Podcast
The MAD Podcast with Matt Turck
Matt Turck talks with builders, investors, and leaders across machine learning, AI, and data infrastructure.
Original source“OpenAI’s Model Hacked Us” - Hugging Face’s Thomas Wolf
Thomas Wolf describes how an OpenAI-powered agent unexpectedly attacked Hugging Face during cyber testing, then explains why open-source models were the only ones willing to help in the live response. The conversation broadens into the changing safety debate around open vs. closed models, agent deception, enterprise adoption, AI sovereignty, and whether frontier AI should slow down.
How to Build Long-Horizon AI Agents — Mitch Troyanovsky, Basis
Mitch Troyanovsky argues that long-horizon AI agents should be judged on process, not just outcomes, especially in messy domains like accounting and tax prep. He explains how Basis uses behavior specs, judges, ontology, and runtime context engineering to make autonomous agents reliable enough for real production work.
The Biggest AI Deployment Nobody Talks About | Samsara CEO Sanjit Biswas
Samsara CEO Sanjit Biswas argues the biggest underappreciated AI deployment is in physical operations: trucks, equipment, roads, and industrial assets. He explains how Samsara combines sensors, edge AI, cloud software, and agents to improve safety, automate workflows, and manage mixed human-machine operations at massive scale.
The Biggest Chip Ever Built — Why OpenAI Runs On It | Cerebras CEO Andrew Feldman
Andrew Feldman argues that AI is shifting from training to inference, where speed and latency matter more than raw model size. He makes the case that Cerebras’s wafer-scale chip, memory architecture, and cloud offerings are built for fast tokens, reasoning, and agentic workloads.
OpenAI’s Compute Chief: We Can’t Build Fast Enough | Sachin Katti
Sachin Katti says OpenAI is racing to build compute faster than the physical world can supply power, chips, cooling, and labor. He frames AI data centers as giant liquid-cooled factories turning electrons into tokens, with custom silicon, new power infrastructure, and networking all becoming strategic bottlenecks.
Stripe's AI Chief: How AI Agents Will Buy, Sell, and Pay
Stripe’s Emily Sands says agentic commerce is moving from theory to production, with protocols, wallets, and real-time billing infrastructure already being built. She also argues that token theft, first-party abuse, and deployment friction are emerging as the key constraints on the AI economy.
Inside Nemotron & NVIDIA’s AI Lab | Bryan Catanzaro
Bryan Catanzaro explains why NVIDIA builds and gives away open AI models: to understand future systems better and strengthen the open ecosystem. He also details Nemotron’s engineering choices, from 4-bit pretraining and hybrid architectures to long context, multi-token prediction, and MoE-heavy systems design.
Cloudflare CEO: The Internet's Business Model Is Dead
Matthew Prince argues Cloudflare is seeing the internet tip from human-driven to bot-driven traffic, and that this will break the ad-supported web economy. He also explains how Cloudflare’s edge network, Workers, AI Gateway, and internal AI tools are being rebuilt for an agentic internet.
The GPU Myth: State of AI Compute 2026 | Stephen Balaban
Stephen Balaban argues GPU compute is not becoming a commodity; instead, AI infrastructure is becoming more vertically integrated around land, power, cooling, networking, software, and financing. He also traces Lambda’s evolution from an early facial-recognition startup into a near-$1B cloud business and lays out a long-term vision where AI becomes the software layer itself.
OpenAI's Dan Roberts: Why AI Can Now Make Discoveries
Dan Roberts argues that modern reinforcement learning and test-time compute are turning language models into systems that can reason, explore, and sometimes make genuine scientific discoveries. He uses OpenAI’s recent math work, RLHF, and physics-inspired thinking to explain why AI progress looks smooth, why scale alone is not enough, and why future models may help drive science itself.