筛选结果
117 篇排序
56:31ENPost-Training Wins Inference: Why Owning the Weights Beats Betting on One Lab主持 Jacob EffronApplied Compute CEO Yash Patil argues that control, private evals, and post-training on the biggest workloads matter more than fearing that frontier labs will turn malicious.
1:56:40EN → 简体中文周期表早已在权重里主持 Nathan LabenzGoodfire 首席技术官 Dan Balsam 向 Nathan Labenz 讲述,概念流形、预测性数据调试,以及每月 1000 美元的研究智能体,如何把可解释性变成一款产品。
2:06:41EN → 简体中文把 J-space 清零,模型的隐藏目标在第一个 Token 上就亮了主持 Nathan Labenz《The Cognitive Revolution》的每周精选:Anthropic 的全局工作空间(global workspace)论文、AI 超级预测者、开放的视频世界模型,以及一款把推理当作数据搬运问题来解的芯片。
50:57ENA Trillion Dollars of GPUs Will Be Worth $100 Billion in Five Years, Says LSTM's Jürgen Schmidhuber主持 Jacob EffronThe AI pioneer tells Unsupervised Learning's Jacob Effron why he is wildly optimistic about the technology and deeply pessimistic about the model companies buying the chips.
1:47:38EN19 Neurons Can Drive a Car. Liquid AI Wants Foundation Models on Every Chip主持 Nathan LabenzRamin Hasani traces liquid neural networks from a worm's 300-cell brain to a 600-megabyte model that will talk from inside a Mercedes, and explains why architecture search now beats architecture avengers.
2:59:44EN → 简体中文两门濒危语言难倒了 Transformer,嵌套学习给模型加上“睡眠”按钮主持 Nathan LabenzGoogle 研究员 Ali Behrouz 向 Nathan Labenz 讲解,语言模型要获得真正的持续学习,不能只堆叠层数,还要堆叠更新频率,并加入一个离线的睡眠阶段。
1:13:33EN → 简体中文一万亿个 Token 之后,概念终于落地:Łukasz Kaiser 谈 Transformer 仍然缺什么主持 Jacob Effron《Attention Is All You Need》合著者在 Unsupervised Learning 节目中表示,Transformer 大概率会继续胜出,但它们学习的方式仍然与人不同。
59:41ENThe GPT moment for video hasn’t arrived, but AGI may already be here主持 Jacob EffronA day after Google I/O 2026, Gemini co-lead Oriol Vinyals argues that world models, file-system memory, and infinite-complexity RL environments will define the next era of frontier models.
1:21:56EN → 简体中文Yann LeCun 离开 Meta:大语言模型通向不了智能主持 Jacob Effron这位图灵奖得主对 Jacob Effron 讲述了他为何从 Meta 分拆出 AMI Labs,为何类 JEPA 的世界模型胜过像素生成器,以及他为何认为 Hinton 在 2023 年改变了看法。
1:48:43EN → 简体中文奖励从 3 涨到 180,Hacker News 上的标题全是谷歌裁员主持 Nathan LabenzCoreWeave 的 Kyle Corbitt 带 Nathan Labenz 讲 GRPO、评分标准(rubric)、强化学习环境,以及为什么大多数奖励作弊(reward hacking)只要知道去找,就很容易发现。
58:46EN → 简体中文OpenAI 首席科学家给出两个时间点:9 月前做到研究实习生,2028 年前做到自动化研究员主持 Jacob EffronJakub Pachocki 在 Jacob Effron 的节目上解释,为什么这些时间表依然成立,为什么持续学习并不是弯路,以及为什么在产品中展示思维链,会让训练信号与监控相互对抗。
1:49:53EN → 简体中文别跟反向传播硬碰硬:Goodfire 押注设计模型学什么主持 Nathan LabenzDan Balsam 和 Tom McGrath 再次做客 The Cognitive Revolution,带来了 1.5 亿美元(150 million)的 B 轮融资、一条名为“意向性设计”(intentional design)的新研究支柱,以及一条给想要引导训练的人的规则:别再试图用蛮力压过梯度下降。
1:34:29ENThree IMO Golds Missed the Same Problem. Harmonic Wants Math to Compile主持 Nathan LabenzVlad Tenev and Tudor Achim built Aristotle to emit Lean proofs a kernel can check—and they argue that is how mathematical superintelligence stays trustworthy.
1:04:51EN125 million cells took a decade. The next billion takes months主持 Nathan LabenzMark Zuckerberg and Priscilla Chan on pairing a frontier biology lab with a frontier AI lab so models train on data that was designed for them.
4:25:13EN → 简体中文架构几乎没变,变的是数据、成本和开放权重主持 Lex FridmanNathan Lambert 和 Sebastian Raschka 带 Lex Fridman 回顾了这一年:训练方法和数据集的变化,远大于模型设计的变化。展望 2026 年,争论的焦点似乎是推理成本、智能体,以及真正开放的模型该归谁。
1:03:00ENA Model That Solves New Math Problems Every Week Still Isn't AGI主持 Jacob EffronFormer OpenAI research VP Jerry Tworek on why scaling still works, why models go hopeless when they fail, and why he left to chase continual learning.
2:25:10EN → 简体中文微调没有死,却把一个乐于助人的模型变得邪恶了主持 Nathan LabenzNathan Labenz 谈为什么大多数工作改由提示词完成而不再微调,涌现式错位(emergent misalignment)到底揭示了什么,以及他如何看待工作、全民基本收入(UBI)和通用人工智能(AGI)。
1:55:12EN → 简体中文28,000 份文献、一个 30 亿参数的模型、22 个人:天主教 AI 的技术账本主持 Nathan LabenzLongbeard 首席执行官 Matthew Harvey Sanders 在 The Cognitive Revolution 节目中讲述,一支使命驱动的团队如何让模型与天主教训导对齐,以及为什么微调还不够用。
1:29:30EN → 简体中文2500 个化合物,51% 的命中率:一个小型图网络如何找到新抗生素主持 Nathan LabenzMIT 的 Jim Collins 向 Nathan Labenz 讲述:一个没有正式经费的 2018 年项目,加上二元生长抑制标签和 Chemprop 图网络,如何从实验室根本无法人工筛查的化学空间中找出新抗生素。
1:01:49ENChatGPT's Co-Creator Left to Put Nature in the Reward Loop主持 Nathan LabenzLiam Fedus and Ekin Dogus Cubuk on why language models still fail at physics, why literature data is too noisy to train on, and why Periodic Labs is aiming at a 200K superconductor.
1:22:07ENEvery payment is a frame in a movie. Stripe’s foundation model took card-testing detection from 59% to 97%主持 Nathan LabenzEmily Sands, Stripe’s head of data and AI, explains how a domain-specific payments model, shared embeddings, and mixed rules turn a $1.4 trillion network into compounding advantage.
2:02:17ENSAEs Find the Nouns. Parameter Decomposition Looks for the Verbs主持 Nathan LabenzGoodfire's Lee Sharkey argues that sparse autoencoders recover a network's variables, not the algorithms those variables run, and shows how stochastic parameter decomposition tries to recover the computations themselves.
57:52ENThey Aimed for 1%. The Lab Hit 20%. Chai Discovery Moved Antibody Design Onto the Computer.主持 Jacob EffronJoshua Meier on three waves of AI drug discovery, why Chai-2 behaves like Photoshop for molecules, and why a 1% hit rate was supposed to be enough to change the industry.
23:39ZH李沐:给文本大模型加入一千万小时语音,开源 Higgs Audio李沐 · Boson AI李沐介绍团队如何在文本大语言模型的训练里加入大量语音数据,做成能听能说的 AudioLLM,并把 Higgs Audio 开源。他讲了架构取舍、token 压缩、数据从哪来,以及怎么在 3090 上跑起来。
加载更多已显示 24 篇,共 117 篇









