AI 与机器学习2015static.googleusercontent.com编辑精选
机器学习：技术债的高息信用卡
Machine Learning: The High-Interest Credit Card of Technical Debt
进阶 · 约 35 分钟读完前置知识：机器学习基础、软件工程、技术债概念
导读
这篇论文讨论机器学习系统在生产环境中引入的隐性技术债，把 ML 比作“高息信用卡”：短期收益快，但长期维护成本高。作者梳理了数据依赖、反馈回路、配置与胶水代码、系统纠缠等 ML 特有债务来源，提醒团队在工程治理上付出持续代价。
HN 关注点：ML 落地后的维护成本、反馈回路污染实验，以及这一比喻是否公平。
后来怎么样
该文系统提出 ML 生产系统中的隐性技术债，“高息信用卡”比喻和 ML 技术债概念被 MLOps/ML 工程实践广泛采用，成为后续 ML 系统与平台讨论的常用框架。
HN 评论精选
这几条评论里：质疑 2 补充 1 亲历 1
亲历A/B 测试中的反馈回路会让旧模型吸收新模型的行为，导致实验提升被稀释甚至消失。
我经历过类似问题。设想一个在线购物推荐服务，使用简单的共现模型，也就是“看过此商品的人还看过”。理想情况下模型最终达到稳态，推荐大体有用。现在有人提出一个几乎完美的替代算法，点击率提升很大，并拿它和原系统做测试。原共现系统能访问所有数据，包括暴露给新算法的会话数据。如果 A/B 测试跑得够久，原系统会学到新系统的行为并模仿它，因为对给定种子商品，很多共现点击会落在新系统推荐的物品上。虽然新系统起初有提升，最终两个系统会趋同显示改进后的推荐，提升趋于零。
bcbrown查看英文原文
I've experienced issues like those mentioned. Imagine a recommendation service for online shopping, using a simple co-occurrence model, aka "People who viewed this also viewed". Hopefully, your model will eventually hit a steady state, with generally helpful recommendations. Now, imagine someone comes up with an almost-perfect replacement algorithm, that provides a high lift in clickthrough rate, that's being tested against the original system. The original co-occurrence system has access to all the data, including that from sessions exposed to the new algorithm. If the A/B test runs for long enough, the original system will learn the new system's behaviour and emulate it, because for a given seed item, a lot of the co-occurring clicks will be on items recommended by the new system. Although initially the new system will show a lift, eventually the two systems will tend towards showing the improved recommendations, and the lift will tend towards zero.
质疑认为论文只讲缺点不讲优点，ML 的维护/重训成本常低于非 ML 方案，“高息信用卡”比喻过重。
论文列出了机器学习的一些缺点，却跳过了优点。比如最近 HN 上的“Google Translate 如何把深度学习塞进手机”，它可能因为世界变化（新字体流行）而效果变差，但相比调整非 ML 方案（如果该问题存在非 ML 方案），直接在新字体上重训可能更容易。把 ML 称为高息信用卡似乎不公平，叫偶尔的服务费可能更贴切。
tim333查看英文原文
The paper lists some of the downsides of machine learning but skips the upsides. If you think of something like "How Google Translate squeezes deep learning onto a phone" featured here recently (https://news.ycombinator.com/item?id=9969000 ) it may stop working well because the world changes in that new fonts get trendy but it is probably easier just to retrain on new fonts than to adjust a non machine learning solution - if there is one for that problem. High-Interest credit card seems unfair. Occasional service charge might be closer.
补充批评把 ML 用在更适合机制模型的场景，并期待概率编程等弥合两类方法。
另一个大缺点是 ML 的错误使用：有些场景更适合机制性模型。最近很多系统生物学论文看起来很荒谬，就像有人试图用 SVM 发现牛顿运动定律。希望概率编程之类能弥合 ML 与经典模型之间的鸿沟。
nextos查看英文原文
A big downside too is incorrect usage of ML, where a more mechanistic model would be more useful. Many recent systems biology research is full of papers that look as ridiculous as someone trying to discover Newton Laws of Motion with an SVM. Hopefully things like probabilistic programming bridge the gap between ML and classical models.
质疑质疑论文是否公平比较 ML 技术债与过度工程化企业系统的长期成本，并批评后半部分像软件工程说教。
有没有论文比较过 ML 方案造成的技术债成本，与过度工程化、抽象沉重、会不断减少能有效维护它的人的企业方案长期上升的成本？也许只是我的感觉，但从“System-level Spaghetti”开始，这篇论文读起来像时薪 1000 美元的 Java 顾问写的：拼命推销一套已被数据库、大量 GPU 服务器、纠缠统计加库胶水代码取代的软件工程玩具箱。它有点道理：NN 应在定义良好的问题上训练，需要更高层系统逻辑来整合各 NN 的专业能力，并监控关键输出漂移；但这听起来就是训练 NN 需要好的“手艺”，而这对几乎一切事情都成立。
Pica_soO查看英文原文
Is there a paper comparing the cost of technical debt, as created by using a machine learning solution- versus the permanent rising cost of having a over-engineered, abstraction heavy enterprise solution - that automatically reduces the people who can work effective on it? Maybe it is just me, but this paper reads from the paragraph "System-level Spaghetti" on-wards like written by one of those java-consultants for 1000 $ a hour. Someone - who is desperately trying to sell this software-engineering toy-toolbox, which has been replaced with a database, lots of graphic-cards server-racks, some admittelty tangled clot statistics plus library glue-code. There is some merits to the point, that a NN should be trained on a well defined problem, and that there is a need for a higher system logic bundling the specialist-abilitys of the NNs, also having a eye on some key-output-parameters drift within boundaries. But that sounds a lot like - you need good "craftsmanship" to train a NN. Which is true- true for nearly all things.
历次讨论
- 2015Hacker News 讨论帖135 分 · 26 评
- 2018Hacker News 讨论帖20 分 · 1 评
同领域还值得读
DeepSeek-R1：通过强化学习激励 LLM 的推理能力
DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via RL
研究如何用强化学习激励大语言模型的推理能力，提出 DeepSeek-R1，并探讨 RL 训练对模型推理表现的影响。
AI 与机器学习2025 · 1,351 分 · 1,056 评 · 引用 30进阶 · 约 80 分钟编辑精选1-bit LLM 时代：面向低成本计算的三值参数
The Era of 1-bit LLMs: ternary parameters for cost-effective computing
提出用三值参数 {-1, 0, 1} 替代 LLM 中的浮点权重，即所谓“1-bit LLM”。
AI 与机器学习2024 · 1,040 分 · 447 评 · 引用 19进阶 · 约 45 分钟编辑精选DeepSeek-V3.2：推进开放大语言模型的前沿
DeepSeek-v3.2: Pushing the frontier of open large language models
这是 DeepSeek-V3.2 的技术报告，延续了 DeepSeek 开源权重模型的路线，目标是让开放模型继续逼近甚至追平前沿闭源模型。
AI 与机器学习2025 · 982 分 · 465 评进阶 · 约 90 分钟编辑精选通过自我对弈与通用强化学习掌握国际象棋和将棋（AlphaZero）
Mastering Chess and Shogi by Self-Play with General Reinforcement Learning
DeepMind 的 AlphaZero 论文：用通用强化学习算法（自我博弈 + MCTS + 神经网络），从零开始仅数小时训练便在国际象棋和将棋上击败了世界顶级引擎 Stock…
AI 与机器学习2017 · 539 分 · 270 评 · 引用 13进阶 · 约 90 分钟编辑精选用强化学习发现更快的矩阵乘法算法
Discovering faster matrix multiplication algorithms with reinforcement learning
论文提出 AlphaTensor，用强化学习搜索矩阵乘法算法。
AI 与机器学习2022 · 518 分 · 111 评专业 · 约 90 分钟编辑精选