推荐VIP6 节课VIP

机器学习算法地图:从 0 到能选型、能解释、能上线

让你面对一个业务问题,能做“问题建模 → 算法选择 → 评估指标 → 上线策略”,而不是只会背一堆名字。课程大纲
  1. 分类/回归/聚类/异常检测怎么选才不翻车
  2. 逻辑回归、SVM、正则化是工业界老牌神器
  3. 决策树、随机森林、XGBoost/LightGBM 为啥这么能打
  4. 无监督与表示学习:聚类、降维、向量化——让数据“可用”
从这里开始共 6 节
  1. STEP 02 · 变现

    AI 搞钱项目分析-变现路线图

    不讲“发财学”,讲可执行的商业模型:需求、成本、交付、渠道、复购、风险。你能用一套框架快速判断一个项目值不值得做。6 节课VIP
  2. STEP 03 · 落地成产品

    AI 落地三件套:绘画 / 视频 / 智能体|从玩具到生产力

    你不仅能“做出来”,还能“做成流程”,能复用、能规模化、能交付给别人用。6 节课VIP
  3. STEP 04 · 懂神经网络

    神经网络核心内功:从感知机到训练稳定的深度模型

    把“网络为什么能学到东西”讲透,并把训练中最常见的坑(不收敛、过拟合、爆炸)变成你能控制的变量。6 节课VIP
  4. STEP 05 · 起点

    AI 数学:把高数、线代、概率变成你能用的算法

    不是“学会公式”,而是把数学变成你写模型、调参、看论文时的直觉武器。学完你能解释:为什么梯度这样算、矩阵为什么这样乘、概率为什么这样更新。6 节课VIP

不只是解锁内容,更是一场认知与效率的跃迁。会员计划正在打磨,即将开放。

加入会员

03 / Prompt

Prompt 咒语

AI Prompt|产品超详细分解视图

Ultra-detailed exploded view of a product, metallic parts and electronic components floating in mid-air, perfectly aligned, revealing inner structure, futuristic technology aesthetic, 8K resolution, soft cinematic lighting, highly realistic.
详情

东方韵味:白发女性的铂金优雅

A hyper-realistic photograph of a stunning East Asian woman with delicate, refined features and long, silky white hair featuring luminous platinum highlights that cascade elegantly over her shoulders. She has large expressive almond-shaped eyes with natural double eyelids, high cheekbones, a small refined nose, and soft full lips, paired with flawless porcelain-smooth fair skin. Captured in a fashion editorial style against a seamless, pure amber-orange backdrop that creates a bold, minimalist color field. She wears a stunning sapphire-blue cutout dress in rich, saturated tones, featuring strategic geometric cutouts along the waist, sides, and décolletage that teasingly reveal smooth skin while the luxurious fabric drapes and clings with perfect tailoring and subtle sheen. A breathtaking amber and silver headpiece crowns her head — delicate metallic filigree intertwined with glowing amber stones and silver chains that catch the light dramatically. She is adorned with matching luxurious accessories: slender silver earrings with amber drops, a fine silver choker, and elegant bracelets that enhance the high-fashion aesthetic. Her posture is exquisitely elegant — spine straight, shoulders relaxed, head slightly tilted with a confident, serene expression and direct gaze. Soft, cinematic studio lighting sculpts every contour with ultra-realistic highlights and gentle shadows, emphasizing her flawless skin texture, fabric drape, and metallic reflections. Photorealistic, ultra-detailed skin texture with natural pores and subtle luminous sheen, lifelike hair strands with individual platinum highlights, intricate dress cutouts and fabric realism, precise headpiece and jewelry details with accurate gemstone refraction, sharp focus on her face and figure, high dynamic range editorial lighting, 8k resolution, highly detailed, no artifacts
详情

复古时尚:玫瑰色调的优雅

A hyper-realistic, low-angle full body shot of a woman with platinum blonde shoulder-length hair, kneeling on a pink carpet and looking down at the camera. She is wearing a monochromatic dusty-rose outfit consisting of a structured satin corset with lace detailing, a draped silk skirt, and sheer latex opera gloves. She is wearing layered delicate silver body chains. The background is a retro wall with vintage beige floral damask wallpaper. The lighting is soft and cinematic, editorial fashion photography style, high detail, 8k resolution.
详情

复古游乐园的黄金时段优雅

{ "prompt": "A fashionable young woman with curly red hair styled in a loose messy updo relaxing inside a metal shopping cart in an empty amusement park parking lot at golden hour. She wears a black-and-white gingham strapless crop top, denim shorts, and bold red platform high-heeled sandals. Her pose is relaxed and confident, with one arm resting on the cart and legs extended outward. Warm sunset light casts long shadows and dramatic highlights on her skin. In the background are colorful fairground structures, a small green ticket booth, tents, and blurred amusement rides. The mood is editorial, cinematic, and slightly nostalgic, with vibrant yet natural color grading and shallow depth of field.", "style": "cinematic fashion editorial photography", "lighting": "golden hour sunlight, warm tones, long shadows", "camera": { "angle": "low three-quarter perspective", "lens": "50mm prime", "depth_of_field": "shallow" }, "quality": "ultra-detailed, photorealistic, cinematic color grading", "aspect_ratio": "2:3" }
详情

暮光下的浪漫梦境

Beautiful woman, coy smile, dewy peach-blush skin, pink bohemian headscarf, flowing blush-pink gown. Twilight with floating lanterns, glowing orbs, soft golden-pink light, bokeh, mist. Dreamy romantic fantasy, cinematic soft lighting, subtle motion.
详情

手持岛屿上的微缩莫斯科

A hyper-realistic close-up shot at a 45-degree angle of a hand holding a miniature 3D diorama on a floating island representing Moscow, Russia. The base of the island is made of layered historic white stone and soil, with detailed textures and an elegant golden inscription “MOSCOW” on the front side. At the top of the island, the main landmarks are depicted in extremely fine detail: the magnificent St. Basil’s Cathedral with its colorful twisted domes, the strict towers of the Moscow Kremlin, and the modern mirrored skyscrapers of Moscow City. Between them, miniature sightseeing riverboats can be seen on the curve of the Moskva River, along with the lush greenery of Zaryadye Park and its floating bridge. Crystal-clear water from the Moskva River surrounds the island, showing soft reflections of city lights and tiny river trams. Decorative lilac bushes and red carnations grow along the edges of the island. In the background, faint silhouettes of Red Square or the embankment at dusk are softly blurred due to cinematic depth of field. The lighting emphasizes the blue hour, with warm city streetlights, ultra-detailed architecture, and photorealistic textures. Technical parameters: Shot with an 85mm lens, f/1.8 aperture, shallow depth of field, HDR, global illumination, 16K resolution, hyper-realistic cinematic photography, realistic skin texture on the hand, masterpiece quality.
详情

化装舞会时尚女性肖像

一张竖直构图的超写实人物肖像照片,呈现出原始时尚摄影和狗仔队闪光灯美学风格。画面主体是一位二十岁出头、极具吸引力且极其美丽的年轻女性。她非常靠近镜头站立,身体从腰部开始微微前倾,双肩自然向后微展,胸部自然挺起,下巴微收。她的面容对称且精致,拥有深邃的大眼睛,眼神直视镜头,展现出强烈的眼神交流与亲密感。眉毛自然放松,嘴唇柔软、带有光泽且微微张开,带着自然自信的微笑。她的浓密深棕色长发略带蓬松与微乱的质感,自然垂落在双肩前方,完美修饰着脸庞。 她的右手举在脸颊高度,手中清晰地握着一个华丽的银黑相间威尼斯化装舞会面具,面具向外倾斜,完全展现出其表面精细复杂的金属丝镂空雕花工艺。她的左手轻轻搭在腰臀处以突显身体曲线。女子身穿一件银色亮片吊带连衣裙,细肩带搭配深邃优雅的领口,自然地展现出锁骨与乳沟轮廓。裙子的亮片在强烈的闪光灯照射下闪烁着耀眼的光芒。她的脖子上佩戴着一条极简风格的钻石项链,刚好位于领口上方。 环境设定在夜晚的豪华屋顶或阳台,背景非常暗且呈现浅景深效果,仅有柔和模糊的城市夜景灯光光斑,没有任何分散注意力的元素。画面采用直接的正面闪光灯照明,光源位于视线稍上方,在女子的眼睛、颧骨、锁骨、领口肌肤以及面具上勾勒出锐利的高光和干净利落的阴影。画面的焦点极为锐利,精准对焦在女子的眼睛和面具上。皮肤质感极度逼真,毛孔清晰可见,亮片的反射光泽、面具的繁复细节以及每一根发丝都展现出极高的清晰度与电影般的写实感。
详情

细节尽显优雅:奢华浴室休憩

{ "task": "image_generation", "output": { "aspect_ratio": "1:1", "resolution": "8K", "style": [ "ultra-realistic", "photorealistic", "luxury editorial" ] }, "identity_reference": { "use_uploaded_image_as_reference": true, "face_preservation": "100%", "body_preservation": "100%", "no_facial_modification": true, "no_body_shape_change": true }, "scene": { "location": "luxury modern bathroom", "environment": [ "white marble vanity", "gold fixtures", "glass shower", "large backlit mirrors" ], "lighting": "soft warm indoor lighting, natural skin highlights, realistic shadows" }, "subject": { "gender": "female", "pose": "sitting or leaning on marble counter, elegant relaxed posture", "expression": "confident, calm, subtle smile", "hair": { "style": "sleek low bun", "part": "center", "texture": "smooth and polished" } }, "outfit": { "dress": { "type": "black bodycon mini dress", "neckline": "deep V-neck", "fit": "form-fitting, elegant" }, "footwear": "black high heels" }, "camera": { "type": "DSLR", "lens": "50mm", "aperture": "f/1.8", "angle": "eye-level", "focus": "sharp subject, soft background depth" }, "quality_notes": { "skin_texture": "natural, detailed, realistic", "fabric_detail": "high-quality dress texture", "no_ai_artifacts": true, "no_over_smoothing": true } }
详情
小杜的标志:戴墨镜的小狮子,额头上有 </>

04 / 关于我

算法攻城狮 · 小杜

“攻城狮”是“工程师”的谐音。这只戴墨镜的小狮子是我的标志,额头上的 </> 是它天天写代码的证明。有问题、有想法,或者只是想聊聊,欢迎来找我。