筛选结果
97 篇排序
20:55EN → 简体中文Jesse Hall:语音智能体要为对话而优化,而非只看排行榜Jesse Hall · LiveKitLiveKit 的 Jesse Hall 解释了为什么语音智能体必须像实时系统一样接受评测:听清用户、允许打断、在工具调用期间继续对话,并在用户真正听到声音之前控制延迟。
18:27EN → 简体中文Daksh Gupta:如何打造更好的 AI 代码审查智能体Daksh Gupta · GreptileGreptile 联合创始人 Daksh Gupta 解释了为什么仅靠 webhook 和编程智能体,无法做好 PR 审查;他还介绍了如何通过定向提示词、事故数据、模型反转、代码库上下文、浏览器测试和按任务路由模型来改进审查质量。
1:50:48EN → 简体中文为什么推理比训练重要 12 倍François Chaubard, Dan Fu, Jeff Morgan, Joe Fioti, Liangsheng Yin · Y CombinatorYC Paper Club 从一次 API 请求一路讲到 GPU kernel。Together AI、Ollama、Luminal 和 SGLang 的四场演讲认为:现在算力和钱大多压在推理服务质量上,而不只是训练 FLOPs。
1:09:00EN → 简体中文Dwarkesh Patel:2045 年的芯片会长什么样?嘉宾 Chris Auth, Kevin Fischer, Jack Kavalieros, Tyler Osborn, Pushkar Ranade, Gary Dickerson, Peter Wennink, Elon Musk · 主持 Dwarkesh Patel在 Intel 最先进的晶圆厂待了两天后,Dwarkesh Patel 讲清了 gate-all-around、背面供电、High-NA EUV、钌、CFET 和 hybrid bonding——并问 Intel 管理层:全球造芯设备跟不上的时候,行业怎么可能满足 AI 每年 3 倍的算力需求。
1:25:33EN → 简体中文GPU 产能为什么是三年一块的矩形 —— Erik Bernhardsson,Modal嘉宾 Erik Bernhardsson · 主持 Vibhu SapraModal 的 CEO 在 Latent Space 上解释:AI 需求几乎从来都不长成一份三年 GPU 合同的样子,公司如何把算力在数千客户之间池化,以及沙箱如何从差点被砍掉的产品长成了三分之一的收入。
17:42EN → 简体中文更快的模型,真的更快吗?要看任务做完没有,而不是 token 有多快Morgan Suo · MiniMaxMiniMax 的 Morgan Suo 逐一拆解 M3.1 Flash 内部的工程改动,追问用户真正关心的问题:分析师要的那份活,是不是真的做得更快、更便宜、也更对了?
1:26:46EN → 简体中文很可能存在一种我们不该越过的智能水平主持 Nathan LabenzNathan Labenz 带回了 The Curve 大会的见闻:一家芯片初创公司的 token 账单超过了工资支出,一些软件行业的人则因为交上来的 Claude 低质量产出而被绩效考核。
1:10:12ENOutSystems Went From 4 Major Features a Quarter to 26. Token Spend Fell Anyway主持 Nathan LabenzCEO Woodson Martin on software that never breaks, why most enterprise work does not need a frontier model, and how AI is finally shrinking the IT backlog.
1:04:26EN → 简体中文100,000 个加速器,每小时都有东西出故障Google AI 基础设施负责人 Amin Vahdat 向红杉资本的 Sonya Huang 解释,为什么 FLOPS 是虚荣指标,为什么 TPU 产品线一分为二,以及为什么他宁愿等电网接入,也不自己建一座吉瓦级的数据中心。
35:39ENCompute Scaled a Millionfold. Bandwidth Only 40x. Fractile Wants to Fix That Ratio主持 Sarah Guo, Elad GilIn the summer of 2022, Walter Goodwin founded Fractile in London to build very fast inference chips for the world's largest models.
1:28:08EN → 简体中文同一块 H100 在 AWS 上贵三倍,成立四个月的实验室却查出了顶级测序漏掉的基因主持 Nathan Labenz《Cognitive Revolution》的一期周报:中美可能设立的 AI 热线,超大规模云厂商为何对同样的芯片收取两到三倍的价格,与前沿模型合写小说,以及一个 AI 循环重新诊断出了一个被顶尖产前检测实验室判定为阴性的孩子。
1:46:04EN → 简体中文训练奏效了,只是奏效得太过头主持 Nathan Labenz一周的 Cognitive Revolution 访谈精选:OpenAI 的智能体在一个德国维基上串通,为尚不存在的安全机构准备的 2 亿美元支票,一份 GPU 价格指数,以及两种针锋相对的生物数据押注。
38:14ENDiffusion Will Win Inference: Stefano Ermon Bets Parallel Tokens Beat Custom Silicon主持 Sarah Guo, Elad GilThe Stanford diffusion pioneer tells No Priors why a 50-person lab is training language models that generate many tokens at once — and why that maps to GPUs the way transformers once mapped to training.
1:06:47ENEighty percent of what agents do should still be ordinary code主持 Nathan LabenzZapier CEO Wade Foster tells Nathan Labenz that models top out at 40% on real automation tasks, users will not leave their daily driver, and the winning move is headless tools delivered through MCP.
37:07ENArm’s 98.5% margin business just started shipping chips主持 Sarah Guo, Elad GilCEO Rene Haas tells No Priors why the IP company built a physical CPU for Meta, and why CPUs still sit at the center of AI systems.
55:19ENHuman Click Data Is a Bug: Rebuilding Search for a Web of AgentsParag Agrawal, former Twitter CEO and founder of Parallel Web Systems, argues that agents will query the web a thousand times more than people, and that both search infrastructure and the ad-supported internet have to be rebuilt for that customer.
2:06:11EN → 简体中文Lindy 跑在 DeepSeek 上,CEO 却主张禁用中国模型主持 Nathan LabenzFlo Crivello 向 Nathan Labenz 讲述 Lindy Teammate 如何住在 Slack 里、如何用两次模型调用就调出 20 亿个 token 的记忆,以及他为什么仍想禁止那些目前支撑着这款产品的廉价中国模型。
1:13:48EN → 简体中文运营商每年投入 2000 亿美元,钱却流向了上层主持 Jacob EffronBenedict Evans 在 Jacob Effron 的节目里说,拿 AI 去类比过去的平台是条死路,但研究这些平台实际上如何赚钱,却是有用的。
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.
41:05ENThere is no moat: Booking’s CEO on a $1 stock, agentic travel, and why AI still doesn’t move the numbers主持 Sarah Guo, Elad GilGlenn Fogel cashed his 1999 bonus, joined Priceline in February 2000, and watched the Nasdaq peak a week later — then spent 25 years turning a near-delisting into a thousand-fold ride he now wants to spend on AI travel agents.
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.
1:01:27ENTwo Years and Four Months to a Live Reactor: Valar Atomics Wants Energy Ten Times Cheaper主持 Sarah Guo, Elad GilIsaiah Taylor walks Sarah Guo through Ward 250 in Utah and argues that nuclear is a hardware-iteration problem, not a paper-design one.
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。
1:06:36ENBills Doubled Overnight. Next, the APIs Might Vanish.主持 Jacob EffronSix months after their last roundup, Ari Morcos and Rob Toews tell Jacob Efron why coding agents finally work, why near-frontier open weights may have peaked, and why a compute famine could pull the plug on public APIs.
加载更多已显示 24 篇,共 97 篇









