用 Workbuddy 啃硬核论文的一些心里话

{"type":"doc","content":[{"type":"paragraph","attrs":{"id":"464167d6-6192-452b-87b4-78455b6fb293","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":"这次拿 DL-SLAM 这篇新文试了 Workbuddy 的深度解析能力，体验下来最大的感受是："},{"type":"text","marks":[{"type":"bold"}],"text":"它不是在“帮我读论文”，而是在“陪我拆论文”"},{"type":"text","text":"。以前用其他 AI 读论文，大多是“扩写摘要 翻译公式”，但 Workbuddy 这次给我的东西，更像是一个浸淫 SLAM 领域多年的师兄写的精读笔记。"}]},{"type":"paragraph","attrs":{"id":"07b4fe79-f2ee-453d-9116-730247dc1d17","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":" "}]},{"type":"horizontalRule","attrs":{"id":"531fd3e7-1a30-4597-b031-f43ea6662474","isHoverDragHandle":false}},{"type":"heading","attrs":{"id":"f8ca3ae2-6ed6-4ccd-8502-b0431051d39b","textAlign":"inherit","indent":0,"level":2,"isHoverDragHandle":false},"content":[{"type":"text","text":"几个超出预期的瞬间 "}]},{"type":"heading","attrs":{"id":"f07ed47d-3dcb-4414-b599-7b1cfa7a5430","textAlign":"inherit","indent":0,"level":3,"isHoverDragHandle":false},"content":[{"type":"text","text":"1. 它真的在“补基础”，而不是“默认你懂” "}]},{"type":"paragraph","attrs":{"id":"224d48e9-4243-463b-9bfd-2c7b1a2ebf51","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":"我印象最深的是它对卡方分布 CDF 的解释——没有上来就甩公式，而是从概率统计课上的“独立标准正态变量平方和”讲起，告诉我这个工具原本是用来做假设检验的，现在被用来做“残差→概率”的校准。甚至连 Sampson 误差为什么比普通代数残差好用，都从“图像坐标尺度各向异性”这种底层原理讲清楚。"}]},{"type":"paragraph","attrs":{"id":"60daedd6-41b2-41c9-adcd-fb0d8f4fc30c","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"id":"c8dd2e0b-1df2-424e-875d-a463cdd47855","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":"这种写法对我这种“数学忘得差不多”的人来说太友好了：它不是默认你已经掌握了前置知识，而是悄悄把台阶给你搭好，让你顺着爬上去就能看懂核心公式。"}]},{"type":"paragraph","attrs":{"id":"6453446e-5f8a-4889-8e34-1245427cd057","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":" "}]},{"type":"heading","attrs":{"id":"398bd7ca-204f-4ed4-bbb9-fcc4fbe49352","textAlign":"inherit","indent":0,"level":3,"isHoverDragHandle":false},"content":[{"type":"text","text":"2. 它会“挖隐含逻辑”，而不是“复述原文” "}]},{"type":"paragraph","attrs":{"id":"da8a8819-85e1-45c7-b798-58d1568950aa","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":"原论文里贝叶斯公式那段其实写得挺简略的，就说了“用渲染概率更新像素概率”。但 Workbuddy 专门点出了一个我第一遍读完全忽略的细节："},{"type":"text","marks":[{"type":"bold"}],"text":"先验和似然的角色分配"},{"type":"text","text":"。它指出最精妙的地方不是贝叶斯公式本身，而是把“地图渲染结果”塞到似然的位置——相当于让地图反过来给感知纠错。这个洞察在原论文里只是一句话，但 Workbuddy 把它拎出来当成了核心创新点。"}]},{"type":"paragraph","attrs":{"id":"5601e32e-5c2c-49db-87c8-282259048071","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"id":"5d91aa71-75ba-4477-a35a-6f50fa21d1ea","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":"那一刻我意识到：好的解析不是“翻译”，而是“再发现”。"}]},{"type":"paragraph","attrs":{"id":"15cbe609-9907-46ea-a9e4-fc19074b2dbf","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":" "}]},{"type":"heading","attrs":{"id":"02e79a71-0e29-4cb3-832e-dd0ce32ee751","textAlign":"inherit","indent":0,"level":3,"isHoverDragHandle":false},"content":[{"type":"text","text":"3. 它敢“挑毛病”，而不是“捧臭脚” "}]},{"type":"paragraph","attrs":{"id":"dbc24580-1e00-499c-8818-ba54b6be320d","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":"最惊喜的是最后的“深度思辨”部分。它没有因为论文来自 arXiv 就一味说好，反而直言不讳地提了一堆问题："}]},{"type":"paragraph","attrs":{"id":"ee1b4b5b-12c9-4f3f-b1ce-d546022e7635","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":" "}]},{"type":"bulletList","attrs":{"id":"37fa49d4-1599-460b-a083-83c4672d5b2c","isHoverDragHandle":false},"content":[{"type":"listItem","attrs":{"id":"e26171ec-e77d-4311-acc1-9d4070e31c74"},"content":[{"type":"paragraph","attrs":{"id":"32a1335f-eeb8-48a9-8372-14c3b391ffcc","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":"三个大模型串联的语义管线，和其他只用单模型的对比方法是不是不公平？"}]}]},{"type":"listItem","attrs":{"id":"4b0f1aff-2541-4d0c-b7a7-1e90f282b6df"},"content":[{"type":"paragraph","attrs":{"id":"6e486348-ef42-4358-a544-e9393cf63cb1","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":"连续3帧确认的机制，会不会在行人突然起步时造成滞后？"}]}]},{"type":"listItem","attrs":{"id":"e86b5632-68d4-4e3c-a9e0-d45570ac4483"},"content":[{"type":"paragraph","attrs":{"id":"8fd89808-308a-4ce1-b593-8e3c5a51566a","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":"长时间静止后离开的人，那些“远古高斯”会不会一直留在地图里没人管？"}]},{"type":"paragraph","attrs":{"id":"558882b6-00b4-43b3-be33-ce3c106306a1","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":"这些问题我在第一遍读的时候完全没考虑到，但 Workbuddy 不仅提了，还分析得头头是道。这种批判性思维，是很多 AI 生成内容里最缺的东西。"}]}]}]},{"type":"horizontalRule","attrs":{"id":"2c320754-6369-4bd8-a75d-8cfa114a66a3","isHoverDragHandle":false}},{"type":"heading","attrs":{"id":"79575754-a083-4915-899f-1a70b09b025c","textAlign":"inherit","indent":0,"level":2,"isHoverDragHandle":false},"content":[{"type":"text","text":"几个实用的使用心得 "}]},{"type":"heading","attrs":{"id":"175e159c-e798-4c76-92c5-ac9c25ed08e0","textAlign":"inherit","indent":0,"level":3,"isHoverDragHandle":false},"content":[{"type":"text","text":"1. Prompt 要“给角色”，不要“要结果” "}]},{"type":"paragraph","attrs":{"id":"1e1a81f9-1ad7-4047-9862-c594dfcae4d7","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":"一开始我只发了“解析这篇 DL-SLAM 论文”，出来的内容就比较平。后来我改了 Prompt，加了具体要求："}]},{"type":"paragraph","attrs":{"id":"4ffb9977-9915-414f-88be-7f0b4b83d865","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":" "}]},{"type":"blockquote","attrs":{"id":"d4388e7e-8879-4625-a12a-42579312e217","textAlign":"inherit","isHoverDragHandle":false},"content":[{"type":"paragraph","attrs":{"id":"0a5c2cef-9526-445a-89c0-5a7edb348c2d","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":"“假设你是一个有 SLAM 研究经验的博士生，要给同门做一次组会汇报级别的精读，需要从大学数学基础讲起，重点分析模块间的逻辑关联，最后要客观评价方法的局限性和实验的公平性。”"}]}]},{"type":"paragraph","attrs":{"id":"d4ccebd0-0cbc-4049-958a-5aa2ff865f97","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":"出来的内容质量立刻上了好几个台阶。"},{"type":"text","marks":[{"type":"bold"}],"text":"Workbuddy 很吃“角色设定 场景限定”这一套"},{"type":"text","text":"，你给它一个明确的“身份”和“任务目标”，它就知道该用什么深度、什么语气来写。"}]},{"type":"paragraph","attrs":{"id":"6fae8bda-5f9f-408e-b89f-3fe02c675bfb","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":" "}]},{"type":"heading","attrs":{"id":"0ec299bc-8be3-408e-9f93-f4c4efe65603","textAlign":"inherit","indent":0,"level":3,"isHoverDragHandle":false},"content":[{"type":"text","text":"2. 让它“讲人话”，比让它“讲专业”更重要 "}]},{"type":"paragraph","attrs":{"id":"8c21680e-72e0-4df3-9df0-d10e86d84b92","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":"我特意要求它“避免堆砌术语，遇到复杂概念要先讲物理意义再给公式”。比如在解释动态感知稠密化的时候，它没有上来就说“语义梯度计算”，而是先讲“在人和背景墙的交界处，语义标签经常乱跳，所以要在这些地方多插几个高斯”——一句话就把动机讲透了。"}]},{"type":"paragraph","attrs":{"id":"8337293f-9386-45dc-9ff6-97db58e8bb8a","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"id":"72ad2329-cde3-4fd1-8a43-34491069e436","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"AI 很容易陷入“术语复读”的陷阱，你得明确要求它“先通俗后专业”，它才会真的去解释概念，而不是搬运概念。"}]},{"type":"paragraph","attrs":{"id":"936a65dd-11d0-42a6-b149-b17ca0dd19dc","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":" "}]},{"type":"heading","attrs":{"id":"12597e84-7a5a-4acf-a23d-7984f5ba130b","textAlign":"inherit","indent":0,"level":3,"isHoverDragHandle":false},"content":[{"type":"text","text":"3. 批判性部分是“调”出来的，不是“生”出来的 "}]},{"type":"paragraph","attrs":{"id":"95967911-dcf2-4024-bfb7-05ec55ff0041","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":"一开始它生成的局限性分析比较温和，我就追问了一句：“你现在是审稿人，找一下这篇论文实验设计和工程落地上的硬伤。”它立刻就列出了“没有真实机器人部署”“缺少大尺度场景测试”“语义管线开销过大”这些更尖锐的点。"}]},{"type":"paragraph","attrs":{"id":"a8150af1-381c-4004-9783-25387152a3b2","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"id":"639b2d0b-937c-4a32-aa5a-5fe1899bde38","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Workbuddy 有“讨好倾向”，如果你不主动要求它批判，它会倾向于总结优点。你需要主动切换视角（比如“假设你是审稿人”“假设你是部署工程师”）来逼它跳出赞美模式。"}]},{"type":"paragraph","attrs":{"id":"bc534e06-313b-4cc4-8dd9-0e8ab0914157","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":" "}]},{"type":"horizontalRule","attrs":{"id":"70a6824f-b27a-4497-8de7-9595fe6029de","isHoverDragHandle":false}},{"type":"heading","attrs":{"id":"104409f4-6110-40fb-8a43-ce8303627eb4","textAlign":"inherit","indent":0,"level":2,"isHoverDragHandle":false},"content":[{"type":"text","text":"我对“AI 辅助科研”的新理解 "}]},{"type":"paragraph","attrs":{"id":"3a606ff2-cffa-48ec-9306-696cdcddf02d","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":"这次体验下来，我觉得 Workbuddy 的定位不是“替我读论文”，而是"},{"type":"text","marks":[{"type":"bold"}],"text":"“把我从繁琐的信息整理中解放出来，让我有更多精力做判断和思考”"},{"type":"text","text":"。"}]},{"type":"paragraph","attrs":{"id":"c4f883d9-57af-4e0b-83f4-f7d2e9cc67d3","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":" "}]},{"type":"bulletList","attrs":{"id":"a0bcdb60-1b82-40dd-b47b-d38c68bf5026","isHoverDragHandle":false},"content":[{"type":"listItem","attrs":{"id":"c88a3cbe-9c7d-4215-bb5a-0b5eba3434c1"},"content":[{"type":"paragraph","attrs":{"id":"fcce6db0-0906-45d5-841c-281a8ab05ce6","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":"它帮我梳理了数据流、整理了公式、对比了实验数据——这些机械性的工作它做得又快又好；"}]}]},{"type":"listItem","attrs":{"id":"63a6eae8-0426-4606-8fcb-b08e9ca8d773"},"content":[{"type":"paragraph","attrs":{"id":"ec816e52-86bd-4776-99b1-78fb0c8ce7ba","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":"但判断“这个创新点到底值不值得跟进”“这个实验漏洞会不会影响结论”“这个方法能不能用在我的项目里”——这些需要领域知识和科研直觉的部分，还是得我自己来。"}]}]}]},{"type":"paragraph","attrs":{"id":"fdf6cacd-ec9e-46a8-98dd-d6cfa1d821e5","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":"说白了，它是个"},{"type":"text","marks":[{"type":"bold"}],"text":"超级高效的“科研助理”"},{"type":"text","text":"：能帮你把论文拆成零件摆好，但怎么组装、怎么改进，还得靠你自己。"}]},{"type":"paragraph","attrs":{"id":"9d0471d3-e2f1-412e-86f9-4adcb36664d2","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":" "}]},{"type":"horizontalRule","attrs":{"id":"23ae63bb-0b03-44a0-aae2-a085f52eaa87","isHoverDragHandle":false}},{"type":"heading","attrs":{"id":"ecc6a842-6886-4780-851a-2c07dceab638","textAlign":"inherit","indent":0,"level":2,"isHoverDragHandle":false},"content":[{"type":"text","text":"一个小吐槽和一个小期待 "}]},{"type":"paragraph","attrs":{"id":"5047fb24-09c3-43c3-931d-5fbbcfc0b40c","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"槽点"},{"type":"text","text":"：它对公式的处理虽然准确，但有时候会把公式和上下文割裂开，需要我手动调整一下排版逻辑。另外，如果论文本身有模糊的地方，它偶尔会“脑补”一些合理的解释，这时候需要我回去核对原文。"}]},{"type":"paragraph","attrs":{"id":"4babc343-43a8-4153-9249-3546563a466c","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":" "}]},{"type":"paragraph","attrs":{"id":"71c80034-dd85-44ec-945a-55d3dbcee708","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"期待"},{"type":"text","text":"：以后如果能支持“上传多篇相关论文，帮我做横向对比”就更完美了。比如这次如果它能同时对比 WildGS-SLAM、DG-SLAM 和 DL-SLAM 的架构差异，那价值就更高了。"}]},{"type":"paragraph","attrs":{"id":"89983518-634d-4054-afa4-1343a2142282","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":" "}]},{"type":"horizontalRule","attrs":{"id":"4728ab45-a948-4d64-8e6a-ddb7706e6906","isHoverDragHandle":false}},{"type":"paragraph","attrs":{"id":"8ed9edcf-fb72-4805-97c4-d0b6ba9a9765","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"总结一句话"},{"type":"text","text":"：Workbuddy 已经不是“能用的 AI 阅读工具”，而是“能打 80 分的科研搭档”——剩下那 20 分，需要你用批判性思维和领域知识去补齐。这大概就是人机协作比较好的状态吧。"}]}]}","createTime":1783254810,"ext":{"closeTextLink":0,"comment_ban":0,"description":"","focusRead":0},"favNum":0,"html":"","isOriginal":0,"likeNum":0,