{"topic":{"id":11838,"title":"对 LLM 的虚假宣传太夸张了","slug":"topic","userId":1047,"categoryId":4,"createdAt":"2024-04-27 18:02:26.185306","bumpedAt":"2024-05-01 16:29:52.598796","views":504,"postsCount":7,"highestPostNumber":7,"likeCount":17,"closed":false,"pinned":false,"pinnedGlobally":false,"pinnedAt":null,"pinnedUntil":null,"isBanner":false,"excerpt":"来自这个报告的图片 https://www.bilibili.com/video/BV1Gz42167i8 GPT4 做简单算法题，宣传：全对。实际测 div2-div4 的 problem A，18 年的题目全对，23 年没见过的【全错】。测试集 ⊂\\subset⊂ 训练集，怪不得宣传这么牛逼呢","deletedAt":null,"isAnonymous":false,"isPrivate":false,"isMembersOnly":false,"category":{"id":4,"name":"谈笑风生","slug":"general","color":"25AAE2"},"author":{"id":1047,"username":"anonymous_coward_old","name":"Anonymous Coward Old","avatarUrl":"/uploads/image/43/32/4332638c11248b8ff07033c11de0759a55205facfbec85e14458245a91991faf.avif"},"realAuthor":{"id":1047,"username":"anonymous_coward_old","name":"Anonymous Coward Old","avatarUrl":"/uploads/image/43/32/4332638c11248b8ff07033c11de0759a55205facfbec85e14458245a91991faf.avif"},"tags":[]},"posts":[{"id":30957,"postNumber":1,"raw":"来自这个报告的图片\nhttps://www.bilibili.com/video/BV1Gz42167i8\n\nGPT4 做简单算法题，宣传：全对。\n\n实际测 div2-div4 的 problem A，18 年的题目全对，23 年没见过的【全错】。\n\n测试集 $\\subset$ 训练集，怪不得宣传这么牛逼呢","cooked":"

来自这个报告的图片

\n $\"\"$ \n Bilibili\n

\n $\"\"$ \n

[FAI] Princeton 吕凯风 | 浅谈神经网络在算法推理上的局限性

讲座标题（中文）：浅谈神经网络在算法推理上的局限性\n讲座标题（英文）：Understanding the Limitations of Neural Networks on Algorithmic Reasoning\n讲座摘要：Deep learning has driven large foundation models to succeed in most cases, yet these models struggle with reasoning. In this talk, I will examine their ability on algorithmic reasoning — the ability to perform reasoning following an algorithm.\nThe first part of the talk is about a surprising "grokking" phenomenon when training a neural net on certain algorithmically generated datasets: the neural net first “memorizes” the training set, resulting in perfect training accuracy but near-random test accuracy, and after training for sufficiently longer, it suddenly transitions to perfect test accuracy. We exemplify in simple yet insightful settings that a dichotomy of early and late phase implicit biases makes grokking provably occur.\nThe second part of the talk is about how the architecture of language models limits their reasoning ability. We focus on inference with Chain-of-Thought (CoT), a standard technique to improve reasoning by outputting intermediate steps. While previous results show that Transformers with CoT have sufficient representation power to simulate Turing machines, we prove that, even with CoT, modern RNNs cannot reach the same power, although they have been increasingly popular due to their higher efficiency. We further show that a key bottleneck lies in the inability of RNNs to perfectly retrieve information from the context, but adding just a single Transformer layer can fix this issue.\n讲者信息：Kaifeng Lyu will be joining the Institute for Interdisciplinary Information Sciences (IIIS) at Tsinghua University as an Assistant Professor in the fall of 2025. His primary research interests include machine learning theory, AI safety/alignment, and optimization. Kaifeng Lyu graduated from Yao class at Tsinghua University in 2019. He is currently a final-year PhD student at Princeton University, advised by Prof. Sanjeev Arora.

GPT4 做简单算法题，宣传：全对。

实际测 div2-div4 的 problem A，18 年的题目全对，23 年没见过的【全错】。

测试集 $\\subset$ 训练集，怪不得宣传这么牛逼呢

\n","createdAt":"2024-04-27 18:02:26.288971","updatedAt":"2024-04-27 23:19:18.635531","replyToPostNumber":null,"replyCount":0,"likeCount":2,"wordCount":65,"liked":false,"bookmarked":false,"polls":[],"reactions":[{"reaction":"❤️","count":2,"users":[{"id":1941,"username":"john","name":"John"},{"id":3273,"username":"anon18815924","name":null}],"hasReacted":false}],"wiki":false,"version":2,"publicVersion":2,"deletedAt":null,"isAnonymous":false,"user":{"id":1047,"username":"anonymous_coward_old","name":"Anonymous Coward Old","avatarUrl":"/uploads/image/43/32/4332638c11248b8ff07033c11de0759a55205facfbec85e14458245a91991faf.avif","admin":false,"moderator":false,"signatureRaw":null,"signatureCooked":null}},{"id":30959,"postNumber":2,"raw":"AI is very stupid. It only remembers everything you feed it. You have seen the solutions before you take the exam, you will probably be the No. 1. That's why I am pessimistic about AI, machine learning, LLM and everything in this scope. Conducting such research does NOT make ANY sense. I have seen many artificial tools that regurgitate information, but I haven't seen even one that has logical mind. I don't like people who are doing spurious AI research just apply the model developed by others to the fields they previously worked on. That's industry, not academia. Do you believe it? They even get these suck papers published in top conferences. How ridiculous.","cooked":"

AI is very stupid. It only remembers everything you feed it. You have seen the solutions before you take the exam, you will probably be the No. 1. That's why I am pessimistic about AI, machine learning, LLM and everything in this scope. Conducting such research does NOT make ANY sense. I have seen many artificial tools that regurgitate information, but I haven't seen even one that has logical mind. I don't like people who are doing spurious AI research just apply the model developed by others to the fields they previously worked on. That's industry, not academia. Do you believe it? They even get these suck papers published in top conferences. How ridiculous.

\n","createdAt":"2024-04-27 18:37:20.685046","updatedAt":"2024-04-27 18:40:46.505814","replyToPostNumber":null,"replyCount":0,"likeCount":3,"wordCount":114,"liked":false,"bookmarked":false,"polls":[],"reactions":[{"reaction":"❤️","count":3,"users":[{"id":601,"username":"m0rsun","name":null},{"id":1941,"username":"john","name":"John"},{"id":3273,"username":"anon18815924","name":null}],"hasReacted":false}],"wiki":false,"version":1,"publicVersion":1,"deletedAt":null,"isAnonymous":false,"user":{"id":3285,"username":"platypus","name":"platypus","avatarUrl":"/uploads/image/8d/c9/8dc95eb138bc6d59e76c2244407a8c6ae1f8cf33049c520cabbde1e4145f38ac.avif","admin":false,"moderator":false,"signatureRaw":null,"signatureCooked":null}},{"id":30961,"postNumber":3,"raw":"逻辑推理确实 shit，但在某些任务确实是最佳选择，比如搜索引擎，机器翻译初稿，语义检索，破防柯洁","cooked":"

逻辑推理确实 shit，但在某些任务确实是最佳选择，比如搜索引擎，机器翻译初稿，语义检索，破防柯洁

\n","createdAt":"2024-04-27 18:52:21.732153","updatedAt":"2024-04-27 18:52:21.732153","replyToPostNumber":null,"replyCount":0,"likeCount":3,"wordCount":44,"liked":false,"bookmarked":false,"polls":[],"reactions":[{"reaction":"magic_wand_grin","count":1,"users":[{"id":30,"username":"一只玉米人","name":"玉米🌽"}],"hasReacted":false},{"reaction":"❤️","count":3,"users":[{"id":30,"username":"一只玉米人","name":"玉米🌽"},{"id":1756,"username":"包图第一猫王","name":"上海第二酱料中等职业技术学校"},{"id":3273,"username":"anon18815924","name":null}],"hasReacted":false}],"wiki":false,"version":1,"publicVersion":1,"deletedAt":null,"isAnonymous":false,"user":{"id":1047,"username":"anonymous_coward_old","name":"Anonymous Coward Old","avatarUrl":"/uploads/image/43/32/4332638c11248b8ff07033c11de0759a55205facfbec85e14458245a91991faf.avif","admin":false,"moderator":false,"signatureRaw":null,"signatureCooked":null}},{"id":30974,"postNumber":4,"raw":"赚的是投资人的钱。\n没有东西可以投资把钱扔在账面上可能更让投资人恐慌","cooked":"

赚的是投资人的钱。
\n没有东西可以投资把钱扔在账面上可能更让投资人恐慌

\n","createdAt":"2024-04-28 01:53:51.611092","updatedAt":"2024-04-28 01:53:51.611092","replyToPostNumber":null,"replyCount":2,"likeCount":3,"wordCount":33,"liked":false,"bookmarked":false,"polls":[],"reactions":[{"reaction":"💯️","count":1,"users":[{"id":1288,"username":"LittleWangInShanghai","name":"小王在上海"}],"hasReacted":false},{"reaction":"❤️","count":3,"users":[{"id":1288,"username":"LittleWangInShanghai","name":"小王在上海"},{"id":1941,"username":"john","name":"John"},{"id":4034,"username":"不赢也是一种善良","name":"不赢也是一种善良"}],"hasReacted":false}],"wiki":false,"version":1,"publicVersion":1,"deletedAt":null,"isAnonymous":false,"user":{"id":178,"username":"Hypocrisy","name":"rotartsinimdA","avatarUrl":"/uploads/image/ce/f3/cef3773e378e161524cc7523c8a63dff0698d22802b0c1ab669894d395ee36a3.avif","admin":false,"moderator":true,"signatureRaw":null,"signatureCooked":null}},{"id":30976,"postNumber":5,"raw":"毕竟 2020s 特色焦虑经济，谁掌握了焦虑谁就掌握了财富。","cooked":"

毕竟 2020s 特色焦虑经济，谁掌握了焦虑谁就掌握了财富。

\n","createdAt":"2024-04-28 02:35:54.179683","updatedAt":"2024-04-28 02:35:54.179683","replyToPostNumber":4,"replyCount":0,"likeCount":2,"wordCount":24,"liked":false,"bookmarked":false,"polls":[],"reactions":[{"reaction":"❤️","count":2,"users":[{"id":1941,"username":"john","name":"John"},{"id":3273,"username":"anon18815924","name":null}],"hasReacted":false}],"wiki":false,"version":1,"publicVersion":1,"deletedAt":null,"isAnonymous":false,"user":{"id":1288,"username":"LittleWangInShanghai","name":"小王在上海","avatarUrl":"/uploads/image/63/d2/63d25f4f2f6693f9ab450f32bd8d04e2d29f96a980b51555a585423f113c4a58.avif","admin":false,"moderator":false,"signatureRaw":null,"signatureCooked":null}},{"id":30992,"postNumber":6,"raw":"哈哈，比如月之暗面还没盈利就开始套现了。凭空变出来一堆钱。","cooked":"

哈哈，比如月之暗面还没盈利就开始套现了。凭空变出来一堆钱。

\n","createdAt":"2024-04-28 04:24:02.660361","updatedAt":"2024-04-28 04:24:02.660361","replyToPostNumber":4,"replyCount":0,"likeCount":4,"wordCount":29,"liked":false,"bookmarked":false,"polls":[],"reactions":[{"reaction":"sweat_smile_pleading_face","count":1,"users":[{"id":1288,"username":"LittleWangInShanghai","name":"小王在上海"}],"hasReacted":false},{"reaction":"🤣️","count":1,"users":[{"id":1688,"username":"ForgottonAge","name":null}],"hasReacted":false},{"reaction":"❤️","count":4,"users":[{"id":1047,"username":"anonymous_coward_old","name":"Anonymous Coward Old"},{"id":1288,"username":"LittleWangInShanghai","name":"小王在上海"},{"id":1688,"username":"ForgottonAge","name":null},{"id":1756,"username":"包图第一猫王","name":"上海第二酱料中等职业技术学校"}],"hasReacted":false}],"wiki":false,"version":1,"publicVersion":1,"deletedAt":null,"isAnonymous":false,"user":{"id":346,"username":"hoist","name":null,"avatarUrl":"/uploads/image/0d/cd/0dcdf69d213b6798002779ad4b729567cd4d889ee4a56c07a8c4454e8b563219.avif","admin":false,"moderator":false,"signatureRaw":null,"signatureCooked":null}},{"id":31244,"postNumber":7,"raw":"https://cn.wsj.com/articles/%E5%B0%9A%E6%97%A0%E4%B8%9A%E5%8A%A1%E7%9A%84ai%E5%88%9D%E5%88%9B%E5%85%AC%E5%8F%B8%E4%B8%BA%E4%BD%95%E8%83%BD%E6%8B%A5%E6%9C%89%E5%A4%A7%E9%87%8F%E7%8E%B0%E9%87%91-3d9e4308","cooked":"","createdAt":"2024-05-01 16:29:52.598796","updatedAt":"2024-05-01 16:29:52.598796","replyToPostNumber":null,"replyCount":0,"likeCount":0,"wordCount":1,"liked":false,"bookmarked":false,"polls":[],"reactions":[],"wiki":false,"version":1,"publicVersion":1,"deletedAt":null,"isAnonymous":false,"user":{"id":1047,"username":"anonymous_coward_old","name":"Anonymous Coward Old","avatarUrl":"/uploads/image/43/32/4332638c11248b8ff07033c11de0759a55205facfbec85e14458245a91991faf.avif","admin":false,"moderator":false,"signatureRaw":null,"signatureCooked":null}}],"hasMore":false,"lastReadPostNumber":5,"hasUrlPostNumber":true}