筛选结果
14 篇排序
1:09:44EN10,000 GB300s for Four Weeks: Reflection’s Open-Model Ledger主持 Sarah Guo, Elad GilMisha Laskin tells No Priors how Reflection trained Beam, why reinforcement learning outspent pre-training, and why the West needs its own open weights.
49:12ENThe People Writing the Transformer’s Eulogy Want a Lab That Can Take Vacation Without Themo1’s reasoning lead and a Gemini pre-training lead argue the architecture, not more scale, is now the bottleneck — and that replacing it starts with automating kernel generation.
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。
1:01:38EN13-Hour Coffee and 2x Throughput: Physical Intelligence Crosses the Deployment LineKarol Hausman and Tobi Springenberg tell Sequoia’s Training Data how a general robot brain, plus real-world RL, turned coffee, laundry, and chocolate boxes into jobs that actually run.
48:51EN → 简体中文OpenAI 前研究负责人:通往 AGI 的关键突破已经齐了Bob McGrew 在红杉资本的 Training Data 节目中说,2025 年是推理之年,智能体的价格会贴近算力成本,而到 2035 年,技术栈上不会再添加什么根本性的东西。
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.
5:06:18EN → 简体中文DeepSeek 宣称训练成本仅 500 万美元,背后是一场关于算力的五年赌注主持 Lex FridmanDylan Patel 与 Nathan Lambert 带着 Lex Fridman 梳理了 DeepSeek-V3 与 R1、出口管制、台积电、超大集群,以及在 AI 竞赛中究竟谁能赚到钱。
1:04:54EN → 简体中文他拿公司股权赌 Transformer 能撑下去,然后为 12000 家公司画出一张 AI 地图主持 Jacob EffronDatabricks 首席 AI 科学家 Jonathan Frankle 告诉企业,什么时候该用提示词、什么时候该微调或预训练,以及他为什么仍然不肯预测下一代架构。
2:24:18EN → 简体中文MMLU 在训练后期才脱离随机水平,loss 只是一个警报主持 Nathan LabenzZyphra 的训练负责人讲解了 Zamba 2 的混合架构、真正重要的评测,以及小型实验室仍需重新摸索的那些技术诀窍。
46:09ENOriol Vinyals: The Days of Burning 90% of Compute on Pretraining Are Numbered主持 Sarah Guo, Elad GilGoogle DeepMind’s VP of research on Gemini, long context, the coming shift from pretraining to RL, and why AGI may never have a finish line.
1:01:39ENGPT-5 Wants 100 Trillion Tokens. Does Earth Have Enough?主持 Nathan LabenzTraining GPT-3 took about a trillion tokens; GPT-4 took about ten trillion. The next jump is a data question as much as a compute one.
1:33:56ENThe AI Scouting Report: 90 Minutes From GPT-4's Exam Scores to Whether Models Understand主持 Nathan LabenzNathan Labenz spends a solo episode walking through what general-purpose models can do, how they are trained, how capabilities unlock, and whether they understand anything at all.
59:28ENFrom $500,000 to $50,000: MosaicML Made Training Your Own Model a Priced Product主持 Nathan LabenzJonathan Frankle and Abhi Venigalla explain who is training custom LLMs, why extraction beats chat, and how ALiBi stretches a 65k context window.
2:23:56EN → 简体中文GPT-4 是非常早期的 AI,历史书会记下的是 ChatGPT嘉宾 Sam Altman · 主持 Lex FridmanGPT-4 发布六天后,Sam Altman 对 Lex Fridman 解释了为什么可用性胜过单纯的规模,为什么对齐与能力并非相反的方向,以及为什么 OpenAI 要趁风险还低的时候推进部署。









