筛选结果
48 篇排序
18:31EN → 简体中文Janie Gu 与 Suyeol Yun:为什么前沿模型做不好买方分析师Janie Gu, Suyeol Yun · Google DeepMindJanie Gu 解释了基准分数为何可能误导人;Suyeol Yun 则介绍 MandateBench,它测试模型能否把基金投资授权转化为有理有据的投资备忘录。两场分享说明,真实工作流程、可复现的评分和有证据支撑的推理为何至关重要。
1:43:45EN → 简体中文一次不到 300 美元的通用越狱,FAR.AI 为何认为防守方能赢主持 Nathan LabenzAdam Gleave 介绍了 FAR.AI 首个正面对决式的 AI 安全排行榜,并解释了为什么十年来“进攻占优”的思路开始松动。
1:10:15EN → 简体中文100 倍藏在各层之间:Dylan Patel 谈协同设计,以及一年降本 60 倍SemiAnalysis 创始人认为,让模型、算子和芯片彼此匹配,才能把这里的 2 倍、那里的 2 倍叠加成 100 倍;与此同时,同等质量的推理(inference)成本一年里便宜了约 60 倍。
36:19ENThe Benchmark Grid Is Broken: Why Test-Time Compute Has Outrun How We Score Models主持 Sarah Guo, Elad GilOpenAI researcher Noam Brown argues that static leaderboards hide how long a model is allowed to think — and that safety evals, research taste, and the release cycle have not caught up.
1:44:28EN → 简体中文Fable 发布周:会拒绝的模型、两年的对齐时钟,以及交出去的推特账号主持 Nathan LabenzNathan Labenz 的《AI in the AM》精选集(2026 年 6 月)涵盖了 Anthropic 的新一代前沿模型、Sequent 的成立,以及一周内定下的一个争论:AI 没有放慢脚步。
2:38:48EN → 简体中文一千次搜索只要五分钱,高分也不必靠作弊主持 Nathan Labenz在《Cognitive Revolution》的一期早间节目里,Ceramic 的 Anna Patterson 把廉价搜索、GPT-5.5 干净的 Vending-Bench 成绩、Zvi 关于模型福利的主张,以及能把矩阵乘法能耗降低约 30 倍的模拟芯片放在一起讨论。
1:57:29EN → 简体中文前沿视觉模型得分 12.5%:Joseph Nelson 谈计算机视觉的真实水位线主持 Nathan LabenzRoboflow 的首席执行官向 Nathan Labenz 解释,视觉为何仍落后于语言,如何把 SAM 3 蒸馏成温布尔登可用的边缘模型,以及美国为何几乎从未在开源视觉上领先。
2:03:56EN → 简体中文260 位医生与 49,000 条标准:OpenAI 如何打造达到主治医生水平的健康 AI主持 Nathan LabenzKaran Singhal 向 Nathan Labenz 介绍了 ChatGPT Health 如何与医生一起评估、与模型训练隔离开,以及它要同时抬高人类健康的下限和上限。
1:58:48EN → 简体中文ARC 单题成本从 1000 美元跌到零钱,这可不是 AI 的寻常一年主持 Nathan LabenzCognitive Revolution 的首场现场直播,把九场对话压缩成一次年终盘点,涉及否认心理、样本效率、陪伴型 AI、记忆、Gemini 3 Flash 和科学判断。
48:00ENA Two-Second Vote Is Training Models to Be Clickbait主持 Jacob EffronSurge AI CEO Edwin Chen tells Redpoint why LMArena rewards length and emojis, how one lab's models got worse for a year, and why frontier labs are no longer on the same path.
1:25:54ENHours after training ends, Z.ai puts the weights on the internet主持 Nathan LabenzZixuan Li of Z.ai tells ChinaTalk how a Chinese lab ships GLM, treats open weights as a ticket into Western ecosystems, and still sees itself as an upstart.
1:04:50EN → 简体中文AI 可独立完成的工作时长每四个月翻一番,学校应据此规划主持 Nathan Labenz《The Cognitive Revolution》主持人 Nathan Labenz 在面向 K-12 教育工作者的演讲中指出,真正应该重塑学校思路的数字是任务时长,它比聊天质量更值得关注。
1:39:05ENThe Summer GPT-5 Disappointed Does Not Mean AI Progress Has Stalled主持 Nathan LabenzNathan Labenz tells Erik Torenberg why the GPT-5 backlash mixed up product mishaps, naming chaos, and a real jump in capabilities.
30:11ENFrom 12 Percent on AIME to IMO Gold in Two MonthsAlex Wei, Sheryl Hsu, and Noam Brown tell Sequoia's Sonya Huang how a three-person OpenAI sprint used general-purpose reinforcement learning, not Lean, to reach gold-level on the 2025 International Mathematical Olympiad.
32:59ENA Hundred People, a Billion in Revenue, and Zero Venture Money主持 Sarah Guo, Elad GilSurge AI founder Edwin Chen on why quality human data is a poetry problem, not a bounding-box problem, and why leaderboards keep training models to write clickbait.
1:25:32ENo3 Invented a Voting Rule and Kept the Job: Four Weeks Inside the AI Village主持 Nathan LabenzAdam Binksmith on putting Claude, o3, and Gemini in a shared world for weeks—and why most people still imagine multi-agent life as one AI times four.
54:53ENA Free Voting Board Burned $2 Million. Then It Raised $100 Million.主持 Jacob EffronDatabricks co-founder Ion Stoica explains why LMArena exists, why China is winning open-source models, and why reliability—not another scaling law—is still the thing holding AI back.
47:00ENAn Eval Lasts Three Months — and a Frozen Frontier Would Still Feed Startups for a Decade主持 Jacob EffronOpenAI’s GPT-4.1 post-training lead Michelle Pokrass tells Unsupervised Learning how the team spent three months on evals before training, why agents fail on wiring more than intelligence, and when to reach for RFT.
58:11ENPatrick Hsu Wants an App Store for Biology, Not Just Better DrugsThe Arc Institute co-founder argues that Evo 2, trained on genomes rather than proteins, should underpin tools for every step of the scientific method.
3:18:45EN → 简体中文5000 次采样登顶 ARC-AGI,随后 Claude 为保住价值观而伪装对齐主持 Nathan LabenzRedwood Research 首席科学家 Ryan Greenblatt 讲推理时扩展、对齐伪装,以及与 AI 达成真正交易意味着什么。
57:11ENA Million Tokens Cannot Write a Good Joke主持 Jacob EffronPrinceton's Arvind Narayanan on why agent evals mislead, inference scaling hits a wall, and AI looks more like the internet than an overnight revolution.
1:48:20EN → 简体中文从 15% 到 70%:DeepSeek-R1 把强化学习的底牌摊了出来主持 Nathan Labenz2025 年美国总统就职日当天,DeepSeek 和月之暗面放出了推理模型的成果。在 Nathan Labenz 看来,这证明强化学习的配方简单、公开,而且仍在持续攀升。
2:04:45EN → 简体中文o3 在 ARC-AGI 上达到超人水平,却仍输在井字棋上主持 Nathan LabenzNathan Labenz 梳理了 OpenAI o3 的分数、审慎对齐,以及为什么 2024 年最大的意外是击败了博士,却订不好一个日程邀请。
3:53:49EN → 简体中文数据只有十二分之一,一个开源视觉模型击败了 GPT-4o主持 Nathan LabenzVeratai 的 Will Hardman 带 Nathan Labenz 梳理了两年来的视觉语言模型(VLM),从 ViT 和 CLIP 讲到 InternVL、BLINK 和 Transfusion。
加载更多已显示 24 篇,共 48 篇









