A User-Centric Perspective on LLM Inference | AM Podcast #3

Peter Norvig is the co-author of Artificial Intelligence: A Modern Approach and a former director of research at Google. He is now a Distinguished Education Fellow at Stanford’s Human-Centered AI Institute and a researcher at Recursive, working on self-improving AI. His work spans AI research, software engineering, and education.Outline:0:00 - Introduction2:32 - What Has (and Hasn’t) Changed Throughout Peter’s Career?7:03 - How Peter Makes Career Pivots11:17 - Is This AI Moment Different From Previous Cycles?15:32 - What’s Still Worth Learning in AI and Programming?19:55 - How Will AI Transform Higher Education?24:10 - Main Bottlenecks to Building Human-Centered AI35:10 - Reward Hacking vs. Creativity41:51 - Continual Learning Remains an Unsolved Challenge44:16 - Open Source AI Models46:32 - AI, Inequality, and Winner-Take-All MarketsReferences:- Peter’s Homepage: https://www.norvig.com/ - Artificial Intelligence: A Modern Approach: https://aima.cs.berkeley.edu/ - Teach Yourself Programming in Ten Years: https://www.norvig.com/21-days.html - The Unreasonable Effectiveness of Data: https://research.google/pubs/the-unreasonable-effectiveness-of-data/- Recursive: https://recursive.com/ Podcast Links:Podcast website: https://augmented-mind.github.io/Apple Podcasts: https://podcasts.apple.com/us/podcast/augmented-mind-podcast/id1868102170Spotify: https://open.spotify.com/show/40KculkYTe2tOpqJm6TAYr?si=PU_UncsMT4mXjVNCRwoXog&nd=1&dlsi=6d9bed7a43d64085RSS: https://anchor.fm/s/10dbf5b7c/podcast/rssAbout the Hosts:The AM Podcast is hosted by Yijia Shao, Shannon Shen, and Michael Ryan, CS PhD students at Stanford University and MIT.

Geoffrey Litt is a design engineer at Notion. Previously, he was a senior researcher at the independent research lab Ink & Switch, and he got his PhD in HCI at MIT. Geoffrey’s work focuses on malleable software and designing environments for thinking.Outline:0:00 - Introduction0:47 - Design Is More Than Better Animation3:32 - Malleable Software and Coding Agents9:06 - Human-Software Interaction is a Full-Stack Problem12:43 - Getting Users On Board and Deeper15:53 - The Trap of Just Building More20:27 - Primitive Design Is a Taste Problem22:20 - Understanding Is the New Bottleneck27:22 - How To Develop Taste?34:49 - Productivity vs. Speed38:10 - Will Better AI Hurt Human-Human Collaboration?44:34 - Advice for Designing Better ProductsReferences:Geoffrey’s Homepage: https://www.geoffreylitt.comMalleable Software: https://www.inkandswitch.com/essay/malleable-software/Understanding is the new bottleneck: https://www.geoffreylitt.com/2026/07/02/understanding-is-the-new-bottleneckNotion Workers: https://developers.notion.com/workers/get-started/overviewPodcast Links:Podcast website: https://augmented-mind.github.io/Apple Podcasts: https://podcasts.apple.com/us/podcast/augmented-mind-podcast/id1868102170Spotify: https://open.spotify.com/show/40KculkYTe2tOpqJm6TAYr?si=PU_UncsMT4mXjVNCRwoXog&nd=1&dlsi=6d9bed7a43d64085RSS: https://anchor.fm/s/10dbf5b7c/podcast/rssAbout the Hosts:The AM Podcast is hosted by Yijia Shao, Shannon Shen, and Michael Ryan, CS PhD students at Stanford University and MIT.

Zora Wang is a CS PhD student at Carnegie Mellon University. Zora’s work lies at the intersection of coding agents and human-ai collaboration.  In this episode we explore how agents work, how humans work, and how we can get more done together.Outline:0:00 - Introduction4:57 - Agent Workflow Memory9:26 - From Program Synthesis to Agent Research11:16 - Tools, Actions, and Skills18:47 - Human Work vs. Agent Work21:19 - Autonomous Agents vs. Collaborative Agents29:47 - Picking Research Problems in the Age of AI Agents34:09 - The Role of Coding Ability in Agent Development36:49 - Cutting Through the Fear Mongering: Building Agents Responsibly45:02 - What’s Next? (The Future of Human-Agent Collaboration)52:42 - Final TakeawayReferences:Zora’s Homepage: https://zorazrw.github.io/ Agent Workflow Memory: https://arxiv.org/abs/2409.07429 How Do AI Agents Do Human Work? Comparing AI and Human Workflows Across Diverse Occupations: https://arxiv.org/abs/2510.22780 Podcast Links:Podcast website: https://augmented-mind.github.io/Apple Podcasts: https://podcasts.apple.com/us/podcast/augmented-mind-podcast/id1868102170Spotify: https://open.spotify.com/show/40KculkYTe2tOpqJm6TAYr?si=PU_UncsMT4mXjVNCRwoXog&nd=1&dlsi=6d9bed7a43d64085RSS: https://anchor.fm/s/10dbf5b7c/podcast/rssAbout the Hosts:The AM Podcast is hosted by Yijia Shao, Shannon Shen, and Michael Ryan, CS PhD students at Stanford University and MIT.

Jeremy Avigad is a professor in the Department of Philosophy and the Department of Mathematical Sciences at Carnegie Mellon University. Jeremy is a pioneer in using AI for Mathematics and the co-creator of the Lean Theorem Prover. Currently, he is the director of the Hoskinson Center for Formal Mathematics at CMU, Dean's Chair in Logic and Philosophy of Mathematics, and the director of the newly-established Institute for Computer-Aided Reasoning in Mathematics under NSF.Outline:0:00 - Teaser1:04 - Monologue2:50 - The Historical Landscape of AI for Mathematics7:28 - Formalization and Computer-Aided Proof11:56 - The Birth of the Lean Project21:21 - Lean Blueprint, Model Training with Lean, Using Lean in Agentic Systems29:48 - Making AI Actually Useful for Mathematicians32:46 - How AI is Changing Mathematics36:29 - "It's Our Mathematics, and Us Doing Mathematics"43:04 - The Verification Gap in Human-AI Collaboration47:46 - The Future of Math Education52:23 - Capital, Startups, and the Mathematicians' Ecosystem1:01:08 - PredictionsReferences:Jeremy’s Homepage: https://www.andrew.cmu.edu/user/avigad/The Lean Theorem Prover: https://lean-lang.org/papers/system.pdfLean projects: https://leanprover-community.github.io/lean_projects.htmlPodcast Links:Podcast website: https://augmented-mind.github.io/Apple Podcasts: https://podcasts.apple.com/us/podcast/augmented-mind-podcast/id1868102170Spotify: https://open.spotify.com/show/40KculkYTe2tOpqJm6TAYr?si=PU_UncsMT4mXjVNCRwoXog&nd=1&dlsi=6d9bed7a43d64085RSS: https://anchor.fm/s/10dbf5b7c/podcast/rssSpecial Thanks to Zixiao Jolene Wang and Rajarshi Mukherjee for their help with this episode!About the Hosts:The AM Podcast is hosted by Yijia Shao, Shannon Shen, and Michael Ryan, CS PhD students at Stanford University and MIT.

Ken Liu is a Stanford CS PhD student and founder of The Open Anonymity Project. Ken’s pioneering work explores the intersection between language models and data & user privacy.Outline:0:00 - Teaser1:08 - Prelude: Introducing Ken Liu1:41 - Monologue: The Open Anonymity Project3:41 - Ken’s Path to Privacy Research6:31 - The Biggest Privacy Concern for LLM Users9:39 - Three Perspectives on Tackling AI Privacy10:57 - “AI presents a Uniquely Worse Privacy Problem”13:44 - The Open Anonymity (OA) Project: Unlinkable Inference17:50 - Blind Signatures as Unlinkable Authentication20:52 - Secure Inference Proxies28:31 - Threat Model in the OA Project31:39 - What If People Give Away Information In Their Prompts35:58 - OpenClaw, Privacy Nightmare In Agents43:00 - The Stories Behind the OA Project50:14 - Intelligence Neutrality52:22 - Safety Concerns in a World with Private AI InferenceReferences:Ken Liu’s Home Page: https://ai.stanford.edu/~kzliu/The Open Anonymity Project: https://openanonymity.ai/Unlinkable Inference as a User Privacy Architecture: https://openanonymity.ai/blog/unlinkable-inference/Podcast Links:Podcast website: https://augmented-mind.github.io/Apple Podcasts: https://podcasts.apple.com/us/podcast/augmented-mind-podcast/id186810217Spotify: https://open.spotify.com/show/40KculkYTe2tOpqJm6TAYr?si=PU_UncsMT4mXjVNCRwoXog&nd=1&dlsi=6d9bed7a43d6408RSS: https://anchor.fm/s/10dbf5b7c/podcast/rssAbout the Hosts:The AM Podcast is hosted by Yijia Shao, Shannon Shen, and Michael Ryan, CS PhD students at Stanford University and MIT.

Woosuk Kwon is CTO of Inferact and creator of the vLLM inference library. Woosuk shares what it takes to build the most popular open-source LLM inference engine from a human-centered perspective.Outline:0:00 - Prelude: Introducing Woosuk and Inferact3:00 - Woosuk’s First PhD Project6:00 - How the vLLM Project Got Started9:18 - AI Infra Needs More Than Just Efficiency14:08 - How AI Infra and Human-centered AI Are Connected15:01 - How to Prioritize Feature Requests for Popular AI Infra18:18 - Streaming Requests and Realtime API24:05 - Multi-turn, Agentic, Proactive LLMs27:03 - How to Design AI Infra in a Principled Way29:13 - How to Design an AI Inference Engine for Continue Learning with RL35:05 - Would LoRA Training Affect RL Infra Design?37:28 - Why Start an AI Inference Infra Startup?40:46 - What Effortless Inference with Open-source Models Means for Developers43:46 - A Vision for On-device AI Inference46:19- Can Today’s Coding Agents Create vLLM?References:Inferact: https://inferact.ai/Efficient Memory Management for Large Language Model Serving with PagedAttention: https://arxiv.org/abs/2309.06180Streaming Requests & Realtime API in vLLM: https://vllm.ai/blog/streaming-realtimeRL’s Razor: Why Online Reinforcement Learning Forget Less: https://arxiv.org/abs/2509.04259Podcast Links:Podcast website: https://augmented-mind.github.io/Apple Podcasts: https://podcasts.apple.com/us/podcast/augmented-mind-podcast/id1868102170Spotify: https://open.spotify.com/show/40KculkYTe2tOpqJm6TAYr?si=PU_UncsMT4mXjVNCRwoXog&nd=1&dlsi=6d9bed7a43d64085RSS: https://anchor.fm/s/10dbf5b7c/podcast/rssAbout the Hosts:The AM Podcast is hosted by Yijia Shao, Shannon Shen, and Michael Ryan, CS PhD students at Stanford University and MIT.

Sherry Wu is a professor at CMU whose research sits at the intersection of human-computer interaction and natural language processing. From making AI work for imperfect humans to making humans work better with AI — Sherry's work challenges us to rethink both sides of the equation.Outline:0:00 - Teaser1:13 - Prelude: Introducing Sherry Wu2:30 - How the AI Field Has Changed in the Last Four Years4:22 - Making AI Systems Work for Imperfect Humans6:54 - Models vs. Scaffolding10:36 - Understanding Human Imperfection in Teaching Contexts19:28 - AI Literacy Skills22:04 - How AI Is Changing CS Education25:38 - Suppose We Have AGI, What Does It Mean to Be Human?29:14 - Training Models to be More Human-centered31:46 - Checklists Are Better Than Reward Models https://arxiv.org/abs/2507.1862436:56 - Challenge in Aligning Models43:22 - Advice for Interdisciplinary Research45:37 - Reflection on Her Own ResearchReferences:Sherry Wu’s Research Homepage: https://www.cs.cmu.edu/~sherryw/Sherry Wu’s course page (PMDS, Spring 2025): https://www.cs.cmu.edu/~sherryw/courses/2025s-pmds.htmlAI Fluency Index: https://www.anthropic.com/research/AI-fluency-indexChecklists Are Better Than Reward Models: https://arxiv.org/abs/2507.18624Not Everyone Wins with LLMs: https://arxiv.org/pdf/2509.21890Podcast Links:Podcast website: https://augmented-mind.github.io/Apple Podcasts: https://podcasts.apple.com/us/podcast/augmented-mind-podcast/id1868102170Spotify: https://open.spotify.com/show/40KculkYTe2tOpqJm6TAYr?si=PU_UncsMT4mXjVNCRwoXogRSS: https://anchor.fm/s/10dbf5b7c/podcast/rssAbout the Hosts:The AM Podcast is hosted by Yijia Shao, Shannon Shen, and Michael Ryan, CS PhD students at Stanford University and MIT.

Omar Shaikh is a Stanford PhD student, HCI and NLP researcher, and author of the award-winning UIST 2025 paper “Creating General User Models from Computer Use”. Omar’s pioneering work aims to bridge the Human-AI grounding gap.Outline:0:00 - Teaser1:21 - Prelude: Introducing Omar Shaikh2:07 - Monologue: Better Context for AI4:22 - Bridging the Human-AI Grounding Gap6:14 - Confidence scores in General User Models (GUMs)7:32 - Calibration of General User Models13:20 - Uses of General User Models15:01 - Mixed Initiative Interactions22:10 - Motivation for GUM25:31 - Tabracadabra: tab everywhere!27:01 - Design decisions in GUM28:26 - Designing Interactive Experiences32:11 - DITTO: https://arxiv.org/abs/2406.0088833:06 - Work on Domains without Existing Benchmarks34:45 - Challenges of the GUM Project37:26 - Privacy and Data Ownership38:57 - Finetuning a User Model44:09 - Mindblowing GUM Inferences49:02 - Social Problems of GUMs50:27 - GUM as a Reflection ToolReferences:Omar Shaikh’s research homepage: https://oshaikh.com/Creating General User Models from Computer Use: https://arxiv.org/abs/2505.10831Tabracadabra: https://x.com/oshaikh13/status/1967626897837494479Aligning Language Models with Demonstrated Feedback: https://arxiv.org/abs/2406.00888Principles of Mixed-Initiative User Interfaces: https://erichorvitz.com/chi99horvitz.pdfVerification of Forecasts Expressed in Terms of Probability: https://journals.ametsoc.org/view/journals/mwre/78/1/1520-0493_1950_078_0001_vofeit_2_0_co_2.xml Podcast Links:Podcast website: https://augmented-mind.github.io/Apple Podcasts: https://podcasts.apple.com/us/podcast/augmented-mind-podcast/id1868102170Spotify: https://open.spotify.com/show/40KculkYTe2tOpqJm6TAYr?si=PU_UncsMT4mXjVNCRwoXogRSS: https://anchor.fm/s/10dbf5b7c/podcast/rssAbout the Hosts:The AM Podcast is hosted by Yijia Shao, Shannon Shen, and Michael Ryan, CS PhD students at Stanford University and MIT.

Introducing The Augmented Mind Podcast (The AM Podcast). We explore techniques for building AI models that collaborate with people and augment human intelligence. In Episode 0, we share who we are, why we started this podcast, and what we're looking forward to.Outline:0:00 - Prelude: the problems we care about1:48 - Host introduction2:03 - Why we started the AM Podcast2:31 - Hot takes on human-centered AI2:45 - Hot take #1: learning on outcome rewards over long horizons will directly solve human-agent collaboration3:00 - The Bitter Lesson3:53 - How to define rewards is a human problem4:50 - Empathetic AI5:48 - Hot take #2: even with an automation-vs-augmentation view, as AI gets stronger, there will be less for us to work on6:09 - Creative Destruction7:21 - Task vs. goal10:45 - Format of our podcast11:28 - Unique technical challenges in human-centered AI11:43 - Example #1: human variation13:58 - Example #2: revolution of annotation and data collection15:10 - Example #3: making sense of noisy data16:45 - Let the journey begin!External Clips Referenced:Eric Horvitz; 1:02:38 - 1:03:07 https://www.youtube.com/watch?v=ddjNTxtyEnwFei-Fei Li ; 12:40 - 12:58 https://www.youtube.com/watch?v=be0gLzeBX5wPodcast Links:Podcast website: https://augmented-mind.github.io/Apple Podcasts: https://podcasts.apple.com/us/podcast/augmented-mind-podcast/id1868102170Spotify: https://open.spotify.com/show/40KculkYTe2tOpqJm6TAYr?si=PU_UncsMT4mXjVNCRwoXogRSS: https://anchor.fm/s/10dbf5b7c/podcast/rssAbout the Hosts:The AM Podcast is hosted by Yijia Shao, Shannon Shen, and Michael Ryan, CS PhD students at Stanford University and MIT.