The AI Black Box Might Finally Be Opening w/ Eric Ho of Goodfire

AI models may not just be predicting words. They may be building rich internal worlds made of features, circuits, and curved geometry. In this episode of The Neuron: AI Explained, Corey Noles and Grant Harvey sit down with Eric Ho, Cofounder & CEO of Goodfire, to talk about what’s actually happening inside neural networks — and why interpretability could be the key to safer, more reliable, and more useful AI. Goodfire is working on AI interpretability: using AI to understand, debug, edit, and eventually design neural networks with more intention. Eric explains why models may “think in shapes,” how neural geometry reveals hidden structure inside models, and why those structures could help researchers reduce hallucinations, steer behavior, improve training data, and build better AI systems. In this conversation: • Why neural networks may contain meaningful internal structures • What “neural geometry” actually means • How models can represent concepts like rabbit ears, confidence, or hallucination • Why interpretability could make model training less trial-and-error • How Goodfire’s Silico aims to help researchers understand and train models • What this means for robotics, multimodal models, safety, and the future of AI Learn more about Goodfire: goodfire.ai Request access to Silico: goodfire.ai Subscribe to The Neuron for more AI conversations that skip the hype and explain what’s actually happening: https://www.theneuron.ai/ Sponsored by SAS AI Governance: Visit https://www.sas.com/en_us/solutions/a... The Neuron Academy helps professionals build practical AI skills they can use right away, with lessons on prompting, workflows, and real workplace use cases. Check out https://theneuronacademy.com/ today! ➤ CHAPTERS 0:00 - Inside the Mind of a Neural Network 2:21 - Intro: SAS AI Governance 2:34 - Eric Ho Joins The Neuron 2:56 - What Goodfire Is Building 4:08 - Using AI to Interpret AI 5:48 - Discovering Concepts Inside Models 6:40 - A message from our sponsor: SAS AI Governance 8:04 - Neurons, Activations, and Neural Geometry 10:18 - Why Models May Be More Interpretable Than Expected 12:30 - Reducing Hallucinations with Feature Rewards 13:30 - The Model's Hidden Inner World 14:31 - Are Models Simulating Intelligence? 19:04 - Compression, Generalization, and Synthetic Intelligence 20:03 - Models May Think in Shapes 20:51 - Check out The Neuron Academy 22:29 - Block Sparse Featurizers and Manifolds 26:01 - Why Linear Representations Aren't Enough 27:41 - The Mountain Car Steering Example 30:09 - Shape Rotators vs. Word People 33:12 - Intentional Design for AI Models 34:13 - Predictive Data Debugging 37:04 - Who Silico Is For 39:00 - Finding Confidence and Hallucination Structures 40:23 - Steering Model Behavior 41:49 - Continual Learning and Safety Guardrails 44:13 - Why More Companies May Train Their Own Models 45:49 - The Biggest Surprise in Interpretability 48:40 - Designing Models Like Software 50:31 - Neural Geometry for Robotics 52:52 - Interpretability Across Modalities 54:13 - Shared Representations in Multimodal Models 56:04 - Where to Learn More About Goodfire 56:57 - Outro: SAS AI Governance 57:02 - Subscribe to The Neuron Hosted by: Corey Noles and Grant Harvey Guest: Eric Ho, Cofounder & CEO of Goodfire Published by: Manique Santos Edited by: Adrian Vallinan