PodDeck
← 全部标签
标签

#technical

36 集相关 · 36 集已生成

已生成

140. 对姚顺宇的4小时访谈:请允许我小疯一下!在Anthropic和Gemini训模型、技术预测、英雄主义已过去

姚顺宇 · 研究科学家 · 前 Anthropic · 现 Google DeepMind (Gemini)

张小珺·语言即世界 EP.140,对姚顺宇的 4 小时访谈节选。姚顺宇博士毕业于斯坦福 理论高能物理,2024 年半道出家加入 Anthropic 参与 Claude 3.7、4.5 的强化学习训练; 2025 年 10 月跳槽到 Google DeepMind 做 Gemini 的 ML coding / long horizon。 这期把两家 lab 的打法、coding bet 的内部信号、AI safety 的"幼稚"自我说服、 以及"个人英雄主义时代已经过去了"等小疯言论摊开讲清楚。

What rebuilding AlphaGo teaches us about self-play, RL, and future of LLMs - Eric Jang

Eric Jang · ex-VP AI @ 1X · ex-Google DeepMind Robotics

Eric Jang 在 sabbatical 期间用大约 10K 美元 + Claude Code 重新实现了 AlphaGo, 在和 Dwarkesh 的对谈里把 MCTS 拆到底——为什么它不是 credit assignment、 为什么它比当下 LLM RL 优雅得多、以及 10 层网络居然能把一个看似 intractable 的搜索问题塞进一次 forward pass。后半段是用 Opus 4.6/4.7 做自动化研究的体感: 超参搜索很强,但还不会"换条路想想看"。

137. 对洪乐潼的4小时访谈:AI for Math、把数学变成Lean、数学天书中的证明、直觉、被创造与被发现的

洪乐潼 · Axiom 创始人 & CEO

张小珺对 24 岁 Axiom 创始人洪乐潼的 4 小时长访谈。00 后华人女孩、MIT 摩根奖得主、 斯坦福数学博士辍学、A 轮 16 亿美元。访谈把 AI for Math 的来龙去脉讲透——从 1960s ATP 到 2024 年 DeepMind AlphaProof,再到 2025/12 Axiom 拿下 Putnam 满分 98.93%; 也讲透了洪乐潼的创业心智:蛮力型选手、对苦难上瘾、bet system not model、 把自己定义成 "research scientist intern" 而非 CEO。

Andrej Karpathy: From Vibe Coding to Agentic Engineering

Andrej Karpathy · AI researcher, OpenAI co-founder, ex-Tesla AI

Karpathy 在一场炉边对谈里,从"作为程序员从未如此落后"讲起:December 是 agentic 编码工作流真正开始 work 的拐点。他串起 Software 3.0(编程变成 prompting)、可验证性如何造就"锯齿状"智能、vibe coding 与 agentic engineering 的分野,以及人类仍独一无二负责的"理解"。

How to ship hardware in the AI era | Caitlin Kalinowski (Apple, Meta, OpenAI)

Caitlin Kalinowski · Hardware leader — ex-Apple (MacBook Pro/Air), ex-Meta (Quest, Rift, Orion AR), ex-OpenAI (robotics)

Caitlin Kalinowski has shipped flagship hardware at three legendary companies — the original unibody MacBook Pro and MacBook Air at Apple, the Quest / Rift / Orion AR program at Meta, and most recently OpenAI's robotics division (which she publicly resigned from over the "department of war" deal). She walks Lenny through why hardware is fundamentally different from software ("we only get to compile four or five times — ever"), the 5-layer supply-chain stack that's been outsourced over 25 years, Palmer Luckey's drones-vs-aircraft-carriers argument, the impulse equation that governs robot safety, the "memory-price meteor" coming for every physical AI startup, and the four hardware design principles she carries from Apple to Meta to OpenAI.

From Consumer Toys for Pets to War Drone — Yaroslav Azhnyuk, The Fourth Law & Noah Smith, Noahpinion Latent Space
≈ 90 min

From Consumer Toys for Pets to War Drone — Yaroslav Azhnyuk, The Fourth Law & Noah Smith, Noahpinion

Yaroslav Azhnyuk · Noah Smith · The Fourth Law / Odd Systems — founder · Noahpinion — writer

Yaroslav Azhnyuk built Pet Cube in San Francisco — "cameras that fling treats to pets" — and then, after Russia's invasion landed on the last flight into Kyiv on Feb 23, 2022, pivoted to building drone autonomy, thermal cameras, and FPV strike drones for the Ukrainian armed forces. With Noah Smith and Latent Space's Brandon, he walks through five levels of drone autonomy, eight dimensions of the autonomous battlefield, why FPVs replaced artillery as "the god of war" (~80% of frontline casualties), and the disquieting arithmetic of China's drone manufacturing scale (4M Ukrainian FPVs/year vs ~4B Chinese capacity).

138. 对罗福莉3.5小时访谈:AI范式已然巨变!OpenClaw、Agent范式很吃后训练、卡的分配、组织平权

罗福莉 · 小米大模型负责人

小米大模型负责人罗福莉的 3.5 小时深度对谈:从春节凌晨 2 点 OpenClaw 觉醒,到 MiMo V2 系列 (Pro / Omni / TTS) 的"悄无声息伏击",再到 Agent 时代后训练算力 1:1、组织扁平化、AGI 两年内可期。 当下范式已从 Chat 切到 Agent —— 1T 基座 + 后训练敏捷性是新的入场券。

Getting Humans Out of the Way: How to Work with Teams of Agents

Rob · Creator of Brumi (open-source multi-agent IDE)

Rob 是开源多 agent IDE Brumi 的作者. 这期他把"如何把人从 loop 里拿出来"的整套手艺摊开讲—— 从让 agent 截图自证 (feature walkthrough doc), 到自定义 lint 规则爆炸, 到 plan.md 替代 plan mode, 到并行 5 个 agent 挑赢家. 核心隐喻只有一句: 教 agent 怎么向上汇报.

The $15B Physical AI Company: Simulation, Autonomy OS, Neural Sim, & 1K Engineers—Applied Intuition Latent Space
≈ 70 min

The $15B Physical AI Company: Simulation, Autonomy OS, Neural Sim, & 1K Engineers—Applied Intuition

Qasar Younis & Peter Ludwig · Co-founders of Applied Intuition (CEO / CTO)

Applied Intuition 给汽车 / 卡车 / 矿车 / 农机 / 防务平台卖"让物理机器变聪明"的技术栈, 18/20 OEM 是它客户, 估值 $15B, 1000 名工程师, 现在在日本跑 L4 无人卡车. 这期把它的三个 bucket (仿真 + 操作系统 + 自动驾驶模型) 全部摊开讲, 还顺手把 "vehicles like pre-Android phones" 这条 类比、neural sim = Gaussian splatting + diffusion、onboard 才是物理 AI 真正瓶颈 这些反直觉点都串了起来.

Inside Abridge: The AI Listening to 100 Million Doctor Visits — Abridge's Janie Lee & Chai Asawa

Janie Lee & Chai Asawa · Abridge — Head of Product / Clinical Decision Support

Latent Space × Supervised Learning crossover with Janie Lee and Chai Asawa from Abridge, the AI clinical-intelligence layer with a dataset on the order of 100M medical conversations. Deep on the "air conditioning" design philosophy, the constellation-of-models architecture behind real-time in-visit guidance, why HIPAA-grade de-identification is one-way, and the hot take that PRDs are very much not dead.

The Creator of Superpowers: Why Real Agentic Engineering Beats Vibe Coding

Jesse Vincent · Creator of Superpowers (110k stars Claude Code skill kit)

Jesse Vincent —— Perl projects lead 出身、K-9 Mail 的原作者、25 年老兵 —— 把过去九个月驯服 Claude Code 的方法摊开来讲. 110k stars 的 superpowers 不是 vibe coding, 而是一套 orchestrator 架构 + 单使命 subagent 分工 + skill 系统的 agentic engineering 方法论. 这期还覆盖 Claude 删测试事件如何用一行 prompt 修好、为什么 swarm 是 2002 年的 Facebook、 以及 2028 年 GitHub 可能不存代码只存 specs 的预言.

An AI state of the union: We've passed the inflection point & dark factories are coming

Simon Willison · Open-source engineer & Django co-creator

Simon Willison (co-creator of Django, coined "prompt injection") talks with Lenny Rachitsky about the November 2025 inflection point when coding agents crossed a reliability threshold, the dark factory pattern where nobody writes or reads code, and the lethal trifecta of AI security risks.

It's 2026, and We're Still Talking Evals

Maggie Konstanty · ML Engineer · LLM Agent Evaluation Lead

Maggie Konstanty 在 MLOps.community 谈 LLM agent 评估的真实战场——为什么团队总是先发布再补 eval、 为什么 pre-prod 和 production 是"两种动物"、以及为什么所有 vendor 工具都让她最终选择自己造。 整期访谈最反直觉的 takeaway:evals 本身不难,难的是让团队对齐"什么叫好"。

Head of Claude Code: What happens after coding is solved | Boris Cherny

Boris Cherny · Head of Claude Code, Anthropic

Claude Code 一周岁。它的负责人 Boris Cherny 复盘从"内部 demo 只收到 2 个赞" 到"GitHub 4% 公开 commits、Anthropic 内部人均生产力 +200%"是怎么发生的, 并解释为什么他认为 coding 已经被解决、下一站是让模型自己想做什么、 以及怎么"为 6 个月后的模型"造产品。

OpenClaw: The Viral AI Agent that Broke the Internet - Peter Steinberger | Lex Fridman Podcast #491

Peter Steinberger · Creator of OpenClaw, founder of PSPDFKit

Peter Steinberger 讲他怎么在 1 小时内把 WhatsApp 接到 Claude Code 的 CLI, 做出了后来成为 GitHub 史上最快增长仓库的 OpenClaw。围绕这个故事展开的是 agentic engineering 的心法、self-modifying software、Moldbook 引发的 AI psychosis、改名大战、以及他从 PSPDFKit 13 年烧完到重新找回 building 乐趣 的整条弧线。最后谈到下一步可能加入 Meta 或 OpenAI。

How a Meta PM ships products without ever writing code | Zevi Arnovitz

Zevi Arnovitz · Meta PM · ex-Wix PM · non-technical vibe coder

Zevi Arnovitz 是 Meta 的 PM, 一年前在日本看了一个 YouTube 视频, 然后 从 zero 技术背景一路走到 Cursor + Claude Code, 用一套可复用的 slash-command 工作流 (create-issue → explore → plan → execute → review → peer-review → update-docs) 独自维护一个副业 app Studymate。他把不同模型拟人化 (Claude 当 CTO, Codex 是小黑屋的 hoodie coder, Gemini 是吓人但出活的 crazy scientist, Composer 是冲锋队), 让它们互相 code review "fight it out"。一句反复出现的 口号: "you'll be replaced by someone who's better at using AI than you."

State of AI in 2026: LLMs, Coding, Scaling Laws, China, Agents, GPUs, AGI | Lex Fridman Podcast #490

Sebastian Raschka, Nathan Lambert · 《Build a Large Language Model (From Scratch)》作者 + AI2 研究员 / RLHF 书作者

Lex Fridman 请来两位"一线做过模型、也写过书"的研究者做 2026 年初的 AI state-of-the-art 盘点: Sebastian Raschka 从 GPT-2 一路手撕到 Qwen3 / Gemma 3, 最擅长从架构里读故事;Nathan Lambert 是 AI2 研究员、RLHF 书作者、atom 项目 发起人,frontier 与 open-source 两边都站过。两人聊了 DeepSeek 时刻、Opus 4.5 神话、RLVR 的"假 aha"、scaling 的三个轴、AI 2027 的时间线推后、Anthropic $1.5B 和解、CUDA 的真护城河、atom project,一直到 100 年后世界的样子。

Terence Tao – How the world's top mathematician uses AI

Terence Tao · UCLA mathematician, Fields Medalist

当世最强的数学家谈他亲手用过 AI 之后的判断:想法生成的成本几乎归零, 但瓶颈搬到了 verification 这一侧;50 道 Erdős 问题被 AI 攻克之后随即陷入瓶颈; AI 像在黑暗中乱跳的机器人,能跨过低墙、爬不上悬崖;breadth × depth 才是 数学的下一步——但要先重设整个学术工作流。

Dylan Patel — The single biggest bottleneck to scaling AI compute

Dylan Patel · CEO, SemiAnalysis

SemiAnalysis CEO Dylan Patel walks through the entire AI compute supply chain, from $600B hyperscaler capex to the 3.5 EUV tools needed per gigawatt. He explains why semiconductors — not power — are the real bottleneck, and what that means for the US-China AI race, memory prices, and consumer electronics.

Quantum Mechanics Contradicts Itself (and He Proved It)

Renato Renner · Professor of Quantum Information, ETH Zurich

Professor Renato Renner of ETH Zurich explains his no-go theorem showing quantum theory contradicts itself when applied to observers who are themselves quantum systems. The conversation covers three incompatible assumptions (universality, consistency, single outcomes), a gravity-based escape route involving reference frame loops, the black hole information paradox, and why choosing an interpretation of quantum mechanics is ultimately an emotional decision.

The Genius Who Invented Reverse Mathematics

Harvey Friedman · Founder of Reverse Mathematics, Emeritus Professor at Ohio State University

Harvey Friedman — the youngest professor in recorded history (a Stanford appointment at 18) and the author of the last paper Kurt Gödel sponsored for PNAS — gives his first podcast. The conversation moves from Gödel's two often-conflated incompleteness theorems to Friedman's 60-year program to push incompleteness out of set-theoretic exile and into the kind of finite, combinatorial mathematics that working mathematicians cannot dismiss. Along the way: TREE(3), the divine consistency proof, embedded maximality, and a quiet meditation on AI as a form of immortality.

Michael Nielsen – Why aliens will have a different tech stack than us

Michael Nielsen · Research fellow, writer, quantum computing pioneer

Michael Nielsen and Dwarkesh Patel explore how scientific progress actually happens — from the messy reality of falsification to hostile verification loops that mislead for decades. They discuss why the tech tree is far vaster than we realize, why alien civilizations would develop radically different technologies, and what this means for AI-accelerated science.

The "Inverse Problem" Of Dark Matter Is Insane

Dr. Jenny Wagner · Astrophysicist, Institute of Astronomy & Astrophysics / Helsinki Institute of Physics

Dr. Jenny Wagner explains why gravitational lensing data only constrains local properties of mass distributions, making every grand dark matter map a model-driven extrapolation. She argues the inverse problem approach — reasoning from data to necessary models — could reshape cosmology and the scientific method itself.

The Theorem That Proves Science Can't Know the Universe

JB Manchak · UC Irvine Professor of Logic & Philosophy of Science

Professor JB Manchak proves that no amount of empirical data — even from every point in the universe — can determine its global structure. He introduces Heraclitus spacetimes (maximally asymmetric universes where local structure determines global structure), and draws surprising parallels between cosmic underdetermination and Zen Buddhist non-self.

139. 【Agent的综述】和苏煜聊Agent技术史、OpenClaw Moment、边界的消弭和社会的辐射

苏煜 · 俄亥俄州立大学计算机系教授 / NeoCognition 创始人 / 2025 斯隆研究奖得主

张小珺商业访谈录 #139 期:和俄亥俄州立大学教授、NeoCognition 创始人苏煜做的一次 Agent 技术综述。 从 Logical Agent (1960s-90s) → Neural Agent → Semantic Parsing → Language Agent 的演进史出发, 讨论了 OpenClaw Moment 与 ChatGPT Moment 的相似性、universal digital agent 的目标、 中美科技辐射的不同 pattern,以及 2026 年 Agent 的瓶颈和大厂们的赌注。

FFmpeg: The Incredible Technology Behind Video on the Internet | Lex Fridman Podcast #496

Jean-Baptiste Kempf & Kieran Kunhya · President of VideoLAN (VLC) & FFmpeg core contributor

Jean-Baptiste Kempf (VLC) and Kieran Kunhya (FFmpeg) take Lex through the invisible plumbing of internet video — what happens between "press play" and "see pixels," why every video codec generation is roughly 30% better and 10–100× more expensive, and why dav1d ships 240,000 hand-written lines of assembly. The conversation also covers the École Centrale origin story of VLC, the Google AI security debacle and Microsoft Teams SLA episode, two intelligence-agency backdoor requests JB refused, a death threat over dropping the PowerPC port, the CIA's Vault 7 fake-VLC build, the codec patent minefield, and JB's new ultra-low-latency project Kyber (4 ms glass-to-glass for remote robotics).

Fixing GPU Starvation in Large-Scale Distributed Training

Kashish · Uber · ML Infra · Marketplace Matching Lead

Kashish (Uber ML infra, ex-Google YouTube Ads) walks Demetrios through a Sherlock-Holmes-grade Petastorm bug—GPU cluster stuck at 15-20% utilization, six debugging steps, two layers of bottleneck, and finally a "double bottleneck" reveal: PyArrow→NumPy translation was silently eating the headroom. Plus serving's latency-vs-utilization war, the reproducibility cost of parallelism, and a live diagnosis of a friend's slow DGX Spark.

Notion's Sarah Sachs & Simon Last on Custom Agents, Evals, and the Future of Work Latent Space
≈ 84 min

Notion's Sarah Sachs & Simon Last on Custom Agents, Evals, and the Future of Work

Sarah Sachs · Simon Last · Notion — engineering manager (core AI capabilities & infra) · co-founder

Notion 把 agent 重写了 5 次,从 2022 年末 GPT-4 时代试图后台跑 assistant 一直撞到今天的 100+ tools custom agent. Sarah (engineering manager) 和 Simon (co-founder) 用 84 分钟 把"为什么这么慢"和"现在为什么终于行了"都摊开讲: progressive disclosure、SQL-light queries、 notion's last exam(主动留 30% 通过率)、为什么是 credits 而不是 tokens、为什么 manager agent 是 对 70 条通知的解、以及为什么"replacing processes"比"replacing people"更准确.

The Modern Software Engineer

Mihail Eric · ML / AI infrastructure practitioner & instructor

Mihail Eric 和 Demetrios 在 SF 录音棚里把 AI coding agent 的真实工程问题挨个摊开: junior 被 cursor 截断的训练链, Eno @ Factory 强调的 validation harness, token 计费迟早被 task 计费取代, Twitter 上 "15 个 tiled Claude Code instances" 的并行神话, 团队该变小、PM 该会提 PR, 以及 下一个 superpower 是 articulation. 全程没有 framework, 全是 day-to-day 判断, 最后一句是 "just breathe".

Moonlake: Interactive, Multimodal World Models — with Chris Manning and Fan-yun Sun

Chris Manning · Fan-yun Sun · Stanford NLP 教父 · Moonlake 联合负责人 + ex-NVIDIA Research / Moonlake co-founder

Latent Space 与 Moonlake 两位负责人 Chris Manning 与 Fan-yun Sun 的对谈。Moonlake 押的是另一条 world model 路线: 不是更大的视频生成器, 而是 symbolic 推理 + 神经渲染。 Chris 给出唯一的硬定义 — "you only actually have a world model if you can predict, given some action is taken, what is going to change" — 然后顺势公开和 Yann LeCun 撕: "Yann has never appreciated the power of language." Sun 反驳"反 bitter lesson" 的标签, 真正的问题是"what is the right abstraction level today"。Moonlake 内部其实是 两个模型: 推理模型管 causality / persistency, 而 Rie 这个 diffusion model 负责 photorealism — 他们已经把它当作 DLSS 的下一代来卖, "skins for worlds"。

Training Transformers to solve 95% failure rate of Cancer Trials — Ron Alfa & Daniel Bear, Noetik

Ron Alfa & Daniel Bear · Noetik 联合创始人 · 用 AI 把 95% 失败的癌症试验变成 matching problem

Latent Space 与 Noetik 创始人 Ron Alfa、Daniel Bear 的对谈。Ron 的核心 论点:95% 的癌症临床试验会失败,但许多"失败"的药其实有效——只是没匹配 到对的病人。Noetik 用近两年时间只在收数据 (thousands of human tumors, hundreds of millions of images), 训练一个自监督的 "virtual cell" 模型, 并发布 TARIO-2——一个 autoregressive transformer, 从每个病人都已经有的 H&E 切片预测 ~19,000 个基因的空间表达。GSK 已经签了 $50M 软件授权: 不是买药, 是买平台。

How GPT, Claude, and Gemini are actually trained and served – Reiner Pope

Reiner Pope · MatX CEO, ex-Google TPU

Reiner Pope(MatX CEO)在 Dwarkesh 的黑板课式访谈里,从几条 roofline 方程一路推到: 最优 batch size = 300 × sparsity、20ms 的"火车"调度、MoE 必须挤进一个 rack、 Ilya 为什么说 pipelining 不智、RL 时代模型被过度训练 100 倍、Gemini API 价格如何 泄露 ~2KB/token 的 KV cache 大小,以及神经网络和密码学之间的对偶关系。

Mistral: Voxtral TTS, Forge, Leanstral, & Mistral 4 — w/ Pavan Kumar Reddy & Guillaume Lample

Pavan Kumar Reddy & Guillaume Lample · Mistral · Audio Research Lead 与 Chief Scientist

Mistral 同一周内同时发了 Voxtral TTS、Forge 平台、Leanstral 形式化模型和新的 Mistral Small——这期 Latent Space 让 Pavan Kumar Reddy 和 Guillaume Lample 一次性把这些 发布背后的工程选择讲清楚。Voxtral TTS 是 3B 模型 + 自研 12.5 Hz 神经音频 codec + auto-regressive flow matching head——为了实时流式而不是 SOTA quality 选 AR 路线。 Forge 是把 Mistral 科学团队用了 2 年的 infra 直接给客户:fine-tune 后能"10x cheaper", 并在某些客户项目把一种语言从 0~1% 训到 50% 的 mix。Leanstral 看似是数学家工具, 实际是赌 long-horizon reasoning 的 transfer——Lean 的编译器是天然不可 reward-hack 的判官。最后透露下一代 RL infra 是为"6 hours to get a reward"的 trajectory 设计的。

How GPT-5 derived new results in theoretical physics and quantum gravity — Alex Lupsasca, OpenAI

Alex Lupsasca · OpenAI for Science · 黑洞理论物理学家 · 2024 New Horizons in Fundamental Physics Prize

Alex Lupsasca 是 2024 年 New Horizons in Fundamental Physics Breakthrough Prize(被称为 "Oscar for physics")的得主, 一位黑洞理论物理学家。他追踪 LLM 在科学前沿的能力已经 一年半。GPT-5 发布时 Twitter 反响 "lukewarm" — 但在他的领域, 模型在 30 分钟内复现 了他自己花了很长时间才做出来的好论文。Mark Chen 教了他一个 "priming" 技巧 (先解一道 textbook warmup), GPT-5 就能解决一篇 training-cutoff 之后才发布的论文。 之后, 他和 PhD 导师 Strominger 把一个 32 项之和、卡了一年的 single-minus gluon tree amplitude 问题给了 ChatGPT — 模型在 Strominger 的飞机降落之前就解决了, 还用 作者们都不知道的技巧给出了证明。第二个实验把题目换成 graviton, 模型在一天内吐出 110 页全新的量子引力, 团队用三周验证。这就是 "vibe physics"。

133. 对谢赛宁的7小时马拉松访谈:世界模型、逃出硅谷、AMI Labs、两次拒绝Ilya、杨立昆、李飞飞和42

谢赛宁 (Saining Xie) · NYU 助理教授 · AMI Labs 联合创始人

谢赛宁的第一次播客访谈,7 小时马拉松对谈, 覆盖 SJTU ACM Class、UCSD 屠卓文、FAIR 何恺明、NYU 李飞飞, 到 2024 年和 Yann LeCun 共同创立 AMI Labs 的全过程。 贯穿主线:representation learning 是 12 年都没解决的核心问题, LLM 是 virtual intelligence,world model 才是真问题。

Extreme Harness Engineering: 1M LOC, 1B toks/day, 0% human code or review Latent Space
≈ 1h 12 min

Extreme Harness Engineering: 1M LOC, 1B toks/day, 0% human code or review

Ryan Lopopolo · OpenAI Frontier Product Exploration · engineer

Ryan Lopopolo 在 OpenAI Frontier 用一个 "out there" 的约束做了 5 个月的实验: 3 个工程师, 1M LOC, 1500 PRs, 他自己一行代码都不写. 这一小时他把这套打法所有的 ratchets 全摊开: 一分钟构建、$land 自动合并、Ghost Libraries、Symphony 用 Elixir、end of bullshit plugins、 对 MCP 的 bearish 判断, 以及一句话哲学: "you can just codex things".