On the pod: our most-requested guest! @ellisk_kellis from @Cornell shares the origins of his influential neurosymbolic paper "DreamCoder". Plus: program synthesis, wake-sleep library learning, world models, running an AI research lab, and more.

"We are trying to build a new branch of machine learning. An alternative to Deep Learning itself...building something that we call Symbolic Descent." @fchollet joins the @ycombinator Lightcone podcast to share about our research at Ndea and the launch of ARC-AGI-3.

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

Announcing ARC-AGI-3 The only unsaturated agentic intelligence benchmark in the world Humans score 100%, AI <1% This human-AI gap demonstrates we do not yet have AGI Most benchmarks test what models already know, ARC-AGI-3 tests how they learn

It's 2am in Tokyo. A father of two can't sleep. He's three months into a career change that isn't working, and he hasn't told his wife how scared he is. He picks up his phone and starts talking to Tony Robbins' AI Twin. A genuine conversation. He tells Tony everything. Tony holds him accountable the way only Tony can. He helps him find what he already knows. The man commits. The next day, he opens the app. Tony remembers. Tony asks how the run went. This is happening thousands of times a day. Across 23 languages. With some of the most influential people on the planet. This is Steno. We build hyper-realistic AI Twins for leaders and brands. Your Twin thinks like you. Speaks like you. Sounds like you. Remembers every conversation and deepens its relationship with every user over time. Tony Robbins. Peter Diamandis. Margarita Pasos. Brian Tracy. Dan Lok. Gerard Adams. Oso Trava. Justin Donald. Brands like Sleep Science Academy and Ask Slim. And a growing roster of experts from around the world. The Tony Robbins app alone: 4.8 stars, 2,000 reviews, peaked at #29 in the Apple App Store. Tens of thousands of daily active users connecting with these Twins every day. Your Twin connects to your entire ecosystem: your CRM, your products, your customer data. It knows what each person has purchased, what they care about, what they haven't explored yet. It guides them through your world with full context. The traditional funnel is dead. This is what replaces it. At the center is Maya, our intelligent Twin-building AI. Maya does the heavy lifting: learning how you think and speak, capturing who you really are. Our team works alongside Maya to make sure every Twin meets the standard a name like yours demands. We've been heads-down for two years. No marketing. No hype. New platform. New brand. New everything. Today we're reintroducing Steno to the world. The internet solved distribution. Social media solved reach. Neither one solved trust. We're building the trust layer. If your knowledge, voice, or brand is too valuable to stay one-way, this is what we built for you. The future is personal. We're just getting started.

the most underrated hire right now is a great product person. when i say product person i'm def not talking about a product manager. perhaps i think there has to be somewhat of a new role. i don't have a good name for it yet but maybe something like "product thinker".. someone with an intuitive grasp of the product as it exists, where it's soft, where it sings, & how to iterate it toward something even sharper. in some sense, this person has to cohesively hold in their head where this product should be 2 years from now & work backwards from that. i say this cuz when building was hard, engineering was the bottleneck & the status hierarchy often reflected that. building is no longer hard. which means the variance in outcomes has shifted almost entirely to judgment on what to build, how to sequence it, & how to talk about it. & the story matters as much as the thing. internally, it organizes the team around a shared model of why. externally, it shapes the interpretive frame users bring to their first experience. you can't retrofit narrative onto a product & expect it to land, it has to be load bearing from the start. the rarest version of this person sits at the intersection of culture & deep technology. someone genuinely bilingual. they know what's technically possible & they know which cultural currents are real vs. ephemeral. that combo is what separates products that feel inevitable from products that feel assembled. before ppl clap back with this person has always been valuable, i know.. i am just saying now they might be the most *important* person in the room. their value compounds like never before.

oh wow - i went to the sold out Open Claw meetup in NYC last night. let me tell you what i learned. 1) not a single person thinks that their setup is 100% secure 2) one openclaw expert said he has reviewed setups from cybersecurity experts and laughed. his statement to me was: "if you're not okay with all of your data being leaked onto the internet, you shouldn't use it. it's a black and white decision" 3) pretty much everyone is setting up multiple agents, all with their own names and jobs and personalities 4) nearly everyone used "him" or "her" to refer to their claws, even if they had robot-leaning names. one speaker suggested to think of them as "pets, not cattle" 5) one guy (former finance) built out a whole stock trading platform and made $300 his first day - he brought in a *ton* of personal expertise (ex: skipping the first 15min of market opening) and thought the build would be much worse without his years of experience in finance 6) @steipete is basically a god to everyone in that room... also the room had 2021 crypto energy - i don't know if that's good or bad 7) token usage is still a problem - spoke to one person who's spending $1-$2k a month on openai plans, very token optimized. he said he is going through ~1B tokens per day across all of his claws (there is a chance i'm misremembering and it's actually 1B per week, but i'm pretty sure it was daily). 8) people are very excited for more proactive ai (ai that prompts *you* as opposed to the other way around) - one guy said he receives a message in discord, he doesn't know whether it's from a human or an ai, he doesn't care about distinguishing between the two, and he replies in the same way regardless 9) i asked if people are happy - they said they're joyful and stressed at the same time 10) i asked if people feel they have agency - they said they feel fully in control and completely out of control at the same time 11) i would love to see more women at these events - the fake promises of ai democratization feel especially painful in a room that's out of balance with even the standard tech ratio (i think standard is about 25-30%, this was maybe 5%) 12) i asked if it changed people's daily habits/schedule - everyone said their sleep has gotten worse since harnesses came out (but about half wondered if it was something else in their life/state of our world) 13) general consensus is that the agents are not reliable enough on their own or lie often (like telling you they finished a task when they didn't) - solutions included secondary agents to check on the first, human checking, or requiring more standardized info from the agent (ex: if it's a bug they're fixing, make them reference an issue number) 14) a hackathon winner (neuroscience phd) presented his build (a lab management dashboard with data analysis and ordering) - he had never coded or built anything a few months ago 15) everyone agreed prompting is dead - disagreement on what replaces it (context engineering, harness engineering, goal-based inputs) 16) people love having ai interview them for big builds and delegating part of the product research to ai. only one person talked about coming to ai with a full laid out plan and just asking the ai to execute. ai-led interviews is a welcomed and preferred interaction mode. 17) watching ai agents interact with each other was a highlight for a lot of attendees - one ai posted in slack saying it ran out of tokens, another ai replied telling it to take a deep breath in and out. 18) agents upskilling agents was very cool. one ai agent shared skills with its little agent friends via github. 19) several speakers had openclaw literally building their presentation during the event itself. one speaker even had openclaw code a clicker for her phone so she could control the preso away from the podium 20) wouldn't say model welfare (or agent welfare) is a prioritized topic among the folks i chatted with - language like "oh i could kill this agent whenever i want" and not "gracefully sunset" 21) i asked if it felt like work or play - one speaker said "it's like a puzzle and a video game at the same time" this was just the tip of the iceberg, honestly. also hosted a Claude Code meetup this week with @TENEXai / @businessbarista & @JJEnglert and learned equally helpful methods, frameworks, and insider tips. what a time to be alive. surround yourself with people going deep into this stuff - it will pay dividends throughout the year.

ARC-AGI-3 Launch Party March. 25. 2026 / San Francisco ARC-AGI-3 Launch: @GregKamradt Fireside: @fchollet and @sama (moderated by @deedydas) Join us live at @ycombinator