The Interlude Show

Vatsal Bajaj + Jake Kang

Interlude is a show about living through the AI transition. As we meander through the in-between of the pre- and post-AI world, we talk through the use cases, implications, and real-world impact. Helping you keep up as it all unfolds.

Episodes

  1. Oct 23

    The Search for Better Learning Algorithms | Yaroslav Bulatov

    Yaroslav Bulatov wants to rethink how we train neural networks. He worked at Google Brain, early OpenAI, Meta, and Together AI, and now runs the Sutro Group at South Park Commons, exploring energy-efficient learning algorithms with AI agents. We get into his case for replacing backpropagation, why moving data can cost more energy than doing arithmetic, and how gradient checkpointing can make a workload faster even while doing more calculations. We also discuss what simple problems like matrix multiplication and sparse parity can teach us about algorithm design, his experience at early OpenAI, and his skepticism about reinforcement learning. We close with the limits of recursive self-improvement, whether more code translates into better outcomes, and AI that helps us pursue our goals versus AI that captures our attention. Hope you enjoy the seventh episode of The Interlude Show! Further reading: Episode fact-check: https://yaroslavvb.github.io/stuff/interlude-show-bulatov-factcheck.html Memory wall research: https://spacesheep.dev/@yaroslavvb/interlude-show-memory-wall-research Chapters: 00:00 - A plan to rethink backpropagation 00:49 - Introducing Yaroslav Bulatov 01:30 - Energy-efficient learning and the 20-year research plan 03:25 - Why focus on software now? 08:05 - Sutro Group: using agents to explore learning algorithms 10:57 - The memory wall: arithmetic versus data movement 17:07 - Why peak memory is only part of the energy story 18:40 - Why Yaroslav wants to replace backpropagation 24:27 - Mixture of experts and disentangling neural networks 27:24 - Toy problems: matrix multiplication, sparse parity, and MNIST 31:21 - Early OpenAI and skepticism about reinforcement learning 34:41 - Gradient checkpointing: doing more math to move less data 38:51 - Power grids, intelligence per joule, and new bottlenecks 43:25 - Do more tokens and pull requests mean better outcomes? 46:21 - Superintelligence, specialization, and physical limits 51:49 - Creative freedom, attention capture, and possible AI futures

    The Search for Better Learning Algorithms | Yaroslav Bulatov

  2. Oct 7

    What AI Unlocks in Low-Income Countries | Daniel Björkegren

    Daniel Björkegren is a development economist and a professor at Columbia studying how AI is being used in low-income countries and what it will take for its benefits to reach the people who need them most. We get into why teachers in Sierra Leone took up an AI chatbot on WhatsApp, how bandwidth costs change the economics of accessing information, and why speaking the right language is only part of making a model useful locally. We also discuss whether AI could make knowledge services more accessible without following the same development path as wealthier countries, what experiments can tell us about its economic impact, and the opportunities and risks that robotics creates for low-income economies. Hope you enjoy the sixth episode of The Interlude Show! Chapters: 00:00 - AI beyond traditional knowledge work 01:00 - Introducing Daniel Björkegren 02:15 - How AI is being used in low-income countries 03:38 - Teachers, WhatsApp, and the cost of information 10:14 - Local languages and model usefulness 17:09 - AI, services, and economic development 20:57 - Measuring AI's economic impact 27:05 - Should AI be a public utility? 33:01 - Robotics and low-income economies 37:58 - What experiments can and cannot tell us 44:05 - What bond markets might tell us about AI 47:44 - Who benefits

    What AI Unlocks in Low-Income Countries | Daniel Björkegren

  3. Sep 21

    Why serving world models is infinitely harder than LLMs | Avik Sethia

    World Models are slated to be the next frontier after LLMs. We had the pleasure of chatting with Avik Sethia, Founding Engineer at Reactor, who recently raised a $59M Seed and Series A to be the Developer Platform for World Models. Tune in to find out what world models are, the unique challenges behind serving them compared to LLMs and exciting future applications! Chapters 00:00 - Introduction 01:09 - What is Reactor? 01:57 - What is a world model? 05:10 - Interactivity differences from Fei Fei Li and Yann LeCun's world models 12:47 - How world models are used today 15:03 - Challenges of serving world models 19:10 - Differences between serving World Models and LLMs 21:17 - KV Cache Challenges 28:06 - How LLM serving experience translate to world models 30:00 - Architecture differences 34:56 - Open Source World Models 39:35 - Robotics Models 40:51 - Future Applications

    Why serving world models is infinitely harder than LLMs | Avik Sethia

  4. Sep 1

    Jasper Gilley: The Brain Region That Predicts Your Own Thoughts

    Jasper Gilley is a member at South Park Commons researching new AI architectures inspired by how the brain works - specifically the cerebellum, and what it would take for a model to introspect on its own computation the way humans do for free. Before that he was at Yutori, building vision-based web agents that emulate human users. We get into why models need expensive post-hoc interpretability while humans can dissect their own thinking automatically, his forward self-modeling work (training a deliberately undersized model to predict a transformer's own activations, then reading the residual as a novelty signal), why weighting grokking data by Zipf's law turns training into an infinite oscillator that discovers the solution and forgets it forever, why vision beats the DOM for browser agents, why software is an "exoskeleton of intelligence" and demand for it is still elastic, and what humans were actually evolutionarily designed for. Chapters: 0:00 Cold open 1:02 Intro 1:43 Researching brain-inspired architectures at South Park Commons 4:05 The cerebellum as a forward model 6:39 The Forward Self-Modeling paper 14:03 Grokking, explained 18:37 Zipfian Grokking: finding the solution and forgetting it 22:26 Yutori: why vision beats the DOM 31:23 Why Jevons' Paradox won't save your software engineering job 41:51 The window where creativity is the whole job 45:39 What humans were evolutionarily designed for 51:09 The future of post-training

    Jasper Gilley: The Brain Region That Predicts Your Own Thoughts

  5. Aug 13

    Ryan Tsang: How to Ship AI Agents When Failure Isn't an Option

    Ryan Tsang is building Harmonize Health - an AI-native operating system for medical clinics. Ryan spent four years at Microsoft working on Azure OpenAI's fine-tuning product, then trained models for open-source AI safety benchmarks at MLCommons. We get deep into how Harmonize Health builds and grades evals for voice agents (including a user-simulator agent that role-plays the patient calling in), why you should get to production before you build the perfect eval set, how Harmonize scrubs patient data out of transcripts, and the tension between forward-deployed engineering and scaling across multiple clinics. Hope you enjoy the third episode of The Interlude Show! Chapters: 00:00 - Move fast and break things doesn't work in healthcare 00:41 - Meet Ryan Tsang and what Harmonize is building 02:40 - Why healthcare, and the path from pre-med to CS 05:45 - Inside a primary care clinic: the work nobody sees 10:50 - Safety is the hardest technical problem 13:55 - Data plumbing vs. AI: EHRs and tribal knowledge 20:36 - Encoding a clinic's operating procedures into agents 23:35 - Inside their eval harness for voice agents 27:07 - What can we really not screw up? 30:16 - Healthcare moves at the speed of trust 38:57 - TempShot: pitching his high school app to us as VCs 49:53 - The YC all-nighter, quitting our jobs, and why compounding wins

    Ryan Tsang: How to Ship AI Agents When Failure Isn't an Option

  6. Jul 25

    Building Inference at Netflix

    Vatsal built ML Serving and Inference architectures at Netflix and Microsoft for 4-5 years. In this episode, we get into personalization, generative recommendations architecture, inference / serving optimizations and China vs. US on ML research. Hope you enjoy the second episode of the Interlude Podcast! Chapters: 00:00 - Intro 02:18 - How Netflix Personalizes Everything 04:23 - Model Serving vs. Inference 06:19 - Generative & Conversational Recommendations 08:13 - Measuring Success: Metrics & A/B Testing 11:29 - China vs. the US: Open Source & the Road Ahead 17:00 - Risks, Incentives & Regulation 20:17 - Serving LLMs vs. Classical ML & Observability

    Building Inference at Netflix

  7. Jul 8

    Building Agentic AI at a YC Startup

    Jake was the first engineer at a YC startup shipping AI in production. We get into how AI actually works, harness engineering vs. post training, the importance of evaluations and war stories from deploying browser agents in the wild.Hear from him directly on the first episode of the Interlude Podcast. Chapters 00:00 - Intro 00:52 - CloudCruise (YC W24) 02:41 - Building a Self-Recovering Browser Agent 04:18 - What Are AI Evaluations? 05:01 - Building Eval Datasets: 90% of the Work 09:46 - Offline vs Online Evals & Distribution Shift 12:06 - Harness Engineering vs Fine-Tuning 15:18 - Owning Your AI in Regulated Industries 17:35 - Hot Take: The OpenAI/Anthropic Threat 20:52 - The Future of Browser Automation

    Building Agentic AI at a YC Startup
  • 7 Episodes

About

Interlude is a show about living through the AI transition. As we meander through the in-between of the pre- and post-AI world, we talk through the use cases, implications, and real-world impact. Helping you keep up as it all unfolds.

Information

  • Creator
    Vatsal Bajaj + Jake Kang
  • Years Active
    2026
  • Episodes
    7
  • Rating
    Clean
  • Copyright
    © Vatsal Bajaj + Jake Kang
  • Show Website
    The Interlude Show