检索并非人工智能的未来——如果真是如此，谷歌早就赢了。
3 分•作者: lamprouge•3 个月前
检索并非人工智能的未来。
如果真是如此，谷歌早已胜出。
还有另一种方法，完全不需要检索。
它每次对话的成本更高，但比糟糕的 GraphRAG 要好。
令我震惊的是，很少有人知道这一点。
但似乎业界尚未过多关注记忆（甚至 Perplexity AI 完全没有长期记忆）。
人类并非真正记住发生的具体经历，
我们记住的是叠加在那之上的一个层面。
是与经历发生在同一天构建起来的东西，
但它不是记录或摘要。
大脑为何要这样做？难道保存真实记忆不更简单吗？
就像大型语言模型一样，如果我们记住每一次经历，我们的大脑就会不堪重负（有照相记忆的人会经历痛苦）。
人工智能代理无法记住一切，否则检索也会变得痛苦。
问问谷歌就知道了；30年来，它一直在努力为你提供最好的简单食谱搜索结果，但仍然无法一次成功。
那又怎样！你取每句话的第一个词和最后一个词，然后稍后重建句子？
不，这同样行不通，无论 Zep AI 多么想让你相信它。
查看原文
retrieval is not the future of Ai<p>if it was, google would have won already<p>there is another way, that doesn't require retrieval at all<p>it costs more per conversation, but its better than a sh*ty graphRAG<p>i am shocked by how few people know this<p>but it seems the industry is not paying much attention to memory yet (hell z.ai doesn't have longterm memory at all)<p>humans don't really remember the exact experience that happened,<p>we remember a layer that sits on top of that<p>something that got built at the same day the experience happened<p>but is not a transcript or a summary<p>why would the brain do that! isn't it easier to save the actual memory?<p>just like LLMs, our brain will be overwhelmed if we remembered every single experience (people with photographic memory go through hell)<p>Ai agents can't remember everything, or retrieval will be hell too<p>just ask google; 30 years of trying to give you the best search results of a simple recipe & it still fails to do that on the first try<p>so what! do you take the first word and the last word of every sentence and reconstruct the sentence later?<p>No, that too doesn't work either, no matter how much Zep AI wants you to believe it