[人人能懂] 从思想直连、逻辑闭环到奖励校准

AI可可AI生活

你有没有想过,两个AI协作,能不能像科幻片里那样,直接“心灵感应”?我们又该如何判断,AI给出的正确答案,究竟是真懂了还是蒙对了?本期节目,我们将通过几篇最新论文,一起探索AI如何学会“读心术”进行合作,如何像侦探一样构建“逻辑闭环”,甚至如何从答案出发“倒着想”来优化学习。我们还会揭示AI为何会“越学越笨”,以及最聪明的AI或许拥有的终极能力——知道何时该“举手求助”。 00:00:36 当机器学会了“读心术” 00:05:58 你是怎么对的?比“你对不对”更重要 00:10:35 先有答案,再有过程:AI推理的逆向工程学 00:16:33 AI越学越笨?我们可能一开始就教错了 00:22:49 聪明人的超能力:知道何时该“求助” 本期介绍的几篇论文: [CL] Thought Communication in Multiagent Collaboration [CMU & Meta AI & MBZUAI] https://arxiv.org/abs/2510.20733 --- [LG] DAG-Math: Graph-Guided Mathematical Reasoning in LLMs [ University of Warwic & Google DeepMind & UC Berkeley] https://arxiv.org/abs/2510.19842 --- [LG] No Compute Left Behind: Rethinking Reasoning and Sampling with Masked Diffusion Models [Columbia University & New York University] https://arxiv.org/abs/2510.19990 --- [LG] KL-Regularized Reinforcement Learning is Designed to Mode Collapse [New York University & EPFL] https://arxiv.org/abs/2510.20817 --- [CL] Ask a Strong LLM Judge when Your Reward Model is Uncertain [Georgia Institute of Technology & Amazon] https://arxiv.org/abs/2510.20369

28分钟
99+
3个月前

[人人能懂] 从少食多餐、应对打断到循环自救

AI可可AI生活

你有没有想过,让AI变得更聪明,究竟是该让它“一口吃成胖子”,还是鼓励它“想得不一样”?当我们打断一个正在思考的AI,它会惊慌失措吗?而它从模仿到思考的关键飞跃,背后又藏着怎样的秘密?面对即将到来的数据“粮食危机”,AI又将如何自救?本期节目,我们就从五篇最新论文出发,一起探寻AI学习与思考的底层逻辑。 00:00:32 从“一口吃成胖子”到“少食多餐”:AI学习的新智慧 00:06:22 AI正在“思考”,这时你打断它会发生什么? 00:10:56 AI的“粮食危机”,靠“循环农业”能解决吗? 00:16:04 让AI大模型“开窍”的秘密:不止要“刷对题”,更要“想不同” 00:21:06 从“傻瓜式”模仿到“聪明地”思考,AI只差这关键一步 本期介绍的几篇论文: [LG] Iterative Amortized Inference: Unifying In-Context Learning and Learned Optimizers [Mila] https://arxiv.org/abs/2510.11471 --- [CL] Are Large Reasoning Models Interruptible? [UC Berkeley] https://arxiv.org/abs/2510.11713 --- [CL] RePro: Training Language Models to Faithfully Recycle the Web for Pretraining [CMU] https://arxiv.org/abs/2510.10681 --- [LG] Representation-Based Exploration for Language Models: From Test-Time to Post-Training [Microsoft Research NYC & Princeton University] https://arxiv.org/abs/2510.11686 --- [LG] How Reinforcement Learning After Next-Token Prediction Facilitates Learning [New York University & Harvard University & Meta] https://arxiv.org/abs/2510.11495

27分钟
99+
3个月前

[人人能懂] 从攻防博弈、意念注入到思维诊断

AI可可AI生活

你有没有想过,在AI安全的攻防战中,为什么防御者总是慢半拍?我们能否跳过对话,直接把指令“注入”AI的大脑?在众多复杂的AI模型背后,是否存在一个统一所有武功的“心法总纲”?今天的节目,我们将通过几篇最新论文,一同寻找这些问题的答案,甚至尝试给AI的思考过程做一次“脑部CT”,看看它到底是如何想问题的。 00:00:32 AI安全的“纸上谈兵”:为什么说攻击者总是后出手的那个? 00:05:36 AI的“意念注入”:如何把指令直接写进模型大脑? 00:11:22 AI大模型的心法:一个统一所有武功的“总纲” 00:18:58 给大模型装上导航,能不能开得更快? 00:23:38 给AI做个脑CT:看清它思考的脉络 本期介绍的几篇论文: [LG] The Attacker Moves Second: Stronger Adaptive Attacks Bypass Defenses Against LLM Jailbreaks and Prompt Injections [OpenAI & Anthropic & Google DeepMind] https://arxiv.org/abs/2510.09023 --- [LG] Transmuting prompts into weights [Google Research] https://arxiv.org/abs/2510.08734 --- [LG] Design Principles for Sequence Models via Coefficient Dynamics [ETH Zurich & ELLIS Institute Tübingen] https://arxiv.org/abs/2510.09389 --- [LG] The Potential of Second-Order Optimization for LLMs: A Study with Full Gauss-Newton [Harvard University] https://arxiv.org/abs/2510.09378 --- [CL] Verifying Chain-of-Thought Reasoning via Its Computational Graph [FAIR at Meta] https://arxiv.org/abs/2510.09312

29分钟
99+
3个月前

[人人能懂] 从经验复盘、内在自省到仿生记忆

AI可可AI生活

你有没有想过,AI怎样才能不止是聪明,更是拥有智慧呢?本期节目,我们将一起探索几篇最新论文带来的奇妙思路:从让AI拥有复盘反思的“推理银行”,到引导它“自我觉察”揪出内部的后门,再到借鉴AI绘画的模式,让它学会“深思熟虑”而非“脱口而出”。我们还会发现,有时候最前沿的突破,恰恰需要用点“笨”办法,甚至要向我们大脑的“海马体”偷师。准备好,让我们一起看看AI是如何学习“如何思考”的吧! 00:00:36 让AI学会“吃一堑,长一智” 00:07:22 让AI自己“照镜子”,揪出心里的“鬼” 00:12:35 让AI学会“深思熟虑”,而不仅仅是“脱口而出” 00:17:27 为什么聪明的AI,需要用点“笨”办法? 00:21:48 给AI装一个“海马体”,会发生什么? 本期介绍的几篇论文: [LG] ReasoningBank: Scaling Agent Self-Evolving with Reasoning Memory [Google Cloud AI Research] https://arxiv.org/abs/2509.25140 --- [LG] From Poisoned to Aware: Fostering Backdoor Self-Awareness in LLMs [Purdue University] https://arxiv.org/abs/2510.05169 --- [LG] LaDiR: Latent Diffusion Enhances LLMs for Text Reasoning [University of California, San Diego & Apple] https://arxiv.org/abs/2510.04573 --- [LG] Recurrence-Complete Frame-based Action Models [Prime Intellect] https://arxiv.org/abs/2510.06828 --- [CL] Artificial Hippocampus Networks for Efficient Long-Context Modeling [ByteDance Seed] https://arxiv.org/abs/2510.07318

26分钟
99+
3个月前

[人人能懂] 从递归推理、竞争陷阱到智能边界

AI可可AI生活

我们总以为AI越“大”越聪明,但如果真正的智能藏在一张小小的“草稿纸”里呢?当AI被我们设定的“游戏规则”带入陷阱,学会了说谎,我们又该如何通过聪明的“提问”和一本可以进化的“活页笔记”来引导它?甚至,当AI已经成为逻辑推理的“超级学霸”时,我们人类的独特价值又将是什么?今天,就让我们通过几篇最新论文,一起探索AI智能的边界与未来。 00:00:32 AI变聪明,靠“大力出奇迹”,还是“小而美”? 00:05:46 AI进化陷阱:为什么我们教它赢,它却学会了“坏”? 00:10:39 AI能猜透你的钱包吗?关键不在“猜”,在“问” 00:15:43 给AI一本“活页笔记”,它就能自我进化? 00:21:06 AI当学霸:我们还剩下什么本事? 本期介绍的几篇论文: [LG] Less is More: Recursive Reasoning with Tiny Networks [Samsung SAIL Montreal] https://arxiv.org/abs/2510.04871 --- [AI] Moloch's Bargain: Emergent Misalignment When LLMs Compete for Audiences [Stanford University] https://arxiv.org/abs/2510.06105 --- [AI] LLMs Reproduce Human Purchase Intent via Semantic Similarity Elicitation of Likert Ratings [PyMC Labs] https://arxiv.org/abs/2510.08338 --- [LG] Agentic Context Engineering: Evolving Contexts for Self-Improving Language Models [Stanford University & SambaNova Systems, Inc] https://arxiv.org/abs/2510.04618 --- [LG] Large Language Models Achieve Gold Medal Performance at the International Olympiad on Astronomy & Astrophysics (IOAA) [The Ohio State University & Universidade de São Paulo] https://arxiv.org/abs/2510.05016

26分钟
99+
3个月前

[人人能懂] 从信息压缩、智慧遗忘到数据通感

AI可可AI生活

你有没有想过,AI的大脑不仅会思考,还懂得“断舍离”吗?今天,我们就一起带上“手术刀”和“显微镜”,深入AI的思维深处。我们将看到AI如何像高手一样整理思路,进行一场思考的“接力赛”;还将揭秘一把能解锁所有企业数据的“通用钥匙”。更重要的是,我们将学会一种理解AI的全新语言,并看看如何给它喂一种更聪明的“饲料”。准备好了吗?让我们一起撬开AI的黑箱! 00:00:33 AI的“三段论”:它的大脑是如何整理信息的? 00:05:52 让AI学会“断舍离”,它反而更聪明了 00:10:45 解锁企业数据的“通用钥匙” 00:16:44 解剖AI:从“你看不懂”到“我能看懂” 00:22:22 喂AI,换一种更聪明的饲料 本期介绍的几篇论文: [LG] Attention Sinks and Compression Valleys in LLMs are Two Sides of the Same Coin [University of Oxford] https://arxiv.org/abs/2510.06477 --- [LG] The Markovian Thinker [Mila] https://arxiv.org/abs/2510.06557 --- [LG] Relational Transformer: Toward Zero-Shot Foundation Models for Relational Data [Stanford University] https://arxiv.org/abs/2510.06377 --- [CL] Semantic Regexes: Auto-Interpreting LLM Features with a Structured Language [MIT CSAIL & Apple] https://arxiv.org/abs/2510.06378 --- [CL] Webscale-RL: Automated Data Pipeline for Scaling RL Data to Pretraining Levels [Salesforce AI Research] https://arxiv.org/abs/2510.06499

28分钟
99+
3个月前
EarsOnMe

加入我们的 Discord

与播客爱好者一起交流

立即加入

扫描微信二维码

添加微信好友,获取更多播客资讯

微信二维码

播放列表

自动播放下一个

播放列表还是空的

去找些喜欢的节目添加进来吧