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
别被AI加传统搜索的表面创新骗了
视频
音频
原始脚本
别被 AI 加传统搜索的表面创新骗了,本质是用法拉利拉磨,本末倒置。 MIT 这篇递归语言搜索的论文,看似是 AI 与传统检索的取长补短,实则是典型的降维使用大模型,用最强大的 AI 大脑,做最底层最廉价的关键字生成监工。 浪费算力,又回避了长上下文问题的核心矛盾。 本质是用新瓶装旧酒,还装错了酒。 一,表面的创新结合藏着致命的本末倒置。 文章的核心逻辑是让大模型生成递归式搜索指令,指挥传统 Grab 反向索引完成检索,再整合结果输出答案。 乍一看既用了 AI 的智能,又借了传统检索的效率,似乎是1+1大于2的完美组合。 但深究本质,这是把大模型的核心价值用错了地方。 传统检索的短板恰恰是大模型的长板,传统关键字搜索解决不了模糊查询、语义理解、线索关联。 比如天龙八部加张无忌的人物错位、宝藏与密宝的同同义词替换,张三说,后续告诉你 的线索传递,而这些正是大模型预训练后具备的核心能力。 能理解语义、修正误差、关联上下文,是解决长文本模糊查找的终极武器。 大模型的工作恰恰是传统工具能替代的。 让大模型生成第5张加张三加藏宝图这类精准关键字指令,本质是低阶的规则匹配与关键词组合。 这类工作用传统的 NLP 工具、规则引擎甚至简单的脚本就能完成,根本不需要调动千亿参数的大模型。 就像用法拉利拉磨,不是不能做,而是完全不合算。 既浪费了法拉利的速度与性能,又解决不了拉磨本身的效率问题。 更讽刺的是,这种 AI 指挥传统检索的模式,还把大模型的优势彻底抵消。 大模型能一次性理解长上下文,关联所有线索,却被拆成多次递归调用,每次只做生成几个关键字的简单工作。 相当于把一个能通盘解题的天才,拆成只会算加减的小学生,反复做重复劳动,最终效果远不如让大模型直接处理。 二、真正的取长补短,是让大模型做高级事,传统工具做辅助事。 对比业界主流的 RAG 方案,就能清晰看出什么是正确的分工。 RAG 的核心逻辑,用向量检索,传统工具的升级,做快速定位,把长文本分块向量化,通过语义相似度召回相关片段,解决长上下文装不下的问题。 再用大模型做深度理解与整合,把召回的片段一次性喂给大模型,让其基于完 完整语义生成答案,解决模糊查询、逻辑关联的问题。 本质是传统工具向量检索,做擅长的高效检索,大模型做擅长的语义理解与推理。 两者各司其职,用低成本的检索放大大模型的能力,实现1+1大于2。 而 MIT 方案的问题是彻底颠倒了分工。 让大模型做传统工具能做的关键字生成,让传统检索做自己做不好的语音模糊查找,最终陷入多次递归调用、加低效检索、加误差传递的死循环。 成本飙升、延迟不可控、效果还不如 RAG,完全是吃力不讨好。 三、拔高维度看大模型的价值,是做决策与创造,而非执行与监工,回到长上下文问题的本质。 我们真正需要的是让大模型发挥高级智能,而非沦为低级工具的附属。 一、让大模型做工作流生成,而非单次指令生成。 面对复杂的长文本检索任务,大模型不该一次次生成搜张三加藏宝图,搜第5张加张三这类零散指令,而应一次性生成完整的递归搜索脚本或工作流。 比如先泛搜宝藏加张三,筛选含线索的段落,提取第5张关键词,再精准搜第5张加张三加藏宝图,若未找到则扩展同义词密宝重试,最后整合所有结果。 这种一次性生成可执行逻辑才是大模型的高级能力。 既减少了多次 API 调用的成本,又能覆盖复杂的检索逻辑,远比反复生成单个指令高效。 二、让大模型做语义推理与筛选,而非关键字匹配。 对于长文本中的模糊线索隐含关联,大模型应直接基于预训练的语义知识库。 成线索提取、误差修正、逻辑关联。 比如自动纠正天龙八部加张无忌的错误,关联宝藏与密宝的同义词,识别张三说后续告诉你的隐含线索,再指挥传统检索工具精准定位,而非让传统检索先做无效的关键字匹配,再让大模型擦屁股。 三,终极方向,内生式记忆而非外挂式检索。 无论是 RAG 还是 MIT 方案,都是外挂式解决长上下文问题,本质是临时借用外部信息,而非真正记住上下文。 而你提到的边训练边推理,持续学习才是内生式的终极解法。 让大模型在对话中实时筛选、消化上下文,把有用信息融入自身参数,形成动态记忆,从根本上突破上下文窗口限制。 这才是大模型该做的高级事,模仿人类的去粗取精、去伪存真,实现真正的智能记忆,而非反复做检索生成的机械劳动。 四、总结,别被表面创新迷惑,本质是方向错了。 MIT 这篇论文的递归语言搜索,看似是 AI 与传统 检索的创新结合,实则是用大模型的高成本弥补传统检索的低效率,既浪费了大模型的核心价值,又解决不了长上下文的核心问题,模糊查询与语义关联。 真正的创新从来不是让强者做弱者的事,而是让强者做强者的事,弱者做弱者的事。 大模型负责决策、推理、创造、生成工作流,传统工具负责高效检索、执行简单指令。 要么就走向内生式持续学习,让大模型真正拥有动态记忆。 而用大模型生成递归关键字,指挥传统 Grab 检索。 本质是用法拉利拉磨,用大炮打蚊子。 不是不能做,而是完全不合算。 既没有学术上的核心突破,也没有工程上的落地价值。 不过是媒体夸大其词的花边新闻,离真正解决长上下文问题还差着十万八千里。
修正脚本
别被 AI 加传统搜索的表面创新骗了,本质是用法拉利拉磨,本末倒置。 MIT 这篇递归语言搜索的论文,看似是 AI 与传统检索的取长补短,实则是典型的降维使用大模型,用最强大的 AI 大脑,做最底层最廉价的关键字生成监工。 浪费算力,又回避了长上下文问题的核心矛盾。 本质是用新瓶装旧酒,还装错了酒。 一,表面的创新结合藏着致命的本末倒置。 文章的核心逻辑是让大模型生成递归式搜索指令,指挥传统 Grab 反向索引完成检索,再整合结果输出答案。 乍一看既用了 AI 的智能,又借了传统检索的效率,似乎是1+1大于2的完美组合。 但深究本质,这是把大模型的核心价值用错了地方。 传统检索的短板恰恰是大模型的长板,传统关键字搜索解决不了模糊查询、语义理解、线索关联的问题。 比如天龙八部加张无忌的人物错位、宝藏与密宝的同义词替换,张三说,后续告诉你的线索传递,而这些正是大模型预训练后具备的核心能力。 能理解语义、修正误差、关联上下文,是解决长文本模糊查找的终极武器。 如今方案里让大模型做的工作恰恰是传统工具能替代的。 让大模型生成第5张加张三加藏宝图这类精准关键字指令,本质是低阶的规则匹配与关键词组合。 这类工作用传统的 NLP 工具、规则引擎甚至简单的脚本就能完成,根本不需要调动千亿参数的大模型。 就像用法拉利拉磨,不是不能做,而是完全不合算。 既浪费了法拉利的速度与性能,又解决不了拉磨本身的效率问题。 更讽刺的是,这种 AI 指挥传统检索的模式,还把大模型的优势彻底抵消。 大模型能一次性理解长上下文,关联所有线索,却被拆成多次递归调用,每次只做生成几个关键字的简单工作。 相当于把一个能通盘解题的天才,拆成只会算加减的小学生,反复做重复劳动,最终效果远不如让大模型直接处理。 二、真正的取长补短,是让大模型做高级事,传统工具做辅助事。 对比业界主流的 RAG 方案,就能清晰看出什么是正确的分工。 RAG 的核心逻辑是用向量检索,作为传统工具的升级,做快速定位,把长文本分块向量化,通过语义相似度召回相关片段,解决长上下文装不下的问题。 再用大模型做深度理解与整合,把召回的片段一次性喂给大模型,让其基于完整语义生成答案,解决模糊查询、逻辑关联的问题。 本质是传统工具向量检索,做擅长的高效检索,大模型做擅长的语义理解与推理。 两者各司其职,用低成本的检索放大大模型的能力,实现1+1大于2。 而 MIT 方案的问题是彻底颠倒了分工。 让大模型做传统工具能做的关键字生成,让传统检索做自己做不好的语义模糊查找,最终陷入多次递归调用、加低效检索、加误差传递的死循环。 成本飙升、延迟不可控、效果还不如 RAG,完全是吃力不讨好。 三、拔高维度看大模型的价值,是做决策与创造,而非执行与监工,回到长上下文问题的本质。 我们真正需要的是让大模型发挥高级智能,而非沦为低级工具的附属。 一,让大模型做工作流生成,而非单次指令生成。 面对复杂的长文本检索任务,大模型不该一次次生成搜张三加藏宝图,搜第5张加张三这类零散指令,而应一次性生成完整的递归搜索脚本或工作流。 比如先泛搜宝藏加张三,筛选含线索的段落,提取第5张关键词,再精准搜第5张加张三加藏宝图,若未找到则扩展同义词密宝重试,最后整合所有结果。 这种一次性生成可执行逻辑才是大模型的高级能力。 既减少了多次 API 调用的成本,又能覆盖复杂的检索逻辑,远比反复生成单个指令高效。 二、让大模型做语义推理与筛选,而非关键字匹配。 对于长文本中的模糊线索隐含关联,大模型应直接基于预训练的语义知识库,完成线索提取、误差修正、逻辑关联。 比如自动纠正天龙八部加张无忌的错误,关联宝藏与密宝的同义词,识别张三说后续告诉你的隐含线索,再指挥传统检索工具精准定位,而非让传统检索先做无效的关键字匹配,再让大模型擦屁股。 三,终极方向,内生式记忆而非外挂式检索。 无论是 RAG 还是 MIT 方案,都是外挂式解决长上下文问题,本质是临时借用外部信息,而非真正记住上下文。 而你提到的边训练边推理,持续学习才是内生式的终极解法。 让大模型在对话中实时筛选、消化上下文,把有用信息融入自身参数,形成动态记忆,从根本上突破上下文窗口限制。 这才是大模型该做的高级事,模仿人类的去粗取精、去伪存真,实现真正的智能记忆,而非反复做检索生成的机械劳动。 四、总结,别被表面创新迷惑,本质是方向错了。 MIT 这篇论文的递归语言搜索,看似是 AI 与传统检索的创新结合,实则是用大模型的高成本弥补传统检索的低效率,既浪费了大模型的核心价值,又解决不了长上下文的核心问题——模糊查询与语义关联。 真正的创新从来不是让强者做弱者的事,而是让强者做强者的事,弱者做弱者的事。 大模型负责决策、推理、创造、生成工作流,传统工具负责高效检索、执行简单指令。 要么就走向内生式持续学习,让大模型真正拥有动态记忆。 而用大模型生成递归关键字,指挥传统 Grab 检索,本质是用法拉利拉磨,用大炮打蚊子。 不是不能做,而是完全不合算。 既没有学术上的核心突破,也没有工程上的落地价值。 不过是媒体夸大其词的花边新闻,离真正解决长上下文问题还差着十万八千里。
英文翻译
Don't be fooled by the superficial innovation of AI plus traditional search—it's essentially using a Ferrari to grind grain, putting the cart before the horse. MIT's paper on recursive language search appears to combine the strengths of AI and traditional retrieval, but in reality, it is a typical case of downgrading the use of large models, employing the most powerful AI brain to act as a cheap supervisor for keyword generation. It wastes computing power while avoiding the core contradiction of long-context problems. Essentially, it's old wine in a new bottle, and the wine is even mislabeled. First, this seemingly innovative combination hides a fatal inversion of priorities. The core logic of the paper is to have the large model generate recursive search commands, directing the traditional Grab inverted index to complete the retrieval, and then integrate the results to output an answer. At first glance, it appears to leverage both AI intelligence and traditional retrieval efficiency—a perfect 1+1>2 combination. But looking deeper, it misuses the core value of large models. The shortcoming of traditional retrieval is precisely the strength of large models. Traditional keyword search cannot solve issues like fuzzy queries, semantic understanding, or clue association. For example, the character mismatch in "Demi-Gods and Semi-Devils" plus "Zhang Wuji", synonym substitution between "treasure" and "cache", or the clue transmission in "Zhang San said, 'I'll tell you later'". These are exactly the core capabilities that large models possess after pre-training: understanding semantics, correcting errors, and linking context—the ultimate weapon for solving fuzzy searches in long texts. In this scheme, the work assigned to the large model is precisely what traditional tools can replace. Having the large model generate precise keyword commands like "Chapter 5 + Zhang San + Treasure Map" is essentially low-level rule matching and keyword combination. Such tasks can be accomplished with traditional NLP tools, rule engines, or even simple scripts—there is no need to mobilize a billion-parameter large model. It's like using a Ferrari to grind grain: not impossible, but completely uneconomical. It wastes the Ferrari's speed and performance while failing to solve the efficiency problem of grinding itself. More ironically, this AI-directed traditional retrieval mode completely cancels out the advantages of the large model. A large model can understand long contexts all at once and correlate all clues, but it is forced into multiple recursive calls, each time only generating a few keywords. This is equivalent to breaking a genius who can solve an entire problem into elementary students who only add and subtract, repeating the same labor—ultimately far less effective than letting the large model handle it directly. Second, true complementary strengths involve letting the large model handle advanced tasks while traditional tools assist. Comparing mainstream RAG solutions clearly shows what correct division of labor looks like. The core logic of RAG uses vector retrieval as an upgraded traditional tool for fast localization: chunking long texts into vectors, retrieving relevant segments via semantic similarity, solving the problem of long contexts that cannot fit entirely. Then the large model performs deep understanding and integration, feeding the retrieved segments at once to generate answers based on complete semantics, solving fuzzy queries and logical association issues. In essence, traditional tools (vector retrieval) do what they are good at—efficient retrieval—while the large model does what it excels at—semantic understanding and reasoning. Each performs its own function, using low-cost retrieval to amplify the large model's capabilities, achieving 1+1>2. The MIT solution, however, completely reverses the division. It makes the large model do what traditional tools can do (keyword generation) and makes traditional retrieval do what it cannot do well (fuzzy semantic search), ultimately falling into a dead loop of multiple recursive calls, inefficient retrieval, and error propagation. Costs skyrocket, latency becomes uncontrollable, and the result is worse than RAG—completely counterproductive. Third, from a higher perspective, the value of large models lies in decision-making and creation, not execution and supervision—returning to the essence of long-context problems. What we truly need is to let large models exert advanced intelligence, not become slaves to low-level tools. One: Let large models generate workflows, not single instructions. For complex long-text retrieval tasks, the large model should not repeatedly generate scattered instructions like "search Zhang San + Treasure Map" or "search Chapter 5 + Zhang San". Instead, it should generate a complete recursive search script or workflow at once. For example: first broadly search "Treasure + Zhang San", filter paragraphs with clues, extract the keyword "Chapter 5", then precisely search "Chapter 5 + Zhang San + Treasure Map"; if not found, expand to the synonym "cache" and retry; finally integrate all results. This one-time generation of executable logic is the advanced capability of large models. It reduces the cost of multiple API calls while covering complex retrieval logic—far more efficient than repeatedly generating single instructions. Two: Let large models perform semantic reasoning and filtering, not keyword matching. For fuzzy clues and implicit associations in long texts, the large model should directly use its pre-trained semantic knowledge base to extract clues, correct errors, and associate logic. For example, automatically correct "Demi-Gods and Semi-Devils + Zhang Wuji", link synonyms like "treasure" and "cache", identify the implicit clue in "Zhang San said, 'I'll tell you later'", and then direct traditional retrieval tools to precisely locate—rather than having traditional retrieval first perform ineffective keyword matching and then having the large model clean up the mess. Three: The ultimate direction is endogenous memory rather than external retrieval. Both RAG and the MIT solution are external approaches to long-context problems—essentially borrowing external information temporarily, not truly remembering context. What you mentioned—simultaneous training and inference, continuous learning—is the ultimate endogenous solution. Let the large model filter and digest context in real-time during dialogue, incorporate useful information into its own parameters, forming dynamic memory—fundamentally breaking through the context window limit. This is the advanced task large models should do: imitating humans' ability to discard the dross and retain the essence, distinguishing truth from falsehood, achieving true intelligent memory—not mechanically repeating retrieval and generation. Fourth, in summary, don't be fooled by superficial innovation—the fundamental direction is wrong. MIT's recursive language search appears as an innovative combination of AI and traditional retrieval, but it is actually using the high cost of large models to compensate for the inefficiency of traditional retrieval. It wastes the core value of large models and fails to solve the core problem of long contexts: fuzzy queries and semantic association. True innovation never makes the strong do the weak's work; it lets the strong do what they excel at and the weak do what they excel at. Large models are responsible for decision-making, reasoning, creation, and workflow generation; traditional tools are responsible for efficient retrieval and executing simple instructions. Alternatively, move toward endogenous continuous learning, allowing large models to truly possess dynamic memory. Using large models to generate recursive keywords and command traditional Grab retrieval is essentially using a Ferrari to grind grain, or using a cannon to kill mosquitoes. Not impossible, but completely uneconomical. It offers neither a core academic breakthrough nor practical engineering value. It's just a sensational gimmick blown up by the media, still tens of thousands of miles away from truly solving long-context problems.
back to top