GroundZero AI Talks

Himanshu Dubey

Your friendly neighborhood creative space shaping the frontier of tech, with occasional conversations and notes.

  1. The Delta of Intelligence is Human Data | Curtis Northcutt (Director of AI Research, Handshake)

    16 MAY

    The Delta of Intelligence is Human Data | Curtis Northcutt (Director of AI Research, Handshake)

    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. TIMESTAMPS 00:00:00 - Intro 00:02:08 - The Snowman Effect: Growing Up in Rural Kentucky to 7 Years at MIT 00:09:50 - Confident Learning, Two Systems of Intelligence & Why Classical ML Still Matters 00:20:37 - The Origin of Cleanlab & the Handshake Acquisition 00:28:00 - 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 Data 00:38:17 - New Kinds of Data, IA & AI 00:42:04 - Scaling Coding Benchmarks & Efficiency in Long-Horizon Tasks 00:49:12 - Pre-Training & Misconceptions Around Human Data 00:57:13 - Open Questions: Taste, Personality & Quick Fire 01:05:15 - Advice to Your 20-Year-Old Self: Skill & Obsession

    1hr 8min
  2. The Never Ending Lore of Harness | Vivek Trivedy (Product Lead, Langchain)

    16 MAY

    The Never Ending Lore of Harness | Vivek Trivedy (Product Lead, Langchain)

    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. TIMESTAMPS 00:00:00 - Intro 00:03:30 - PhD at Temple & First Startup in Vision 00:12:42 - Bio Internship at Lockheed Martin & Mike Mill Lore 00:16:18 - Building at LangChain, DeepAgents & R&D in Open Source 00:25:09 - Harness Design, Filesystems & Foundational Harness Primitives 00:42:25 - Trajectories for Continual Learning, Where Not to Use RL, Skills & Context Rot 00:56:30 - Harness Engineering: OpenClaw, Hermes, Custom Harnesses & Quick Fire 01:12:15 - Meta-Harnesses, Self-Improvement Loops, Simulation-as-a-Service & RL Environments 01:28:30 - Advice to Anyone 20 / Starting Out in College

    1hr 33min
  3. The Paradigm of Autonomous Research | Francesco Pappone (CEO, Paradigma)

    16 MAY

    The Paradigm of Autonomous Research | Francesco Pappone (CEO, Paradigma)

    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. TIMESTAMPS 00:00:00 - Intro 00:01:03 - Dropping Out of a PhD, Saturation in Academia & Building from Europe 00:13:15 - AutoResearch, The Idea Behind Flywheel & the ex-OpenAI Co-Founder 00:19:18 - Evals for AutoResearch & Research Taste 00:25:00 - How Research Labs Use Flywheel 00:31:15 - The Thesis Behind AutoResearch & Adoption Across Top Labs 00:41:00 - Working with Flywheel & the Combinatorial Explosion Problem 01:02:03 - Science as a Public Graph, New Research Loops & Quick Fire 01:06:23 - The Toughest AutoResearch Challenge Today & Future Directions 01:11:53 - Automating Science: Sakana, DeepMind & the Trust Problem 01:21:07 - Business Model, Automating R&D & Model Capabilities 01:27:20 - Dropping Out of College to Live on the Frontier

    1hr 31min
  4. Building Frontier Open Reasoning Models | Lucas and Varun (Arcee AI)

    16 MAY

    Building Frontier Open Reasoning Models | Lucas and Varun (Arcee AI)

    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. TIMESTAMPS 00:00:00 - Intro 00:00:59 - Varun’s Transition from SWE to Pre-Training Lead 00:04:20 - The Trinity Manifesto & OpenClaw Ecosystem 00:12:15 - Arcee’s Post-Training to Pre-Training Pivot 00:23:45 - Varun’s First Pre-Training Run (“You Can Just Do Things”) 00:27:33 - Saturation in Pre-Training? Mid-Training Explained 00:37:00 - Tweaking Training Architecture, Adam vs Muon & Evals 01:09:07 - Inference Engineering, Quick Fire & Post-Training Recipes 01:18:02 - Alpha in RL Environments & Harness Design 01:23:00 - Why American Open Source Is Trailing Chinese Competitors & Trinity Adoption 01:29:25 - Hiring at Arcee & Advice to Your 20-Year-Old Self

    1hr 39min
  5. Agents, Verifiability and First Principles Thinking | Vishnu Patankar (CTO, EigenCloud)

    16 MAY

    Agents, Verifiability and First Principles Thinking | Vishnu Patankar (CTO, EigenCloud)

    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. TIMESTAMPS 00:00:00 - Intel Monopoly, Chips & Compilers 00:15:10 - OpenClaw, Guardrails & Harnesses 00:25:45 - First Principles of Writing Code 00:32:18 - Privacy & Security While Running AI Systems 00:39:10 - Evals, Benchmarks & Taste 00:44:15 - Deterministic vs Non-Deterministic Inference & Floating Point Math 01:03:45 - Self-Improving Agents, RL Environments & Rollouts 01:15:10 - AGI, ASI & Continual Learning 01:20:50 - Open Source & The Chinese Frontier 01:26:30 - The K-Shaped Economy & Why Curiosity Is the Frontier

    1hr 33min
  6. Tiny models will run the World | Vikhyat K (Co-founder, Moondream AI)

    16 MAY

    Tiny models will run the World | Vikhyat K (Co-founder, Moondream AI)

    Vik is the Co-founder of Moondream AI. TIMESTAMPS 00:00:00 - Vik’s Story & Moondream 00:18:50 - The Core Thesis of Small & Scale 00:28:35 - The Data Problem 00:33:35 - Deciding the Training Architecture 00:43:20 - Post-Training & RL Performance 00:46:06 - Post-Training Recipes of Moondream 3 00:47:40 - Open Source & VLM Development Priorities 00:52:05 - AI War: America vs China 00:55:07 - Moondream Acquisition & The Future 01:04:08 - Community Questions 01:16:40 - Trivia, Lore & Opinions on the Current State of AI 01:28:07 - Advice to Your 20-Year-Old Self

    1hr 31min
  7. The Story of Dhravya Shah | 20yo raised $3M to build SuperMemory, Dhravya Shah

    16 MAY

    The Story of Dhravya Shah | 20yo raised $3M to build SuperMemory, Dhravya Shah

    Dhravya Shah is the Founder of Supermemory. TIMESTAMPS 00:00:00 - Teaser 00:01:39 - Introduction 00:02:42 - What is SuperMemory? Explaining the Product 00:04:43 - Coolest Use Cases & Customer Stories 00:07:48 - Early Days: Growing Up in Mumbai & Learning to Code 00:09:24 - First Success: Discord Bot & Twitter Screenshot Tool Acquisition 00:13:11 - The IIT Story: Myth vs Reality 00:14:38 - The 40-Week Building Streak 00:17:37 - Learning Strategy & Resources 00:20:18 - From AnyContext to SuperMemory: The Origin Story 00:21:16 - Failed Projects & Lessons Learned 00:25:04 - Getting Attacked & Accidentally Joining Cloudflare 00:26:30 - Relationship Support & Building While in College 00:27:57 - How to Sell Your Projects & Acquisitions 00:29:23 - Working at Mem0 & Differences with SuperMemory 00:33:51 - Cloudflare Experience & Working with CEO Dane Knecht 00:36:00 - The Fundraising Journey: From Buildspace to a $3M Round 00:40:51 - Why Skip Y Combinator? 00:42:16 - O-1 Visa Story: Becoming “Officially Extraordinary” 00:44:14 - Being a Solo Founder: Challenges & Benefits 00:47:20 - Hiring Philosophy & Team Culture at SuperMemory 00:51:46 - India vs Bay Area: Ecosystem Differences 00:53:10 - Vision vs Profit: What Matters at the Early Stage 00:54:26 - Thoughts on Joining College 00:55:18 - What’s Next for SuperMemory (Local-First & Nova) 00:57:38 - Advice for Aspiring Builders & Students 00:59:00 - Closing Thoughts

    1hr 42min
  8. Model is the Product | Common Corpus, Mid-Training, Open Science | Pierre-Carl Langlais, Pleias

    16 MAY

    Model is the Product | Common Corpus, Mid-Training, Open Science | Pierre-Carl Langlais, Pleias

    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. TIMESTAMPS 00:00:00 - TEASER 00:01:12 - INTRO 00:02:03 - Who is Alexander Doria [Pierre-Carl Langlais]? 00:04:10 - Early career: From humanities to AI research 00:07:50 - Meeting influential people in computational humanities 00:10:00 - How the idea of Pleias came about 00:13:30 - Building Pleias: Infrastructure and compute challenges in Europe 00:17:06 - Team structure and work culture at Pleias 00:19:06 - What is "open science" and why it matters 00:21:53 - Big announcement: OpenSynthetic initiative 00:25:25 - Synthetic data experiments and surprising results 00:28:11 - "The Model is the Product" - explained 00:31:56 - Implications for companies building on top of models 00:35:25 - Differentiation in a world of shared base models 00:38:40 - Common Corpus: Origins and development 00:44:12 - The lack of open, legally clear datasets 00:47:03 - Anthropic's use of Common Corpus for mechanistic interpretability 00:50:20 - What makes good post-training? 00:54:00 - Reasoning under 400M parameters in SLMs 00:56:35 - 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 models 01:06:29 - How inference cost drops change monetization models 01:09:12 - New value layers beyond token marketplaces 01:11:38 - Major technical obstacles to embedding workflows in models 01:13:40 - How smaller labs can compete on training infrastructure 01:15:36 - 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 models 01:22:17 - RL in domains with weak or delayed rewards 01:24:50 - Multi-step training loops: Draft, verify, refine, backtrack 01:26:38 - The scarcity of agentic data and bootstrapping solutions 01:29:32 - Making agent training tractable at scale 01:31:44 - What is mid-training and why it matters 01:34:55 - Deployment, use cases, and hybrid model architectures 01:37:37 - Human-in-the-loop for regulated domains 01:39:48 - Advice for startups positioning in this transition 01:41:58 - Europe's structural challenges in AI 01:45:52 - Tokenizers: The overlooked competitive frontier 01:49:59 - Training LLMs on personal data and dead languages 01:52:12 - World models and JEPA architectures 01:53:50 - Building agentic systems: Stack and RL environments 01:55:34 - The art of training good RL models 01:58:49 - Trivia: Underrated habits and mindsets in research 02:00:09 - AI Twitter community and its impact 02:01:40 - Advice for folks starting in AI research 02:03:27 - Final thoughts and wrap-up

    2h 5m
See All (13)

About

Your friendly neighborhood creative space shaping the frontier of tech, with occasional conversations and notes.

Information

  • Creator
    Himanshu Dubey
  • Years Active
    2k
  • Episodes
    13
  • Rating
    Clean
  • Copyright
    © Himanshu Dubey
  • Show Website
    GroundZero AI Talks