筛选结果
20 篇排序
2:00:51EN → 简体中文在每个 Token 处重新采样,就能看见模型怎样做决定主持 Nathan LabenzGoodfire 的 Eric Bigelow 谈分叉路径、上下文学习,以及为什么越来越难信任思维链(chain-of-thought)监控。
51:55EN → 简体中文AI 解不出的最后一道 IMO 题Grant Sanderson · 3Blue1BrownGrant Sanderson 带你走一遍 2025 年 IMO 第 6 题——前沿模型没做出来的铺砖谜题——以及要证明一种构造最优,需要怎样的视觉耐心。
1:34:29EN → 简体中文三个 IMO 金牌级系统漏掉了同一道题,Harmonic 想让数学能够编译主持 Nathan LabenzVlad Tenev 和 Tudor Achim 打造了 Aristotle,让它输出内核能够检验的 Lean 证明。他们认为,这正是数学超级智能保持可信的方式。
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.
1:14:11EN → 简体中文推理曲线其实一直很平滑:Gemini 2.5 的 Jack Rae 谈思考、记忆与 AGI主持 Nathan Labenz谷歌 DeepMind 的思考与推理负责人解释:为什么推理用的强化学习(RL)显得来得突然,为什么 Gemini 会展示原始思维链,以及为什么一百万 token 的窗口仍然不等于记忆。
1:10:23EN → 简体中文每美元一百万个 Token:为什么 Noam Shazeer 认为推理算力(inference compute)仍未被充分挖掘主持 Jacob EffronTransformer 共同发明人 Noam Shazeer 与 Gemini 研究员 Jack Rae 讲述测试时算力(test-time compute)、Ilya Sutskever 对通用人工智能(AGI)的批评,以及一个四岁孩子如今能给老师讲蜥蜴的故事。
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。
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。
1:33:01ENRare Diseases Are Common: How a Harvard Gene Hunter Uses o1-pro on the Cases Sequencing Already Missed主持 Nathan LabenzDr. Catherine Brownstein of Boston Children's Hospital explains why cheap genomes made her job harder, why data sharing still beats any model, and where ChatGPT Pro actually gives her hours back.
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。
1:07:32ENo1 Is a New Generation — and the Day AGI Arrives Will Feel Boring主持 Jacob EffronOpenAI’s former chief research officer Bob McGrew on why o1 is a 100x leap, why agents stall on reliability, and why AGI may land as office tedium rather than a single dramatic moment.
48:18ENNoam Brown: There's No Hard Wall—Just Eight Orders of Magnitude of Headroom主持 Jacob EffronA ChatGPT query costs about a penny. The problems society cares about most, he says, are worth a million dollars.
1:06:48EN → 简体中文o1 又慢又难用,Percy Liang 仍然认为它是个转折点主持 Jacob Effron斯坦福大学的 Percy Liang 把 OpenAI 的 o1 一分为二来看:作为产品令人失望,作为研究信号,却可能是这几年最重要的一个。
1:56:03EN → 简体中文人类接近满分,最强 AI 仍只拿到 39%主持 Nathan LabenzZapier 联合创始人 Mike Knoop 和 François Chollet 为 ARC-AGI 拿出了 100 万美元奖金。ARC-AGI 是 2019 年提出的、考察样本效率的智能测试,在每题一美分的算力预算下,语言模型至今仍无法破解。
1:22:59ENIf Your Agent Is Right 80% of the Time, Ten Steps Leave You With 10%主持 Nathan LabenzImbue CTO Josh Albrecht tells Nathan Labenz how a 27-person, pre-product research company is putting $200 million and 10,000 H100s toward agents that can actually finish a long job.
2:00:19EN1.5 Million GPT-4 Fairy Tales Let a 33-Million-Parameter Model Out-Reason GPT-2主持 Nathan LabenzTry this fill-in: Lily likes cats and dogs. She asked her mom for a dog. Mom said no, so she asked for a…
1:45:49ENTrust at Scale: Elicit's Founders on Reasoning Languages, World Models and Automating the Company主持 Nathan LabenzAndreas Stuhlmüller and Jungwon Byun explain how Elicit keeps frontier models checkable for high-stakes science, and where they think AI reasoning is heading.
58:57ENWhen the Model Learns to Look Around: Apollo Research on Testing o1 for Scheming主持 Nathan LabenzApollo Research founder Marius Hobbhahn explains how his team probed OpenAI's o1 for deception, what it found, and why the trajectory worries him more than the model itself.
57:25ENFrom RAG to Reasoning: Why the Future of AI Is a System, Not a Model主持 Jacob EffronDouwe Kiela explains what OpenAI's o1 reveals about test-time compute, why enterprise AI must be specialized, and where post-training and evaluations are heading.
1:03:10ENPause Tokens, Analogical Prompts and Ring Attention: A Tour of the Transformer's Next Upgrades主持 Nathan LabenzNathan Labenz and Trey Kollmer walk through five recent papers on pausing, backspacing, prompting, streaming and long-context training, and sketch how they might combine into a more capable model.









