AI可可AI生活

fly51fly

来自 @爱可可-爱生活 的第一手AI快报,用最通俗的语言,聊最前沿的人工智能科研进展~ #人工智能# #科技前沿#

  1. 11小时前

    [人人能懂] 本质洞察、内在罗盘与认知多样性

    我们都希望学得更聪明,但到底怎样才算“聪明”?本期我们就从几篇最新论文出发,看看AI是如何被教导着实现真正的“开窍”:它要如何学会看透不同知识表象下的本质,如何为自己打造一个用于自我提升的“进度条”,又是如何从只追求唯一的最优解,到学会欣赏整个“高分区”的所有好答案。这些AI的“内功心法”,或许正是我们自我成长的关键钥匙,让我们一探究竟! 00:00:32 AI 学习的“升维”之路:从“对答案”到“懂原理” 00:05:32 机器人的“开窍”秘诀:从抄作业到上补习班 00:11:18 AI训练的“内功心法”:当数据成了稀缺品 00:16:59 AI的“开窍”心法:从单打冠军到全能高手 00:21:36 从一锅粥里,尝出每一粒米的味道 本期介绍的几篇论文: [CL] LLM-JEPA: Large Language Models Meet Joint Embedding Predictive Architectures   [Atlassian & NYU & Brown University]   https://arxiv.org/abs/2509.142   --- [LG] Self-Improving Embodied Foundation Models   [Google DeepMind & Generalist AI]   https://arxiv.org/abs/2509.15155   --- [LG] Pre-training under infinite compute   [Stanford University]   https://arxiv.org/abs/2509.14786   --- [LG] FlowRL: Matching Reward Distributions for LLM Reasoning   [Shanghai Jiao Tong University & Renmin University of China & Microsoft Research]   https://arxiv.org/abs/2509.15207   --- [LG] Optimal Learning from Label Proportions with General Loss Functions   [Google]   https://arxiv.org/abs/2509.15145

    27 分钟
  2. 1天前

    [人人能懂] 从熔炼答案、系统权衡到逻辑自洽

    有没有想过,AI高手不是靠找答案,而是靠熔炼所有错误尝试,“创造”出全新答案来当自己的老师?本期节目,我们将揭秘AI如何完成这种不可思议的“自我修炼”,甚至在想象的梦境中为自己安排一套动态升级的“学习课程表”。我们还会一起探讨,如何为AI的“精准手术”建立一套体检标准以防“副作用”,并教会它在“信心一跃”的瞬间果断停止思考,拒绝无效内耗。最后,我们将看到AI如何在一个严厉教练的指导下,学会“瞻前顾后”的严谨逻辑。准备好了吗?让我们一起探索这些最新论文中,那些让AI变得更聪明、更靠谱的成长心法。 00:00:45 AI的“自我修炼”心法:高手不是靠找答案,而是靠造答案 00:06:54 AI的“精准手术”难题:治好了头疼,会不会引发脚气? 00:12:35 AI的“梦中修炼”法:高手是在想象中自我迭代的 00:18:37 AI的“偷懒”智慧:想明白了,就别再想了 00:23:26 怎么让AI做事靠谱?教它学会“瞻前顾后” 本期介绍的几篇论文: [LG] Compute as Teacher: Turning Inference Compute Into Reference-Free Supervision   [Meta Superintelligence Labs]   https://arxiv.org/abs/2509.14234  --- [LG] SteeringControl: Holistic Evaluation of Alignment Steering in LLMs   [University of California, Santa Cruz & Washington University in St. Louis]   https://arxiv.org/abs/2509.13450  --- [LG] Imagined Autocurricula   [University College London AI Centre & University of Oxford]   https://arxiv.org/abs/2509.13341  --- [CL] Early Stopping Chain-of-thoughts in Large Language Models   [University of Delaware & Peking University]   https://arxiv.org/abs/2509.14004  --- [LG] Teaching LLMs to Plan: Logical Chain-of-Thought Instruction Tuning for Symbolic Planning   [MIT CSAIL]   https://arxiv.org/abs/2509.13351

    30 分钟
  3. 2天前

    [人人能懂] 从任务分解、元认知到精准剪枝

    有没有想过,为什么AI会像武林高手一样,有时会练伤自己的“七伤拳”,甚至在减肥时反而越减越“胖”?本期节目,我们将一同潜入AI的“思想世界”,去看看它是如何学会“四两拨千斤”的巧劲,又是如何为自己编写“武功秘籍”来高效成长的。我们还会见证科学家如何像外科医生一样,用一把“手术刀”精准切除AI的坏念头。准备好了吗?让我们从这几篇最新的论文中,发现让AI和我们自己都变得更聪明的“章法”! 00:00:35 AI的“四两拨千斤”:高手做事,不靠蛮力  00:05:28 AI的“武功秘籍”:高手是怎么炼成的? 00:10:53 AI的“外科手术刀”:如何精准“切除”一个坏念头? 00:16:16 AI的“七伤拳”:学得越多,忘得越快? 00:22:08 给AI“减肥”,为何越减越“胖”? 本期介绍的几篇论文: [LG] Small Models, Big Results: Achieving Superior Intent Extraction through Decomposition   [Google]   https://arxiv.org/abs/2509.12423   --- [LG] Metacognitive Reuse: Turning Recurring LLM Reasoning Into Concise Behaviors   [Meta]   https://arxiv.org/abs/2509.13237   --- [LG] RepIt: Representing Isolated Targets to Steer Language Models   [University of California, Santa Cruz & UC Berkeley]   https://arxiv.org/abs/2509.13281   --- [LG] RL Fine-Tuning Heals OOD Forgetting in SFT   [Polytechnique Montreal & University of Montreal & McGill University]   https://arxiv.org/abs/2509.12235   --- [LG] Reasoning Models Can be Accurately Pruned Via Chain-of-Thought Reconstruction   [LinkedIn]   https://arxiv.org/abs/2509.12464

    28 分钟
  4. 3天前

    [人人能懂] AI江湖的武林秘籍、操盘学徒与瘦身魔法

    你有没有想过,那些让人眼花缭乱的AI新方法,背后可能藏着同样的“武功心法”?我们又要如何像培养学徒一样,把AI训练成能炒股、会分析的“操盘手”,甚至给它请个“小助理”来把工作效率提升七倍?本期节目,我们就从几篇最新的论文出发,聊聊如何给AI“瘦身”、“加速”,并看清它究竟是全能学霸,还是个连数数都会搞错的“偏科生”。 00:00:30 AI武林秘籍:天下武功,同出一门? 00:06:02 AI炒股机器人进化论:从“学徒”到“操盘手” 00:11:49 AI模型的“瘦身”魔法:让老模型焕发新生 00:16:57 AI画画慢?给它请个“小助理” 00:22:42 给AI当“统计学老师”:学霸还是偏科生? 本期介绍的几篇论文: [LG] Opal: An Operator Algebra View of RLHF   [Microsoft]   https://arxiv.org/abs/2509.11298   --- [LG] Trading-R1: Financial Trading with LLM Reasoning via Reinforcement Learning   [University of California, Los Angeles & University of Washington]   https://arxiv.org/abs/2509.11420   --- [LG] PHLoRA: data-free Post-hoc Low-Rank Adapter extraction from full-rank checkpoint   [Amazon AGI & EdgeRunner AI]   https://arxiv.org/abs/2509.10971   --- [LG] SpeCa: Accelerating Diffusion Transformers with Speculative Feature Caching   [Shanghai Jiao Tong University & The Hong Kong University of Science and Technology]   https://arxiv.org/abs/2509.11628   --- [CL] Reasoning Under Uncertainty: Exploring Probabilistic Reasoning Capabilities of LLMs   [University of Maryland]   https://arxiv.org/abs/2509.10739

    29 分钟

关于

来自 @爱可可-爱生活 的第一手AI快报,用最通俗的语言,聊最前沿的人工智能科研进展~ #人工智能# #科技前沿#

你可能还喜欢