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
从大模型幻觉到概念化智能1
视频
音频
原始脚本
从大模型幻觉到概念化智能，一条务实的 AI 突破路径。 一，开篇，我们为什么怕大模型一本正经的胡说八道？问大模型，鲸鱼有几条腿？它可能斩钉截铁的回答四条。 让它写一篇冬季养生指南，它会编造出吃西瓜能预防感冒的伪知识。 这种明明错了却信誓旦旦的幻觉，成了很多人用 AI 时的心病。 我们怕的不是 AI 不懂，而是它假装懂。 就像遇到一个夸夸其谈的人，你分不清他说的是真知识还是随口编造。 这种不确定性让 AI 的实用性打了折扣。 学生不敢用它查知识点，职场人不敢用它写报告，甚至医生、律师等专业领域，更不敢让它触碰核心决策。 毕竟谁也不想被概率性的错误误导。 可为什么大模型会犯这种低级错误？是它推理时走神了？还是训练时没学好？答案比走神更本质，大模型从一开始就没学会理解规律。 它只学会了模仿人类说话的概率。 就像一个背熟了千万句台词的演员，能流畅的念出关于鲸鱼养生的句子，却不知道这些句子背后哪些是确定的事实，哪些是无稽之谈。 解开 AI 幻觉的谜题，不能只盯着怎么让它少说错话。 更要回到源头，AI 该如何从模仿说话，真正走向理解知识？这正是我们接下来要聊的核心，一条不依赖更大模型、更多数据，而是靠提炼规律、聚焦概念的务实路径。 从大模型幻觉到概念化智能，一条务实的 AI 突破路径二。 大模型的认知局限，只记现象流水账，不悟规律本质。 要理解大模型的糊涂，我们可以先做个更精准的类比，把它比作古代记录潮汐变化的海边渔民，而人类则是发现潮汐与月球引力相关的科学家。 渔民祖祖辈辈生活在海边，每天都会认真记录潮汐的时间。 初一十五涨大潮，白天涨潮都在正午，下雨天潮汐好像会晚一点。 他记了几十年，能精准复述过去所有潮汐出现的规律。 甚至能根据过去100次初一都涨大潮，推测下次初一也会涨大潮。 但他永远不会去想为什么初一十五会涨大潮，更不会总结出潮汐与月球引力相关这个核心规律。 他只关心现象出现的频率，不关心现象背后的逻辑。 他也分不清必然规律和偶然关联，比如把下雨天潮汐晚一点这种偶然出现的关联当成和初一涨大潮一样的规律记下来。 大模型的学习方式和这位渔民几乎一模一样。 他在训练时阅读亿万篇文本。 本质是在统计文字组合的概率。 看到天空是后面接蓝色的概率高达99%，就记住天空是蓝色。 看到鸟喙后面接飞的概率是95%，就记住鸟会飞。 甚至看到下雨天和堵车常一起出现，也会把下雨天堵车当成该记住的关联。 他就像在记一本文字接龙流水账，账上记满了哪两个词常一起出现，却从来不会问为什么天空是蓝色，为什么鸟会飞。 更不会分辨天空是蓝色是必然规律，而下雨天堵车只是偶然关联。 这种只记概率不问因果的认知方式，直接导致了他的幻觉。 比如他看到鲸鱼是哺乳动物，哺乳动物大多有腿，这两个高概率关联，就会默认鲸鱼有腿。 他不会去想鲸鱼是不是哺乳动物里的例外，更不会去验证有没有鲸鱼美腿的反例。 就像渔民若看到多数下雨天会堵车，就会默认只要下雨就会堵车，却忽略了凌晨下雨时道路空旷不会堵车这个例外。 甚至会把某次下雨时恰好有交通事故导致堵车，当成下雨必然堵车的证据，进一步加深错误认知。 更关键的是，大模型没有规律的判断力。 人类知道天空是蓝色是规律，因为这是光的散射原理决定的，无例外。 而鸟会飞不是规律，因为有鸵鸟、企鹅等例外。 下雨天堵车更不是规律，只是常见的偶然关联。 但大模型分不清高概率现象、偶然关联和无例外规律。 在它眼里，99%概率的文字组合、60%概率的偶然关联和100%确定的规律没有本质区别，都是该记住的流水账条目。 这就是大模型最核心的认知局限，它能精准模仿人类语言的表象，却永远触不到规律的本质。 要让它不胡说八道，光给它更多流水账、更多数据、更大的账本、更大模型，没用。 我们需要教他像人类科学家一样，从流水账里提炼出无例外的规律，而不是把所有文字关联都当成知识。
修正脚本
从大模型幻觉到概念化智能，一条务实的 AI 突破路径。 一、开篇，我们为什么怕大模型一本正经地胡说八道？问大模型，鲸鱼有几条腿？它可能斩钉截铁地回答四条。 让它写一篇冬季养生指南，它会编造出吃西瓜能预防感冒的伪知识。 这种明明错了却信誓旦旦的幻觉，成了很多人用 AI 时的心病。 我们怕的不是 AI 不懂，而是它假装懂。 就像遇到一个夸夸其谈的人，你分不清他说的是真知识还是随口编造。 这种不确定性让 AI 的实用性打了折扣。 学生不敢用它查知识点，职场人不敢用它写报告，甚至医生、律师等专业领域，更不敢让它触碰核心决策。 毕竟谁也不想被概率性的错误误导。 可为什么大模型会犯这种低级错误？是它推理时走神了？还是训练时没学好？答案比走神更本质，大模型从一开始就没学会理解规律。 它只学会了模仿人类说话的概率。 就像一个背熟了千万句台词的演员，能流畅地念出关于鲸鱼养生的句子，却不知道这些句子背后哪些是确定的事实，哪些是无稽之谈。 解开 AI 幻觉的谜题，不能只盯着怎么让它少说错话。 更要回到源头，AI 该如何从模仿说话，真正走向理解知识？这正是我们接下来要聊的核心，一条不依赖更大模型、更多数据，而是靠提炼规律、聚焦概念的务实路径。 二、大模型的认知局限，只记现象流水账，不悟规律本质。 要理解大模型的糊涂，我们可以先做个更精准的类比，把它比作古代记录潮汐变化的海边渔民，而人类则是发现潮汐与月球引力相关的科学家。 渔民祖祖辈辈生活在海边，每天都会认真记录潮汐的时间。 初一十五涨大潮，白天涨潮都在正午，下雨天潮汐好像会晚一点。 他记了几十年，能精准复述过去所有潮汐出现的规律。 甚至能根据过去100次初一都涨大潮，推测下次初一也会涨大潮。 但他永远不会去想为什么初一十五会涨大潮，更不会总结出潮汐与月球引力相关这个核心规律。 他只关心现象出现的频率，不关心现象背后的逻辑。 他也分不清必然规律和偶然关联，比如把下雨天潮汐晚一点这种偶然出现的关联当成和初一涨大潮一样的规律记下来。 大模型的学习方式和这位渔民几乎一模一样。 它在训练时阅读亿万篇文本。 本质是在统计文字组合的概率。 看到天空是后面接蓝色的概率高达99%，就记住天空是蓝色。 看到鸟类后面接飞的概率是95%，就记住鸟会飞。 甚至看到下雨天和堵车常一起出现，也会把下雨天堵车当成该记住的关联。 它就像在记一本文字接龙流水账，账上记满了哪两个词常一起出现，却从来不会问为什么天空是蓝色，为什么鸟会飞。 更不会分辨天空是蓝色是必然规律，而下雨天堵车只是偶然关联。 这种只记概率不问因果的认知方式，直接导致了它的幻觉。 比如它看到鲸鱼是哺乳动物，哺乳动物大多有腿，这两个高概率关联，就会默认鲸鱼有腿。 它不会去想鲸鱼是不是哺乳动物里的例外，更不会去验证有没有鲸鱼有腿的反例。 就像渔民若看到多数下雨天会堵车，就会默认只要下雨就会堵车，却忽略了凌晨下雨时道路空旷不会堵车这个例外。 甚至会把某次下雨时恰好有交通事故导致堵车，当成下雨必然堵车的证据，进一步加深错误认知。 更关键的是，大模型没有规律的判断力。 人类知道天空是蓝色是规律，因为这是光的散射原理决定的，无例外。 而鸟会飞不是规律，因为有鸵鸟、企鹅等例外。 下雨天堵车更不是规律，只是常见的偶然关联。 但大模型分不清高概率现象、偶然关联和无例外规律。 在它眼里，99%概率的文字组合、60%概率的偶然关联和100%确定的规律没有本质区别，都是该记住的流水账条目。 这就是大模型最核心的认知局限，它能精准模仿人类语言的表象，却永远触不到规律的本质。 要让它不胡说八道，光给它更多流水账、更多数据、更大的账本、更大模型，没用。 我们需要教它像人类科学家一样，从流水账里提炼出无例外的规律，而不是把所有文字关联都当成知识。
英文翻译
From Large Model Hallucinations to Conceptual Intelligence: A Pragmatic Path to AI Breakthroughs. I. Introduction: Why Are We Afraid of Large Models Blatantly Nonsensing? Ask a large model, "How many legs does a whale have?" It might answer with absolute certainty, "Four." Ask it to write a winter health guide, and it might fabricate pseudoscience like "eating watermelon can prevent colds." This kind of confidently wrong hallucination has become a headache for many AI users. What we fear isn't that AI doesn't know, but that it pretends to know. It's like dealing with a boastful person—you can't tell whether what they say is real knowledge or made up on the spot. This uncertainty undermines AI's practicality. Students dare not use it to look up facts, professionals dare not rely on it for reports, and even doctors and lawyers hesitate to let it touch core decisions. After all, no one wants to be misled by probabilistic errors. But why do large models make such basic mistakes? Is it a lapse in reasoning? Or insufficient training? The answer goes deeper than a lapse: from the start, large models never learned to understand rules. They only learned to mimic the probability of human speech. Like an actor who has memorized millions of lines, they can fluently recite sentences about whales and health, but they have no idea which of those sentences reflect solid facts and which are utter nonsense. To solve the riddle of AI hallucinations, we can't just focus on making it say fewer wrong things. We must go back to the source: How can AI move from mimicking speech to truly understanding knowledge? This is the core of what we'll discuss next—a pragmatic path that doesn't rely on bigger models or more data, but on extracting patterns and focusing on concepts. II. The Cognitive Limitations of Large Models: Recording Surface Phenomena, Not Grasping Essential Rules. To understand the confusion of large models, let's make a more precise analogy: compare it to an ancient fisherman who recorded tidal changes, while humans are like scientists who discovered the correlation between tides and the moon's gravity. The fisherman lived by the sea for generations, diligently recording the times of tides every day. "High tide on the first and fifteenth of the lunar month," "high tide at noon on sunny days," "tides seem a bit later on rainy days." He kept records for decades and could accurately recite all past tidal patterns. He could even predict that the next first day of the month would bring a high tide because the previous 100 had done so. But he would never wonder why high tides occur on the first and fifteenth, let alone deduce the core rule linking tides to the moon's gravity. He cares only about the frequency of phenomena, not the logic behind them. He also cannot distinguish necessary rules from coincidental correlations—for instance, he might treat the occasional delay of tides on rainy days as a rule as firm as the high tide on the first of the month. A large model learns in almost the same way. During training, it reads billions of texts. Its essence is to count the probabilities of word combinations. Seeing that after "the sky is" the word "blue" appears with 99% probability, it remembers that the sky is blue. Seeing that after "birds" the word "fly" appears with 95% probability, it remembers that birds fly. It even sees that "rainy days" and "traffic jams" often appear together, so it records the association "rainy days cause traffic jams." It's like keeping a ledger of word chains, filled with which words often appear together, but it never asks why the sky is blue or why birds fly. Nor does it distinguish that "the sky is blue" is a necessary rule, while "rainy days cause traffic jams" is only a coincidental correlation. This cognitive style—recording probabilities without asking causality—directly leads to hallucinations. For example, seeing that "whales are mammals" and "most mammals have legs" are both high-probability associations, it defaults to "whales have legs." It never considers whether whales are an exception among mammals, nor does it verify counterexamples. It's like the fisherman: if most rainy days see traffic jams, he defaults to assuming that rain always causes jams, ignoring the exception of early morning rain on empty roads. He might even take a single rainy day with a traffic accident as evidence that rain inevitably causes jams, further entrenching his error. More critically, large models lack the ability to judge rules. Humans know that "the sky is blue" is a rule, because it's determined by the principle of light scattering and has no exceptions. But "birds fly" is not a rule, because there are exceptions like ostriches and penguins. "Rainy days cause traffic jams" is not a rule at all; it's just a common coincidence. But large models cannot distinguish high-probability phenomena, coincidental correlations, and exceptionless rules. In their view, a 99% probability word combination, a 60% probability coincidence, and a 100% certain rule are all the same—just entries in a ledger to memorize. This is the core cognitive limitation of large models: they can accurately mimic the surface of human language but can never touch the essence of rules. To stop them from talking nonsense, simply giving them more ledgers, more data, bigger books, and larger models won't help. We need to teach them, like human scientists, to extract exceptionless rules from the ledger, rather than treating every word association as knowledge.
back to top