Over the past year, we've been building our own internal agent infrastructure at YC: over 350 tools, self-improving skill loops, and a shared organizational brain that gets smarter overnight. In this episode of the @LightconePod, we sat down with YC General Partner Pete @koomen to talk about how he led the effort from the ground up. We cover how giving agents unrestricted access to one database was the key unlock, the self-improving skill loops that get smarter overnight, and why he thinks we've arrived at the personal computer moment for AI. 00:39 — YC's AI Stack 02:15 — The Finance Team Problem That Started It All 05:07 — SQL Access Changes Everything 07:20 — One Database to Rule Them All 09:14 — Jevons Paradox 10:07 — Denormalizing for Agents 12:15 — The Single-Player Era of Agents 14:16 — 350 Tools and a Shared Registry 16:24 — Skillify, DRY, and MECE Resolvers 18:23 — The Self-Improving Dream Cycle 20:26 — The Two-Sentence Pitch Skill 23:06 — How Super Intelligence Compounds 25:10 — Recording Everything as a Building Layer 27:10 — The Shared Organizational Brain 29:18 — Trust-Default Culture as a Requirement 30:44 — Raising the Floor for New Employees 32:35 — Horseless Carriages 34:24 — Why Chat Is the Best Interface for Agents 38:50 — Just-in-Time Software 40:49 — Centralizing vs. Decentralizing AI 43:32 — The Personal AI Revolution

Physical Intelligence (@physical_int) is building a foundation model that can control any robot to do any task — what the team describes as the GPT moment for robotics. The company's cross-embodiment approach trains across many different robot platforms, and recent results show tasks being performed zero-shot that last year required hundreds of hours of data collection. In this episode of the @LightconePod , co-founder Quan Vuong (@QuanVng) sat down with @garrytan, @snowmaker, @sdianahu, and @harjtaggar to talk about why robotics is finally ready for its scaling moment, how PI runs its models in the cloud rather than on-device, and the playbook for what Quan sees as a Cambrian explosion of vertical robotics companies. 00:00 — Robotics just got cheaper 00:41 — The GPT moment for robotics 02:24 — Why robots didn’t work before 05:30 — The breakthrough that changed everything 09:12 — The data problem 13:33 — Robots learning without data 15:05 — Robots folding laundry (for real) 22:18 — From engineering problem → ops problem 29:12 — The startup playbook 38:46 — Thousands of robotics startups are coming

François Chollet (@fchollet) has spent years asking a different question than most of the AI world. Instead of scaling what already works, he’s trying to understand what intelligence actually is and how to build it from first principles. In this episode of the @LightconePod, he traces that path from his early work on deep learning to the creation of the @arcprize, and the launch of ARC V3, a new benchmark designed to measure something deeper than performance: the ability to learn, adapt, and reason efficiently in entirely new environments. He explains why today’s systems may be hitting limits, what recent breakthroughs really mean, and why reaching true general intelligence may require a fundamentally different approach. 00:00 - AGI by 2030? 00:31 - Introducing Ndea: A New Path Beyond Deep Learning 01:08 - A New ML Paradigm 01:30 - Replacing neural nets with compact symbolic programs 03:04 - Why Ndea Isn’t Competing With Coding Agents 05:20 - Why Everyone Might Be Wrong About Scaling LLMs 07:22 - Why Coding Agents Suddenly Work So Well 08:50 - The Limits of LLMs in Non-Verifiable Domains 10:48 - What AGI Actually Means (And Why Most Definitions Are Wrong) 13:30 - Why Deep Learning Hits a Wall 14:00 - ARC’s Origin Story 18:20 - ARC Benchmarks Explained: From V1 to V3 22:49 - The RL Loop Powering Coding Agents Today 27:03 - ARC-AGI V3: Measuring “Agentic Intelligence” 31:14 - Inside the ARC Game Studio 35:31 - Could AGI Fit in 10,000 Lines of Code? 44:01 - Building Ndea: From Idea to Compounding Research Stack 46:46 - The Future of ARC: Benchmarks That Evolve With AI 47:21 - Why There’s Still Huge Opportunity for New AI Paradigms 53:37 - How to Build a Breakout Open Source Project - Lessons From Keras 56:39 - Advice For How To Think About AI

With the takeoff of OpenClaw and MoltBook, a new agent-driven economy is taking shape. On the @LightconePod, we took a look at the explosive growth of AI dev tools and whether the time has come for builders to make something agents want. 00:00 - Intro 02:12 - No human involvement is changing the experience 04:55 - Does YC need to change its motto? 07:48 - Email tools and agent infrastructure 09:36 - Agent-driven documentation 13:00 - Swarm intelligence 15:36 - Content generation and dead Internet theory 18:12 - Growth, rules, and founder insights

Wondering why your maker-turned-manager suddenly seems distracted in meetings? Maybe they're addicted to coding agents! In this episode of Lightcone, @calvinfo — a co-founder of Segment and former engineer on OpenAI's Codex team — joins us to talk about why coding agents suddenly feel so powerful, the differences between Codex, Claude Code, and Cursor, and what the future of work will look like. 0:00 Intro 1:15 Garry can’t stop using Claude Code! 4:00 Contrast with IDE’s, context-splitting 6:23 Distribution models, top down vs bottom up 9:11 Licensing and optimization 12:28 Tips on becoming a top 1% user of coding agents 17:36 When can the agents work 24-48 hour running jobs on their own? 21:34 Can the agent teach things like architecture? 26:27 Will the next generation have even better taste and multitasking ability? 29:58 Maker vs manager schedules 31:36 How would Calvin build Segment now? 35:52 The importance of testing 38:52 The Claude bots are talking (to each other!) 40:10 Examples of complex issues, how will the tools evolve 43:00 Outro

.@SpenserSkates has spent more than a decade building Amplitude from a YC startup into a public company, and in that time, he's had to reinvent himself just as much as the product. Joining the @LightconePod, he talks through the shift from founder to large-company CEO, the skepticism his team initially had toward AI, and the moment they realized the next wave of analytics would require a full reset. He walks through the hard reorgs, the bottom-up experiments, and the mindset changes that let Amplitude move fast again. 03:40 - Embracing AI at Amplitude 11:00 - Product roadmap, AI native priorities 15:32 - Org changes & hierarchy 18:40 - “AI killing SaaS”, changing the Amplitude roadmap 23:30 - The “features, not companies” debate, advantages for incumbents 32:14 - Finding good mentors, hyper focus 38:09 - The difference between being a founder vs a big company executive

Logistics is a multi-trillion-dollar industry that quietly powers the entire global economy — and it's shockingly manual. Ryan Petersen (@typesfast), founder & CEO of Flexport, joins the @LightconePod to break down how AI is finally touching the physical world: making shipping cheaper, speeding up global trade, and automating work that used to live inside emails, spreadsheets, and phone calls. 03:17 - When did AI tools become serious at the company 06:27 - The benefit of internal hackathons 12:03 - What internal AI projects have been most impactful at Flexport 14:40 - What software can do better and faster in logistics 19:08 - Goods get cheaper if more logistics get automated 21:18 - The spiritual/philosophical implications of AI in society 23:51 - How does AI change the model structure for companies? 26:38 - Would Ryan have built Flexport differently today?

MIT's new State of AI in Business report went viral for claiming that 95% of enterprise AI projects fail. But the real story isn't that AI doesn't work — it's just big companies can't build it. On the @LightconePod, @garrytan, @harjtaggar, @sdianahu, and @snowmaker break down what the study really says, why in-house enterprise AI efforts keep stalling, and how startups are filling the gap with products that learn, integrate, and actually deliver value. 2:08 - The enterprise AI adoption gap and why the failure rate is high 3:32 - Even Apple can be bad at software 4:30 - Why getting enterprise software to actually work is so hard 11:08 - The Reducto case study 13:39 - The type of enterprise employee you should find as a founder 14:39 - Meet founders who’ve been acquired by enterprises 15:25 - Enterprise/startup tension and symbiosis

Nine out of ten people might tell you you're crazy. The tenth might see what you see. This week on the Lightcone, @garrytan, @harjtaggar, @snowmaker, and @sdianahu discuss contrarian bets — the ideas that look impossible until they work. From Uber and Coinbase to DoorDash and Flock Safety, they share how founders find opportunity where others see dead ends. 0:00 - Intro 2:00 - AI verticals are becoming more crowded 6:22 - Non-obvious successes 9:50 - End users can get regulations changed 18:05 - Finding contrarian ideas, what founders should look for 25:10 - Flock Safety and selling to local governments 33:40 - The sci-fi founder and "impossible" big ideas 36:42 - Outro

In the early days, the only moat that startups have is speed. Once you make something people want, the question becomes what deeper moats can you build on to defend against the competition. On the @LightconePod, @garrytan, @harjtaggar, @sdianahu, and @snowmaker dive into Hamilton Helmer’s Seven Powers framework to find out how these moats show up in practice today in AI startups. 00:00 - The Moat Problem 01:30 - The Seven Powers Framework 04:20 - When to Think About Moats 08:40 - Forward Deployed Engineering 10:18 - Process Power 14:34 - Cornered Resources 19:30 - Switching Costs 24:54 - Counter Positioning 31:24 - The Workforce Displacement Reality 34:00 - Brand & Speed as Moats 37:30 - Network Economies 41:00 - Scale Economies 43:44 - Final Advice

"It's harmless if reporters and know-it-alls dismiss your startup. They always get things wrong. It's even ok if investors dismiss your startup; they'll change their minds when they see growth. The big danger is that you'll dismiss your startup yourself."

Bob McGrew (@bobmcgrewai) helped build some of the most influential technologies of the past two decades. He was an early engineer at PayPal, an early executive at Palantir—where he helped pioneer the Forward Deployed Engineer (FDE) model— and was recently Chief Research Officer at OpenAI - where he led the development of ChatGPT, GPT-4, and o1. On this episode of @LightconePod, he explains how FDEs became central to today's startups, why "doing things that don't scale at scale" works, and where he sees the biggest opportunities for founders working in AI. 00:29 – From PayPal to Palantir to OpenAI 02:19 – The Role of a Forward Deployed Engineer 03:19 – How Palantir Invented It 07:56 – Product Discovery in the Field vs. Sales 09:51 – Echo and Delta Teams Explained 13:34 – Training Ground for Founders 14:35 – Consulting or Real Software? 17:54 – The Birth of Palantir’s Ontology 23:04 – Why AI Companies Adopt It 36:17 – What Success Metrics Look Like 41:14 – Building with Demo-Driven Development 44:56 – Joining the US Army Reserve 47:43 – Opportunities for Founders

Anthropic Co-Founder Tom Brown: Why Anthropic Models Are The Best at Coding "The benchmarks are so easy to game. All the other big AI labs have teams whose job it is to make the benchmark scores good. We don't have such a team. That is the biggest factor." @AnthropicAI's @nottombrown on @ycombinator's @LightconePod with @garrytan, @harjtaggar, @sdianahu, and @snowmaker

Tom Brown (@nottombrown) co-founded Anthropic after helping build GPT-3 at OpenAI. A self-taught engineer, he went from getting a B-minus in linear algebra to becoming one of the key people behind AI's scaling breakthroughs. Today, Anthropic's Claude is the go-to choice for developers, and his team is overseeing what he calls "humanity's largest infrastructure buildout ever." On the @LightconePod, he discusses his unconventional path from YC founder to AI researcher, the discovery of scaling laws that changed everything, and his advice for young engineers entering AI today. 0:00 - From Failure to Success 2:30 - Early Startup Days at Linked Language 4:12 - The Grouper Dating Experiment 6:10 - Making the Leap to OpenAI 8:42 - First Product Launch Challenges 10:12 - Self-Teaching AI Research 12:44 - Building GPT-3 Infrastructure 15:44 - The Anthropic Spinoff 18:23 - Early Days of Building Claude 20:21 - The ChatGPT Wake-Up Call 22:08 - Claude 3.5 Sonnet Breakthrough 24:13 - Why Benchmarks Don't Tell the Whole Story 26:20 - Claude Code's Secret Sauce 28:51 - Building for the AI Agent 31:11 - The Largest Infrastructure Buildout Ever 32:46 - Multi-Chip Strategy 34:38 - Advice for the Next Generation

AI has upended the once "safe" CS career path. New grads are facing unemployment rates twice those of art history majors, and a CS degree is no longer a surefire ticket to wealth. At the same time, small, focused teams are scaling from zero to eight-figure revenue in months. In a special Lightcone Live at AI Startup School, Garry, Diana, Harj, and Jared discuss why it's now more important than ever to focus on building real skills, domain expertise, and agency rather than just chasing credentials. 04:18 - The Inverted Career Risk Paradigm 05:16 - AI's Impact on Education and Skills 07:08 - Agency vs. Credential Maxing 08:28 - Motivation: Fear or Excitement 09:43 - The Accelerated Growth of AI Startups 10:50 - Real Success over Fake Credentials 12:55 - Domain Expertise and Technical Expertise 15:05 - Gaining Domain Expertise as a Student 18:51 - Breaking the Student Mindset 20:39 - The Dangers of Entrepreneurship Programs 22:52 - Social Media Strategy for Startups 27:30 - The College Dropout Question 32:33 - When to Quit Your Job

Alexandr Wang (@alexandr_wang) started Scale AI to help machine learning teams label data faster. It started as a simple API for human labor, but behind the scenes, he was tackling a much bigger problem: how to turn messy, real-world data into something AI could learn from. Today, that early idea powers a multi-hundred-million-dollar engine behind America's AI infrastructure—fueling everything from Fortune 500 workflows to real-time military planning. Just last week, Meta agreed to invest over $14 billion in Scale, valuing the company at $29 billion. Alexandr joined us on @LightconePod to share how Scale evolved from a scrappy YC startup into the backbone of some of the world's most advanced AI systems, how he thinks about competition with Chinese AI labs, and what it takes to build infrastructure that shapes the frontier. 01:15 - Alexandr’s early days at YC 07:25 - Dialing in on what worked 10:24 - Model improvements, evals 19:18 - The techno-optimist view of work 27:47 - The turning points for Scale AI 37:37 - Agentic workflows 41:55 - “Humanity’s Last Exam” 47:48 - U.S. vs China in AI and hard tech 56:57 - How to be hardcore

Prompting AI agents to consistently do what you want is becoming the most important skill for founders to learn and build their companies around. We share some of the more advanced techniques we've learned from founders, and building AI agents ourselves at YC, in the latest @LightconePod episode.

At first, prompting seemed to be a temporary workaround for getting the most out of large language models. But over time, it's become critical to the way we interact with AI. On the @LightconePod, Garry, Harj, Diana, and Jared break down what they've learned from working with hundreds of founders building with LLMs: why prompting still matters, where it breaks down, and how teams are making it more reliable in production. They share real examples of prompts that failed, how companies are testing for quality, and what the best teams are doing to make LLM outputs useful and predictable. 0:58 - Parahelp’s prompt example 4:59 - Different types of prompts 6:51 - Metaprompting 7:58 - Using examples 12:10 - Some tricks for longer prompts 14:18 - Findings on evals 17:25 - Every founder has become a forward-deployed engineer (FDE) 23:18 - Vertical AI agents are closing big deals with the FDE model 26:13 - The personalities of the different LLMs 27:26 - Lessons from rubrics 29:47 - Kaizen and the art of communication