GroundZero AI Talks

This is Episode 15 on Ground Zero!Ludwig is 31yo programmer with no formal education. He dropped out in 7th grade - started programming around age 10 - began working at 15 - and later worked on distributed systems for a company with hundreds of millions of users. You might know Ludwig for systems programming, ML, intelligence, mathematics, software-engineering culture and his X presence.Ludwig [Guest]: https://x.com/ludwigABAPHimanshu [Host at GroundZero]: https://x.com/himanshustwtsGroundZero: https://groundzeroai.inSPONSOR: Haize Labs (https://haizelabs.com) is building expert-level agents for mission-critical work, powered by proprietary Reliability Harness. They are doing some ambitious and interesting work. Checkout Leonard Tang (Cofounder and CEO) articles to learn more: https://x.com/leonardtang_/articlesTo sponsor future episodes, visit https://www.groundzeroai.in/partnerTIMESTAMPS:0:00:00 - INTRO0:01:06 - What are you most excited about these days?0:03:31 - All arcs are isomorphic, Dropping out in 7th grade0:09:31 - Does skipping formal education make you naturally broader?0:10:11 - Growing up on 4chan, Mathematics is humbling0:14:52 - Mentors, First programming job, Internet microcultures0:21:47 - Culture in SF, Why he won't start a company0:25:21 - Does AI make self-teaching easier or trap you in infinite explanations?0:29:49 - The 20% doing 80% inside a company0:33:24 - Company culture and Building subcultures0:35:56 - How do you define understanding, compression equals prediction0:45:44 - Understanding vs knowing, Michael Levin, Cognitive light cones0:54:39 - Starting a research lab, Sheaf theory and Grothendieck0:58:43 - Mapping Levin's biology onto proof space with MCTS and Lean41:02:49 - Active inference, Distributed systems and Morphogenesis1:12:28 - Signal vs noise on X, The unit distance proof, Flanderization1:17:39 - Community Questions1:21:39 - Shape rotator vs wordcel, Tenstorrent, How to hire great engineers1:28:01 - Building mental space to do things that will be left to do1:33:42 - Advice to a 20-year-old. Be a capable thinker, you have 20,000 days left

Curtis Northcutt joins us on GroundZero.He is Director of AI Research at Handshake, founder of Cleanlab (acquired by Handshake) and invented confident learning during his PhD at MIT. In this conversation, we dig into why the delta between model generations is almost entirely human data, the four quadrants of the AI market, how expert data labeling is deeply misunderstood, the long tail of human knowledge that AI still can't touch, RL environments, coding benchmarks and nuances around Human Data market.PS: Handshake’s revenue has risen to nearly $1 billion up from $550 million in January and $5 million a year ago.TIMESTAMPS00:00:00 - Intro00:02:08 - The Snowman Effect: Growing Up in Rural Kentucky to 7 Years at MIT00:09:50 - Confident Learning, Two Systems of Intelligence & Why Classical ML Still Matters00:20:37 - The Origin of Cleanlab & the Handshake Acquisition00:28:00 - How Does GPT-5 Become GPT-6?00:31:20 - The Fastest Growing Data Lab: What Makes Handshake Different?00:35:32 - The Difference Between Good Data & Bad Data00:38:17 - New Kinds of Data, IA & AI00:42:04 - Scaling Coding Benchmarks & Efficiency in Long-Horizon Tasks00:49:12 - Pre-Training & Misconceptions Around Human Data00:57:13 - Open Questions: Taste, Personality & Quick Fire01:05:15 - Advice to Your 20-Year-Old Self: Skill & Obsession

In this episode, Viv joins us on Ground Zero. He leads open source agents and harnesses at LangChain. In this conversation, we go deep into what a harness actually is, Deepagents, Context rot, Why agents should be more opinionated, Harness as a Service, File System, how different frontier models behave in agent workloads, Meta Harness, RL Envs and beyond that.TIMESTAMPS00:00:00 - Intro00:03:30 - PhD at Temple & First Startup in Vision00:12:42 - Bio Internship at Lockheed Martin & Mike Mill Lore00:16:18 - Building at LangChain, DeepAgents & R&D in Open Source00:25:09 - Harness Design, Filesystems & Foundational Harness Primitives00:42:25 - Trajectories for Continual Learning, Where Not to Use RL, Skills & Context Rot00:56:30 - Harness Engineering: OpenClaw, Hermes, Custom Harnesses & Quick Fire01:12:15 - Meta-Harnesses, Self-Improvement Loops, Simulation-as-a-Service & RL Environments01:28:30 - Advice to Anyone 20 / Starting Out in College

Francesco Pappone (@tensorqt) is building Paradigma - an infrastructure for autonomous research. Their product Flywheel replaces papers with directed acyclic graphs where every experiment is a node with tracked lineage, and AI agents traverse and extend the graph via MCP. In this conversation, we go through Francesco's path from physics to founding Paradigma, how Flywheel works, building from Rome, how it connects to Karpathy's autoresearch and the broader auto-research movement, the hard problems around trust and verification when AI runs experiments, and where autonomous research is headed.TIMESTAMPS00:00:00 - Intro00:01:03 - Dropping Out of a PhD, Saturation in Academia & Building from Europe00:13:15 - AutoResearch, The Idea Behind Flywheel & the ex-OpenAI Co-Founder00:19:18 - Evals for AutoResearch & Research Taste00:25:00 - How Research Labs Use Flywheel00:31:15 - The Thesis Behind AutoResearch & Adoption Across Top Labs00:41:00 - Working with Flywheel & the Combinatorial Explosion Problem01:02:03 - Science as a Public Graph, New Research Loops & Quick Fire01:06:23 - The Toughest AutoResearch Challenge Today & Future Directions01:11:53 - Automating Science: Sakana, DeepMind & the Trust Problem01:21:07 - Business Model, Automating R&D & Model Capabilities01:27:20 - Dropping Out of College to Live on the Frontier

Arcee AI went from post training company to building their own frontier reasoning model from scratch + and open-sourcing all of it. In this episode, Lucas Atkins (Co-Founder & CTO) and Varun Singh (Pre-Training Lead) take us through the full story — the decision to start pre-training, how they built and trained Trinity on a fraction of the budget that frontier labs spend, the data and engineering challenges along the way, and their views on where open-source AI is headed. A really honest conversation about what it actually takes to build at the frontier.TIMESTAMPS00:00:00 - Intro00:00:59 - Varun’s Transition from SWE to Pre-Training Lead00:04:20 - The Trinity Manifesto & OpenClaw Ecosystem00:12:15 - Arcee’s Post-Training to Pre-Training Pivot00:23:45 - Varun’s First Pre-Training Run (“You Can Just Do Things”)00:27:33 - Saturation in Pre-Training? Mid-Training Explained00:37:00 - Tweaking Training Architecture, Adam vs Muon & Evals01:09:07 - Inference Engineering, Quick Fire & Post-Training Recipes01:18:02 - Alpha in RL Environments & Harness Design01:23:00 - Why American Open Source Is Trailing Chinese Competitors & Trinity Adoption01:29:25 - Hiring at Arcee & Advice to Your 20-Year-Old Self

In this episode, Vishnu Patankar (CTO, EigenCloud) joins us for a wide-ranging conversation on AI, infrastructure, and the economy — from chips and compilers to agent harnesses, privacy/security, evals, deterministic inference, RL environments, open source, and where the frontier is headed.TIMESTAMPS00:00:00 - Intel Monopoly, Chips & Compilers00:15:10 - OpenClaw, Guardrails & Harnesses00:25:45 - First Principles of Writing Code00:32:18 - Privacy & Security While Running AI Systems00:39:10 - Evals, Benchmarks & Taste00:44:15 - Deterministic vs Non-Deterministic Inference & Floating Point Math01:03:45 - Self-Improving Agents, RL Environments & Rollouts01:15:10 - AGI, ASI & Continual Learning01:20:50 - Open Source & The Chinese Frontier01:26:30 - The K-Shaped Economy & Why Curiosity Is the Frontier

Vik is the Co-founder of Moondream AI.TIMESTAMPS00:00:00 - Vik’s Story & Moondream00:18:50 - The Core Thesis of Small & Scale00:28:35 - The Data Problem00:33:35 - Deciding the Training Architecture00:43:20 - Post-Training & RL Performance00:46:06 - Post-Training Recipes of Moondream 300:47:40 - Open Source & VLM Development Priorities00:52:05 - AI War: America vs China00:55:07 - Moondream Acquisition & The Future01:04:08 - Community Questions01:16:40 - Trivia, Lore & Opinions on the Current State of AI01:28:07 - Advice to Your 20-Year-Old Self

Dhravya Shah is the Founder of Supermemory.TIMESTAMPS00:00:00 - Teaser00:01:39 - Introduction00:02:42 - What is SuperMemory? Explaining the Product00:04:43 - Coolest Use Cases & Customer Stories00:07:48 - Early Days: Growing Up in Mumbai & Learning to Code00:09:24 - First Success: Discord Bot & Twitter Screenshot Tool Acquisition00:13:11 - The IIT Story: Myth vs Reality00:14:38 - The 40-Week Building Streak00:17:37 - Learning Strategy & Resources00:20:18 - From AnyContext to SuperMemory: The Origin Story00:21:16 - Failed Projects & Lessons Learned00:25:04 - Getting Attacked & Accidentally Joining Cloudflare00:26:30 - Relationship Support & Building While in College00:27:57 - How to Sell Your Projects & Acquisitions00:29:23 - Working at Mem0 & Differences with SuperMemory00:33:51 - Cloudflare Experience & Working with CEO Dane Knecht00:36:00 - The Fundraising Journey: From Buildspace to a $3M Round00:40:51 - Why Skip Y Combinator?00:42:16 - O-1 Visa Story: Becoming “Officially Extraordinary”00:44:14 - Being a Solo Founder: Challenges & Benefits00:47:20 - Hiring Philosophy & Team Culture at SuperMemory00:51:46 - India vs Bay Area: Ecosystem Differences00:53:10 - Vision vs Profit: What Matters at the Early Stage00:54:26 - Thoughts on Joining College00:55:18 - What’s Next for SuperMemory (Local-First & Nova)00:57:38 - Advice for Aspiring Builders & Students00:59:00 - Closing Thoughts

Pierre-Carl Langlais (aka Alexandar Doria) is Co-founder of Pleias. We'd discussed about pre-training recipes, common corpus, mid-training, agentic systems, good post-training and everything AI.TIMESTAMPS00:00:00 - TEASER00:01:12 - INTRO00:02:03 - Who is Alexander Doria [Pierre-Carl Langlais]? 00:04:10 - Early career: From humanities to AI research00:07:50 - Meeting influential people in computational humanities00:10:00 - How the idea of Pleias came about00:13:30 - Building Pleias: Infrastructure and compute challenges in Europe00:17:06 - Team structure and work culture at Pleias00:19:06 - What is "open science" and why it matters00:21:53 - Big announcement: OpenSynthetic initiative00:25:25 - Synthetic data experiments and surprising results00:28:11 - "The Model is the Product" - explained00:31:56 - Implications for companies building on top of models00:35:25 - Differentiation in a world of shared base models00:38:40 - Common Corpus: Origins and development00:44:12 - The lack of open, legally clear datasets00:47:03 - Anthropic's use of Common Corpus for mechanistic interpretability00:50:20 - What makes good post-training?00:54:00 - Reasoning under 400M parameters in SLMs00:56:35 - Generalist scaling is stalling - where are the diminishing returns?00:59:40 - Will specialization always win over scale?01:02:00 - Opinionated and task-specialized models01:06:29 - How inference cost drops change monetization models01:09:12 - New value layers beyond token marketplaces01:11:38 - Major technical obstacles to embedding workflows in models01:13:40 - How smaller labs can compete on training infrastructure01:15:36 - Should startups raise capital for AI training?01:17:16 - What new capabilities do models need for orchestration?01:19:50 - Designing verifier functions for agentic models01:22:17 - RL in domains with weak or delayed rewards01:24:50 - Multi-step training loops: Draft, verify, refine, backtrack01:26:38 - The scarcity of agentic data and bootstrapping solutions01:29:32 - Making agent training tractable at scale01:31:44 - What is mid-training and why it matters01:34:55 - Deployment, use cases, and hybrid model architectures01:37:37 - Human-in-the-loop for regulated domains01:39:48 - Advice for startups positioning in this transition01:41:58 - Europe's structural challenges in AI01:45:52 - Tokenizers: The overlooked competitive frontier01:49:59 - Training LLMs on personal data and dead languages01:52:12 - World models and JEPA architectures01:53:50 - Building agentic systems: Stack and RL environments01:55:34 - The art of training good RL models01:58:49 - Trivia: Underrated habits and mindsets in research02:00:09 - AI Twitter community and its impact02:01:40 - Advice for folks starting in AI research02:03:27 - Final thoughts and wrap-up

Kalomaze is 20 yo Researcher at Prime IntellectTIMESTAMPS00:00:00 - Teaser00:01:06 - Intro00:01:49 - Organic Influence00:02:46 - “Extremely Silly Jester”00:04:04 - School Days & Diving into ML Research00:09:27 - Developing “Research Taste” & The Art of Selection00:11:45 - Dropping College, Parents’ Reaction & The Plan Ahead00:18:13 - How a Reddit Post Led to a Job at Shopify00:23:53 - Why He Chose Prime Intellect Over Other Offers00:29:27 - A Day in the Life: Workload at PI00:31:53 - Verifiers, GRPO & Semi-Verifiable Rewards00:40:02 - Progress on Synthesizers00:43:28 - The Environment Hub: A Hugging Face for RL?00:47:47 - How Research Ideas Emerge00:51:38 - Defining “Taste” in Research00:54:15 - The Future of Supervised Fine-Tuning01:01:48 - The Flaw in “Pure Reasoner” Models & Thoughts on GPT-OSS01:05:34 - Perspective on the Chinese Open-Source AI Surge01:07:58 - The Art of Training a Good RL Model01:10:08 - How to Learn a New Field: A Hacker’s Approach01:12:33 - Handling Disagreements with Experienced Researchers01:14:49 - Recipe for RLHF01:19:47 - Scaling Bigger Models vs. Designing Better Rewards01:26:31 - Rethinking Progress & The Post-AGI Narrative01:29:27 - The Most Unexpected Part of Working at Prime Intellect01:30:09 - Trivia Round Begins01:30:25 - Code Reviews from Will?01:33:43 - A Hidden Secret About Will01:34:26 - The Underrated Mindset That Gives Him an Edge01:36:18 - Take on AI Tech Twitter / TPOT01:38:16 - How Kalomaze Shaped Up + Advice01:41:17 - Podcast Wrap-Up

Daniel Han is Co-founder and CEO of Unsloth.(00:00:00) - TEASER(00:01:00) - INTRO(00:02:10) - Early Career(00:04:47) - Hyperlearn and Sciblox(00:09:30) - Bug Bounty Guy(00:12:32) - Questions for Daniel(00:30:20) - Scaling RL, GRPO and Transformer moment(00:42:35) - Unsloth: GTM, YC aid, next big thing and more(00:59:10)- General: learning process, multi-gpu, advices and more

Early July, X users have shared dozens of stories about encounters with Soham Parekh, a software engineer who seems to have been simultaneously working at multiple Silicon Valley startups. We'd discussed his whole journey, experiences and reflections in this episode. Hope you'll find it insightful.(00:00:00) - TEASER(00:00:45) - INTRO(00:02:40) - Define Soham Parekh, Early Career(00:05:10) - Current routine and Interests(00:06:38) - Soham - a Generalist?(00:08:05) - How has been life since twitter blowout?(00:09:17) - Academics, Dropping out, Grad School(00:10:38) - Acing Tech Interviews, Open Source(00:11:44) - Curious about Low-level internals(00:14:30) - Mistakes, Tutorial hell, Interviews(00:17:12) - What was the first immediate reaction?(00:18:45) - Reaction of family and friends(00:20:10) - What were the misconceptions floating around?(00:21:15) - Dev-shop of 20 Interns?, Interviews, Contributions(00:25:03) - Did Soham give any company his 100%, working parallely?(00:27:55) - How picky Soham was choosing startups to interview?(00:29:58) - How was a typical day with peak workload?(00:31:20) - Why should anyone trust your words/work ethic?(00:32:32) - Incurred interest on integrity, Work Trials(00:35:10) - Crazy story with a startup(00:36:25) - How much Soham used Cursor/AI tools with jobs?(00:38:25) - Did Soham interview at Frontier AI Labs?(00:40:02) - Sharing MVPs among different companies?(00:42:38) - Mutual understanding/friendships with founders?(00:44:38) - Managing onboarding at multiple startups?(00:48:12) - Surviving background checks(00:49:56) - Difference between culture in SF and India (00:52:08) - Thoughts on Agentic AI space(00:58:13) - Which one startup excited Soham most?(01:01:00) - Contradictory statement in TBPN Interview, self perception now(01:02:14) - Negative stereotype reinforced for remote work(01:04:45) - What is one thing startups are completely missing now?(01:07:15) - What kind of tech product Soham will build in current era?(01:09:30) - How would Soham react with moonlighting as a CEO?(01:12:05) - What excites Soham more with current startup (Darwin)(01:14:05) - Final thoughts(01:15:21) - CHEERS!

Will Brown is Research Lead at Prime Intellect.TIMESTAMPS(00:00:00) - INTRO(00:02:00) - Career Trajectory, Love for Maths and Algorithms(00:06:24) - Early Motivation and PhD experience(00:12:22) - Teaching, Open Source, GenAI Handbook(00:17:27) - Rubrik Engineering and Approach to Design(00:23:06) - RL over Instruction Tuning(00:26:07) - Journey behind VERIFIERS (00:33:30) - RL experiments on SLMs(00:37:06) - Decentralized Training and Inference(00:40:28) - Self-Improving agents(00:44:05) - GRPO in Multimodal settings, Robotics(00:49:00) - Challenges in Curriculum Learning(00:52:02) - Kimi Dev 72B?(00:54:55) - Reinforcement Pre-training(00:57:53) - Do LLMs have hit a wall (in novelties)?(01:02:06) - LLM and Classical Agents(01:05:10) - LLMs as Evaluators or Doers(01:07:19) - Do models know when they're Reward Hacking?(01:12:42) - Training agents in diverse environments(01:14:55) - What if solving reward hacking accidentally gives us alignment?(01:16:28) - Will AGI collapse decentralized system into centralized one?(01:19:24) - The Illusion of the Illusion of Thinking(01:20:21) - LLMs in Quant Trading(01:22:21) - Thoughts on Pattern Matching in Intelligence?huhhh, let's relax!(01:23:55) - Kalomaze's hidden secret(01:24:30) - Which frontier lab Will is bullish on?(01:26:57) - Finest teams in AI Research(01:30:45) - Is xAI late compared to other frontier labs?(01:32:35) - How to become an Experimentalist?(01:33:49) - Underrated mindset that has been an edge in Will's research?(01:34:41) - "Cheat Code" in RL Research that has been game changing?(01:35:45) - What Will think about AGI?(01:38:01) - Will's advice to 20yo high school/college grads(01:39:23) - Classical ML in current age of AI Research(01:41:10) - What Will have to say for a general audience?(01:42:29) - Concluding!

Tokenbender on Post-Training, RL and Reasoning, Experiments, Post AGI Landscape, Indic AI and ExperimentsTokenbender [Guest]: https://x.com/tokenbender(00:00:00) - INTRO(00:01:30) - Career Trajectory and Motivation(00:10:00) - Non-CS Background and Building Intuitions(00:14:00) - Journey with Codecherrypop and Small Models(00:21:50) - Partner-In-Crime and Roleplay Series(00:28:45) - Post-Training and how it is evolved?(00:36:45) - Is pre-training actually dead?(00:45:10) - RL over next-token-predictors?(00:52:10) - Reliable agents, RL in training workload(00:58:07) - Weak priors and Reward Sparsity(01:01:20) - What's new RL sauce?(01:06:11) - RL from Zero Pre-train, Coherent text and Beyond(01:12:01) - Intelligence isn't flat, Optimizing for one sharp spike?(01:16:37) - Sampling and Creating Data for Models, New approaches?(01:20:55) - Role of failures(01:24:26) - Obsession over next number(01:27:05) - Shallow safety alignment(01:31:58) - RL over Diffusion Models, 'aha' moments(01:37:40) - 50x in productivity?(01:40:18) - How do you build the mindset to keep experimenting?(01:48:20) - Writing papers on AI research(01:51:10) - How you look up to open source models, what next?(01:53:45) - Finding or Creating synthetic datasets(01:56:08) - TRIVIA(02:07:06) - Indic AI Landscape, Challenges(02:12:32) - ADVICE FOR STUDENTS(02:16:50) - FINAL THOUGHTS FROM TOKENBENDER