我的征尘是星辰大海。。。
The dirt and dust from my pilgrimage forms oceans of stars...
-------当记忆的篇章变得零碎，当追忆的图片变得模糊，我们只能求助于数字存储的永恒的回忆
作者:黄教授
手机视频列表
AI训练中的闻过则喜
视频
音频
原始脚本
AI 训练中的闻过则喜。 古人云，闻过则喜，意指乐于正视自身过错，从不足中汲取成长力量。 西方也有谚语， failure builds Characters 点明挫折与失败才是淬炼能力的关键。 这套历经时间验证的处世智慧，放在人工智能大模型的训练与迭代中，同样高度契合。 相较于常规的正确样本，错误案例、边界复利才是推动模型能力进阶的核心宝藏。 我们可以从日常学习、软件工程两大场景。 直观理解这一道理。 在校园学习中，老师从不会反复讲解人人都能答对的基础题目，教学重心永远集中在高频错题上。 一道典型错题，集中暴露了知识盲区、思维误区。 把这类问题钻研透彻，收获远大于重复梳理标准答案。 答对题目只能证明掌握了基础内容，甚至有可能是侥幸蒙对。 而错题直指短板，是补齐能力的最优抓手。 这套逻辑在软件工程的测试体系里被运用的淋漓尽致。 软件测试分为正向用例与反向用例。 正向测试仅用来验证基础功能是否达标，是产品上线的最低门槛。 倘若连通顺运行、基础功能这类正向场景都无法通过。 软件便没有继续打磨的意义。 真正考验产品稳定性，决定产品上线的是海量反向测试用例。 工程师会专门针对过往出现的 bug 设计测试项，持续验证问题是否复现，也就是业内所说的回归测试。 就漏洞、异常场景。 极端边界条件是测试工作的重中之重。 一个软件能否稳定可靠，不在于它能完美完成常规任务，而在于它能否抵御异常，规避历史问题。 错误案例的权重在此刻远超正向案例。 将视角拉回大模型训练，若把大模型看作一个不断拟合真实世界的高维函数。 这一规律会更加清晰。 海量常规正确样本本质是让模型沿着既定的数据分布做平滑的解析延拓。 模型在这类场景下输出正确结果是理所应当的事。 只能证明他掌握了主流规律，停留在舒适区里。 不断堆砌同类正向数据，仅仅是加固已有的认知，很难让模型突破能力边界。 而模型产出错误结果的场景，恰恰对应着函数曲线上的拐点。 真实世界从来不是一条平滑简单的直线，而是充满转折、例外、歧义与复杂规则的曲面。 当模型按照固有逻辑泛化却遭遇失败时，就意味着它触碰到了现有认知的边界。 原有的推理模式、知识体系已经无法适配当下的现实场景。 每一个错误案例都是一次精准提醒，标记出模型的能力缺口、数据盲区与逻辑缺陷。 正如围棋兵法中所讲，顺势而为的胜利本就是偶然，充满变数的博弈与困境才是常态。 映射到 AI 领域也是如此，模型依靠简单泛化得出正确答案。 属于偶然情况，需要辨析规则，权衡条件，应对特殊场景的复杂问题，才是落地应用里的常态。 如果只依赖正向样本训练。 模型永远只能应对简单场景，一旦走入现实的拐点便极易出错。 如今行业内也逐渐形成共识，大模型的迭代早已不再单纯追求纸面准确率。 挖掘 hard case 构建负样本数据集，开展对抗性训练，强化回归验证。 所有进阶优化手段，核心都是围绕错误案例展开。 正视错误、收集错误、剖析错误、修正错误，已经成为模型从能用走向好用、耐用的必经之路。 所谓 AI 训练里的闻过则喜。 本质就是摒弃对完美正确率的片面追求，主动拥抱错误与不足。 正确样本筑牢模型的基础底线，而错误案例才是驱动模型突破瓶颈。 贴近真实世界，实现能力跃升的核心动力。 学会从失败中找方向，才是 AI 持续进化的底层逻辑。
修正脚本
AI 训练中的闻过则喜。 古人云，闻过则喜，意指乐于正视自身过错，从不足中汲取成长力量。 西方也有谚语， failure builds Characters 点明挫折与失败才是淬炼能力的关键。 这套历经时间验证的处世智慧，放在人工智能大模型的训练与迭代中，同样高度契合。 相较于常规的正确样本，错误案例、边界案例才是推动模型能力进阶的核心宝藏。 我们可以从日常学习、软件工程两大场景，直观理解这一道理。 在校园学习中，老师从不会反复讲解人人都能答对的基础题目，教学重心永远集中在高频错题上。 一道典型错题，集中暴露了知识盲区、思维误区。 把这类问题钻研透彻，收获远大于重复梳理标准答案。 答对题目只能证明掌握了基础内容，甚至有可能是侥幸蒙对。 而错题直指短板，是补齐能力的最优抓手。 这套逻辑在软件工程的测试体系里被运用得淋漓尽致。 软件测试分为正向用例与反向用例。 正向测试仅用来验证基础功能是否达标，是产品上线的最低门槛。 倘若连通顺运行、基础功能这类正向场景都无法通过，软件便没有继续打磨的意义。 真正考验产品稳定性，决定产品上线的是海量反向测试用例。 工程师会专门针对过往出现的 bug 设计测试项，持续验证问题是否复现，也就是业内所说的回归测试。 就漏洞、异常场景、极端边界条件是测试工作的重中之重。 一个软件能否稳定可靠，不在于它能完美完成常规任务，而在于它能否抵御异常，规避历史问题。 错误案例的权重在此刻远超正向案例。 将视角拉回大模型训练，若把大模型看作一个不断拟合真实世界的高维函数。 这一规律会更加清晰。 海量常规正确样本本质是让模型沿着既定的数据分布做平滑的解析延拓。 模型在这类场景下输出正确结果是理所应当的事。 只能证明它掌握了主流规律，停留在舒适区里。 不断堆砌同类正向数据，仅仅是加固已有的认知，很难让模型突破能力边界。 而模型产出错误结果的场景，恰恰对应着函数曲线上的拐点。 真实世界从来不是一条平滑简单的直线，而是充满转折、例外、歧义与复杂规则的曲面。 当模型按照固有逻辑泛化却遭遇失败时，就意味着它触碰到了现有认知的边界。 原有的推理模式、知识体系已经无法适配当下的现实场景。 每一个错误案例都是一次精准提醒，标记出模型的能力缺口、数据盲区与逻辑缺陷。 正如围棋兵法中所讲，顺势而为的胜利本就是偶然，充满变数的博弈与困境才是常态。 映射到 AI 领域也是如此，模型依靠简单泛化得出正确答案，属于偶然情况，需要辨析规则，权衡条件，应对特殊场景的复杂问题，才是落地应用里的常态。 如果只依赖正向样本训练，模型永远只能应对简单场景，一旦走入现实的拐点便极易出错。 如今行业内也逐渐形成共识，大模型的迭代早已不再单纯追求纸面准确率。 挖掘 hard case 构建负样本数据集，开展对抗性训练，强化回归验证。 所有进阶优化手段，核心都是围绕错误案例展开。 正视错误、收集错误、剖析错误、修正错误，已经成为模型从能用走向好用、耐用的必经之路。 所谓 AI 训练里的闻过则喜。 本质就是摒弃对完美正确率的片面追求，主动拥抱错误与不足。 正确样本筑牢模型的基础底线，而错误案例才是驱动模型突破瓶颈，贴近真实世界，实现能力跃升的核心动力。 学会从失败中找方向，才是 AI 持续进化的底层逻辑。
英文翻译
In AI training, be glad to hear of one's mistakes. An ancient saying goes, "rejoice when hearing of one's faults," meaning one should be willing to face their own errors and draw strength for growth from shortcomings. Similarly, a Western proverb says, "Failure builds character," highlighting that setbacks and failures are the true crucibles for honing abilities. This time-tested wisdom of life is highly relevant to the training and iteration of large AI models. Compared with regular correct samples, error cases and boundary cases are the core treasures that drive model capability advancement. We can intuitively understand this principle through two scenarios: daily learning and software engineering. In academic learning, teachers never repeatedly explain basic questions that everyone can answer; the teaching focus always centers on high-frequency error-prone questions. A typical wrong question exposes knowledge gaps and thinking pitfalls. Thoroughly studying such problems yields far more than simply reviewing standard answers. Answering correctly only proves mastery of basic content, and might even be a lucky guess. But wrong questions directly point to weaknesses, making them the optimal lever for filling competency gaps. This logic is fully demonstrated in the testing system of software engineering. Software testing is divided into positive test cases and negative test cases. Positive tests only verify whether basic functions meet standards, serving as the minimum threshold for product launch. If even positive scenarios like smooth operation and basic functionality fail, the software has no value for further refinement. What truly tests product stability and determines launch readiness is a massive set of negative test cases. Engineers specifically design test items targeting past bugs to continuously verify whether issues reappear—this is known in the industry as regression testing. Vulnerabilities, abnormal scenarios, and extreme boundary conditions are the top priorities of testing. Whether a software is stable and reliable depends not on its flawless completion of routine tasks, but on its ability to withstand anomalies and avoid historical problems. At this point, the weight of error cases far exceeds that of positive cases. Shifting the perspective back to large model training, if we view the model as a high-dimensional function continuously fitting the real world, this rule becomes even clearer. A massive volume of regular correct samples essentially allows the model to perform smooth analytical continuation along the established data distribution. It is only natural for the model to output correct results in such scenarios, proving only that it has mastered mainstream patterns and remains in its comfort zone. Continuously piling up similar positive data merely reinforces existing knowledge and hardly pushes the model beyond its capability boundaries. On the other hand, scenarios where the model produces incorrect results exactly correspond to inflection points on the function curve. The real world is never a smooth, simple straight line but a surface full of turns, exceptions, ambiguities, and complex rules. When a model generalizes based on its inherent logic but encounters failure, it means it has hit the boundary of its current knowledge. The existing reasoning patterns and knowledge systems can no longer adapt to the actual situation. Every error case serves as a precise reminder, marking the model's capability gaps, data blind spots, and logical flaws. As the Go strategy says, victories that follow the flow are inherently accidental; games full of variables, challenges, and difficulties are the norm. The same applies to AI. When a model relies on simple generalization to produce correct answers, it is a coincidence. Distinguishing rules, weighing conditions, and handling complex scenarios with special requirements—these are the norm in real-world applications. If training relies solely on positive samples, the model will only ever handle simple scenarios. Once it encounters real-world inflection points, errors are highly likely. Gradually, the industry has formed a consensus: iteration of large models no longer simply pursues paper accuracy. Mining hard cases to build negative sample datasets, conducting adversarial training, and strengthening regression validation—all advanced optimization methods center on error cases. Facing errors, collecting errors, analyzing errors, and correcting errors have become the necessary path for a model to go from barely usable to truly useful and durable. This is what "be glad to hear of one's mistakes" means in AI training. In essence, it means abandoning the one-sided pursuit of perfect accuracy and actively embracing errors and shortcomings. Correct samples build the model's baseline foundation, while error cases are the core driving force that pushes the model to break through bottlenecks, get closer to the real world, and achieve capability leaps. Learning to find direction from failure is the underlying logic of AI's continuous evolution.
back to top