Latent Space

Why Agents Love Supabase — with Paul Copplestone and Ant Wilson

5.4K views · 2d ago · 1:16:39

Guest: Paul Copplestone

Why do AI agents keep choosing Supabase—and what changes when the agent becomes the user? Supabase co-founders Paul Copplestone (CEO) and Ant Wilson (CTO) join swyx for their first podcast together after a $0 to $10B speedrun as the preeminent open source, fully remote Postgres platform company.

Topics: Welcome and why agents love Supabase · Paul Copplestone and Ant Wilson · Compute beside the database · OrioleDB and Multigres · Scaling up, scaling out, and Turso · Developer marketing and memes

How Anthropic ships the #1 LLM in the World — with Katelyn Lesse and Angela Jiang

25K views · 3d ago · 1:46:41

Guest: Katelyn Lesse

An agent’s sandbox can die without taking the whole agent with it. That’s the idea behind “the brain and the hands,” Anthropic’s architecture for Claude Managed Agents. Now that the dust has settled on Opus 5.5, it is universally agreed to be the best model in the world. In this conversation with Katelyn Lesse and Angela Jiang, Anthropic’s Heads of Engineering and of Product, goes inside the systems that turn model capabilities into products developers can use.

What 90,000 Rollouts Reveal About AI Agents Doing Science - Ian Quigley, Leash Bio

5.4K views · 4d ago · 16:34

Leash Biosciences built a platform to physically measure thousands of proteins against millions of molecules. In this followup to our Leash Bio podcast with cofounder Ian Quigley, he returns to discuss the release of Minos, Leash’s framework of pre-measured experimental worlds that scientific agents can explore.

Why AI Compute Comes in Rectangles — Erik Bernhardsson, Modal

7.8K views · 1w ago · 1:25:33

AI workloads don’t run in neat, three-year rectangles. But much of the GPU market still sells capacity that way. Modal CEO Erik Bernhardsson joins Latent Space to explain how Modal pools compute, scales AI workloads and thinks about the financial risks of building a GPU cloud.

Topics: What Modal does · Founding before the generative AI boom · First customers and fundraising · Building for inference and autoscaling · Where Modal gets GPU capacity · Renting vs. owning infrastructure

The Recommendation Engine for Commodity Traders — Anekha Sokhal, Moshi

7.1K views · 1w ago · 26:41

What would you get if Bloomberg, Claude, and Netflix had a baby? Moshi founder and CEO Anekha Sokhal joins swyx at AI Engineer NYC to discuss building AI tools for traders: connecting market data, handling conflicting sources, and turning insights into recommendations people actually use.

Topics: AI Engineer NYC and the one-person team · Meet Anekha Sokhal · From math to trading · What is Moshi? · Structured, unstructured, and private data · Building a connected data graph

AI in NYC: How Tenex Builds an AI-Native Engineering Team — Alex Lieberman & CJ Hess

36K views · 1w ago · 43:33

How do you organize an engineering team when AI agents can do more of the implementation? Alex Lieberman and CJ Hess of Tenex join swyx at AI Engineer New York to discuss paying for outcomes, the MetaHarness process, enterprise AI adoption, and where AI-built software can—and cannot—replace SaaS.

Topics: Title sequence · Introductions · What Tenex does · Why CJ joined Tenex · Paying engineers for outcomes · Inside MetaHarness

The Programmable Data Center — Evan Conrad, SF Compute

29K views · 1w ago · 36:44

GPU compute is bought and sold like a commodity, but delivering it is a lot messier than trading a contract. Evan Conrad of SF Compute joins swyx at AI Engineer New York to explain the market they’re building—and why he thinks it needs an independent operator.

Topics: Why can’t NVIDIA build this? · Welcome back to Latent Space · From selling a room to running the hotel · Physical settlement: actually delivering compute · Liquidity and the GPU order book · Trading nonstandard compute

High Taste Testers: OpenAI Shipped Everything. What Actually Works?

11K views · 1w ago · 1:15:21

Guest: High Taste Testers

Six AI power users compare what actually works—and what falls apart once you start using it. Ben Davis, Ben Hylak, Dex Horthy, Allison Durham, Alex Volkov, and swyx discuss OpenAI DevDay, coding agents, personal assistants, model preferences, and the gap between impressive AI capabilities and useful products.

Topics: High Taste Testers · OpenAI DevDay: Dots, Spaces, and product execution · Great models versus great products · Agents versus assistants · Building a personal assistant with Pi · Decision models, Jev, and predictive interfaces

From Wet Lab to RL Env: 10,000 Drug-Discovery Campaigns in a Weekend — Ian Quigley, Leash Bio

43K views · 2w ago · 1:49:57

From a four-billion-measurement Kaggle competition to models that recognize a molecule’s chemist, Leash Bio is testing what AI actually learns about chemistry. Co-founder and CEO Ian Quigley joins Brandon Anderson and RJ Honicky to explain how Leash makes DNA-encoded libraries (DELs) produce clean, reproducible protein–small-molecule binding data—and why collecting the right experiments matters as much as choosing a model.

Topics: Hook: Learning the Chemist, Not the Chemistry · How DNA-Encoded Libraries Work · Sequencing as a Scalable Biological Readout · Barcodes, Controls, and DNA-Binding False Positives · DNA Tags, Off-DNA Validation, and Reproducibility · Why Leash’s DEL Strategy Works

Robots That Predict the Future: Dyna-2 & Physical AI — Jason Ma, Dyna Robotics

60K views · 2w ago · 1:21:06

From training a world-action model on over one million hours of human video to deploying robots that fold thousands of napkins daily at Din Tai Fung, Dyna Robotics is pushing physical AI beyond impressive demos and into real-world production.

Topics: Hook · Diffusion Transformers, Flow Matching, and Generalist Robots · The Robotics Data Bottleneck · Synthetic Data and Self-Improving World Models · Real-Time Inference and Model Distillation · Full-Stack Robotics: Hardware and Supply Chains

How to kill your SaaS: what we learned replacing a $43K/yr tool with AI agents

58K views · 2w ago · 28:05

A SaaS vendor quoted us $43,000 a year to run speakers for my @AIdotEngineer conferences. That would be fine if it delivered the qualitySo we asked builders to clone it with coding agents, open source, eval by the people who'd actually use it: our human events team.

Topics: Why we tried to kill a $43k SaaS · Why this software is hard · The brief and the rules · What the operator saw · The AI judge · How they built it

Why AlphaFold Didn't Solve Protein Folding — Pushmeet Kohli, Google DeepMind & Sal Candido, Biohub

13K views · 2w ago · 32:42

From the Bitter Lesson of AI scaling to the unsolved mysteries of protein folding, Google DeepMind’s Pushmeet Kohli and Biohub’s Sal Candido are rethinking what it takes to build AI that truly understands biology.

Topics: Hook: AlphaFold Hasn't Solved All of Protein Folding · Finding Scaling Laws and the Right Data for Biology · DeepMind's Bitter Lesson: Start with the Problem · AlphaFold's Data Limits and Building the Virtual Cell · Handcrafted Architectures vs. Scaling Compute · Good Data, Inductive Bias, and Model Design

AI Scientists Are Here: Autonomous Labs & Synthesis Superintelligence — Periodic Labs

141K views · 3w ago · 1:25:21

Guests: Periodic Labs, AI Scientists Are Here

From building AI systems that reason over noisy physical experiments to creating laboratories where every instrument can become intelligent, Periodic Labs is betting that the next frontier of AI won’t come from simply training on more internet data — it will come from letting models experiment with the real world.

Topics: AI and Reinforcement Learning in the Physical World · The End-to-End Materials Discovery Loop · Why Physics Isn’t “Solved” · The Matter Compiler and AI Characterization · DFT, Simulation, and Experimental Ground Truth · Dream Materials, Compute, and Quantum Computing

Which GPU Clouds Are Actually Good? | ClusterMAX 3.0

17K views · 4w ago · 57:52

Which GPU clouds are worth trusting with your AI workloads? This Lightning podcast episode explores SemiAnalysis’s ClusterMAX 3.0 and the questions that matter when evaluating managed GPU clusters. About the research: ClusterMAX evaluates configuration, workload performance, reliability and response to failures through hands-on testing, documentation review and customer feedback.

Recursive Language Models — Alex Zhang, MIT PhD

77K views · 4w ago · 1:43:25

Guest: Recursive Language Models

From GPU kernels and KernelBench to Recursive Language Models, agent harnesses, and massive multi-agent swarms, Alex Zhang is exploring how much capability we’re leaving on the table by wrapping increasingly powerful models in primitive systems.

Topics: Hook · GPU Mode, KernelBench, and AI-Written Kernels · Human Expertise vs. Brute-Force AI Search · Research Taste and Taking Big Bets · GEV and Rethinking the Language Model · Video Game Agents and the Harness Problem

OpenAI’s New Agent Stack: Computer Use, Decisions API, UltraFast, Dots—Ari Weinstein & Nikunj Handa

84K views · 1mo ago · 40:24

From giving AI agents their own cloud computers to making Computer Use faster than humans at real-world tasks, OpenAI is pushing agents much closer to actually operating software end-to-end.

Topics: Hook · OpenAI DevDay: Dots, GPT-6.1, Agents API, and Decisions API · Dots and Personal Cloud Computers · Why Computer Use Is “180 Degrees Different” · From Sky to Self-Debugging Computer Use Agents · How Computer Use Sees and Operates Software

Claude Code’s Bitter Lesson: Prompts, Harnesses, Mods, and Agents — Thariq Shihipar, Anthropic

149K views · 1mo ago · 1:34:32

From the rapid rise of Claude Code to a future where agents can rewrite their own harnesses, collaborate across teams, and operate across cloud and local environments, the way we build software is changing extraordinarily fast. In this episode, Anthropic’s Thariq Shihipar joins swyx and Vibhu to unpack how power users are actually working with Claude Code today, why prompting remains a high-skill discipline, and where Anthropic thinks the agent harness is headed next.

Topics: Hook · Ask User Question and the Future of Agent Interfaces · Artifacts, Projects, and Multiplayer Agents · Prompting as the Core Claude Code Skill · Context, Effort, and Smarter Model Usage · Is Claude.md Going Away?

The $10 Trillion Token Economy — Alex Atallah, OpenRouter & Anjney Midha, AMP

49K views · 1mo ago · 1:22:41

From the earliest days of open-weight models to becoming the neutral routing layer for more than 10 million developers, OpenRouter is one of the clearest bets that the future of AI will be multi-model.

Topics: Hook · Alpaca, Llama, and the Multi-Model Bet · Why “One Model Wins” Was the Wrong Bet · Why Model Labs Struggle With Distribution · “Just a Wrapper”: Why VCs Misunderstood OpenRouter · Bootstrapping OpenRouter Through Community

Runway’s Bet Beyond Video: World Models, Robotics, and the Neural OS — Anastasis Germanidis

52K views · 1mo ago · 1:38:06

Guest: Anastasis Germanidis

From making one of the earliest bets on generative video to building world models that simulate physics, power robots, and render software interfaces directly as pixels, Runway is pushing far beyond AI video generation. In this episode, Runway co-founder Anastasis Germanidis joins swyx and Vibhu to unpack how the company went from experimental creative tools to frontier video models — and why Runway now sees video generation as a path toward general-purpose world simulation.

Topics: Hook · Runway’s Origins and the Bet on Generative Video · The Stable Diffusion Story · Gen-2, Controllability, and the Weekend Hack · From Video Generation to World Models · Learning From the World, Not Just Language

🔬Bio-security is an AI Arms Race - Eric Nguyen (CEO, Radical Numerics)

5.8K views · 1mo ago · 1:31:59

Guest: Eric Nguyen

What happens when AI stops merely reading DNA and starts learning how to write it? Eric Nguyen, CEO and co-founder of Radical Numerics and a key figure behind Evo, joins us to explain the leap from early genome language models to Omni—a system that can outperform specialized models on disease-variant prediction, reason across biological sequences, and even begin to “show its work” when designing better molecules.

Topics: What Is a Genome Language Model? · From Reading DNA to Writing It · The First AI-Generated Genome · Why Omni Is Different From Evo 2 · Predicting Disease-Causing Mutations · Can DNA Models “Reason”?

🔬 Google's AI Scientist Started as an Attempt to Automate Kaggle — John Platt, Google Fellow

7.7K views · 1mo ago · 2:01:26

From inventing textbook machine learning algorithms to winning an Academy Award and discovering two asteroids, John Platt has spent his career at the intersection of computation and science. In this episode, the Google Fellow joins Latent Space to unpack how AI can accelerate scientific discovery, what it takes to turn a research problem into something an agent can optimize, and why better models still need scientists who know how to avoid fooling themselves.

Topics: John Platt: Google Fellow, ML Pioneer, and Academy Award Winner · ERA: Turning Scientific Problems into Scorable Tasks · Inside ERA: Tree Search, Parallel Experiments, and Shared Context · Reward Hacking, Scoring Functions, and the Gemini 2.5 Breakthrough · Predictive vs. Descriptive Models: The Limits of AI Discovery · Overfitting, Auto-Kaggle, and Keeping Scientists in the Loop

Why I couldn't build Jev at OpenAI — Diogo Almeida, TypeSafe Co-founder & CEO

356K views · 1mo ago · 2:22:22

From helping build instruction-following models at OpenAI to spending years trying to answer why AI can solve extraordinarily hard problems yet still automate so little real work, Diogo Almeida is betting that the next wave of AI won’t look like chat. In this episode, the TypeSafe founder and CEO joins swyx days after Jev’s breakout launch to explain System One Models: machine-native AI designed for code, calibrated decisions, reliability, and intelligence per dollar.

Topics: Jev Launch Week and the AI Economic Revolution · What Is Jev? System One Models and Programmable AI · RLHF, Mode Collapse, Calibration, and Yann LeCun · Programmatic AI, Refusals, and Safety Alignment · Why TypeSafe Rejects Public Benchmarks · The Bitterest Lesson: Data, Compute, and the Right Task

A Worm With 302 Neurons Inspired Their Architecture — Ramin Hasani, Liquid AI

7.3K views · 1mo ago · 1:09:52

Tiem for a deep dive into the evolution of AI architectures at Liquid AI, led by CEO Ramin Hasani. The conversation covers the origins of their technology and their approach to building efficient foundation models. Key takeaways include

Topics: Biological inspiration from C. elegans worm DNA · Defining "Liquid" for flexibility and neural networks · Continuous time dynamical systems and recurrence · Overcoming scaling bottlenecks in recurrent neural networks · The "Star" framework: Synthetic design of hybrid architectures · Evolution of Liquid Foundation Models (LFM)

The Watchdogs of AGI — Rune Kvist of AI Underwriting Company

8.1K views · 1mo ago · 1:26:58

From their seed led by Nat Friedman and Dan Gross, AIUC has now announced a monster $40M Series A! From being Anthropic’s first product hire to building the standards, testing, and insurance infrastructure meant to make frontier AI deployable, Rune Kvist is betting that the biggest constraint on AI adoption won’t be capability — it will be trust.

Topics: Hook · AIUC’s $40M Round and the Risk Bottleneck for AI · From Scaling Laws to Early Anthropic · Why Trust, Not Capability, Could Limit AI Adoption · Founding AIUC and Building AIUC-1 · How AI Agents Are Audited and Stress-Tested

Recursive Self-Improvement: from Auto Research to Superintelligence — Richard Socher, Recursive

219K views · 1mo ago · 1:33:27

Guest: Recursive Self-Improvement

From helping pioneer ideas that foreshadowed prompt-based, general-purpose NLP to now building AI systems that can automate AI research itself, Richard Socher is betting that the next major step in AI is recursive self-improvement.

Topics: Hook · The Eureka Machine and Superintelligence · AI Optimism, Slow Takeoff, and Regulation · AI Safety, Reward Hacking, and Anthropic's Constitution · Alignment, Personalization, and Open Source AI · Why Richard Started Recursive

Orbs: Shifting Coding to Cloud — Quinn Slack, Amp Code

12K views · 1mo ago · 40:14

swyx catchs up with Quinn Slack, CEO of Amp, discusses how the company has fundamentally changed software development by embracing AI agents and removing traditional processes like mandatory code reviews (7:03). Key Takeaways from the Discussion

Topics: The shift to AMP and the new 'orb' model · The death of local development environments · Moving past mandatory code review · Personalizable end-user software and dashboards · The legacy of Sandstorm and personal clouds · AI agents for operational tasks and migrations

Trillion Token Context. No, Really — Anima Anandkumar & Benedikt Jenik, Accelerated Understanding

14K views · 1mo ago · 27:02

What would a GPT-style foundation model for the physical world look like? Note: this is a followup to Just days after emerging from stealth, Accelerated Understanding co-founders Anima Anandkumar and Benedikt Jenik, join us to explain their bet: that the universality and scale we’ve seen in language models can also emerge across physics.

Topics: Anima returns & Accelerated Understanding comes out of stealth · A foundation model for the physical world · Can knowledge transfer across different kinds of physics? · Scaling to trillion-context physical models · Evidence for transfer learning across physics domains · Neural operators, geometry & resolution invariance

The Inference Frontier: from 100 to 10,000 tokens per second — Sean Lie, Cerebras CTO

413K views · 1mo ago · 44:03

From pushing inference beyond 4,000 tokens per second to working with OpenAI on a new generation of ultra-fast AI infrastructure, Cerebras is betting that speed doesn’t just make models faster, it makes entirely new kinds of AI possible. In this episode, Cerebras co-founder and CTO Sean Lie joins swyx and Vibhu fresh off Hot Chips to unpack CS4, preview CS5, and explain why 100–200 tokens per second may soon feel like “batch mode.”

Topics: Hook · CS4 and 4,400 Tokens Per Second · From “Impossible” Wafer Scale to OpenAI · CS5 and the Road to 10,000 TPS · Cerebras Is Sold Out — and OpenAI Wants the Capacity · Why OpenAI Is Sharing Ultra-Fast Inference

🔬 Why Transformers Hit a Wall the Moment Physics Shows Up — Anima Anandkumar, Caltech

121K views · 2mo ago · 1:23:32

Most of AI is pointed at language. Anima Anandkumar points it at the physical world, weather, fusion, materials, the systems physics writes down as equations, and finds the usual playbook breaks. There isn't enough data, the resolution is brutal, and no transformer is big enough. Her way through is older than deep learning: build the structure of the world into the model. It already produces weather forecasts that rival the supercomputers on a single GPU.

Topics: Anima's thesis: AI for science, not language models · TorchLean: proving what a neural network will do · Why physics-informed nets fail, and how neural operators fix it · What a neural operator actually is · Fourier neural operators, and why Fourier · Old math meets deep learning, and why scaling breaks

⏭️ Forward Deployed: Voice AI on what works in 2026

3K views · 2mo ago · 36:31

Voice agents are one of the hottest use cases in enterprise right now, but also one of the hardest to actually take live. Getting latency low enough to feel human without dumbing down the responses, making reliable tool calls to a CRM without dropping the customer mid-call, building fallback models for when your main provider goes down. None of it is as simple as the demos make it look.