Podcast
Training Data
Sequoia Capital conversations with AI founders, researchers, and builders about company building and technology shifts.
Original source
Google's AI Infrastructure Chief, Amin Vahdat, on the Physics & Economics of Frontier AI
Amin Vahdat argues that frontier AI infrastructure should be measured by goodput, not FLOPS, because large accelerator fleets fail constantly and workload-level performance depends on the whole system. He also walks through Google’s TPU strategy, power and networking constraints, optical switching, and why the long-term shape of AI data centers may include more specialized, tightly integrated racks and even orbital compute.
Box's Aaron Levie: On Reinventing Yourself in the AI Age and Enterprise Diffusion
Aaron Levie argues that Box’s AI opportunity is not just model access, but the application layer that connects models to real enterprise workflows, permissions, and data. He says Box is building a model-agnostic agent harness for document-heavy work and predicts that most enterprise tokens will soon be generated by agents users never directly start.
Making Cities Awesome: Peregrine’s Nick Noone & Ben Rudolph
Peregrine’s founders argue that public-safety software should connect the data cities already own, not collect more of it, and that safety, privacy, and local data sovereignty have to be designed in from the start. They also show how forward-deployed engineering and AI agents can turn messy municipal data into search, analysis, and even new operational tools for police, fire, EMS, and emergency management.
Parallel’s Parag Agrawal: Building a New Web for AI Agents
Parag Agrawal says Parallel is rebuilding web search for agents, not humans, by treating click data as a bug and optimizing for agent feedback, latency, and token efficiency. He also argues the web’s ad economy will need a new pricing model for AI-era usage, likely built on Shapley-style attribution and differential pricing for content owners.
Rich Sutton and Khurram Javed: Why AI Models Stop Learning, and How to Start It Again
Rich Sutton and Khurram Javed argue that intelligence should be understood as continual learning, not a one-time training phase. They say today’s LLMs are powerful but fundamentally incomplete, and they outline a research program for agents that keep updating after deployment without catastrophic forgetting.
Chai Discovery's Bitter Lesson: Drug Design Is Another Scaling Problem
Chai Discovery argues drug design is a scaling problem: build simpler models, scale data and compute, and verify results in the lab. The episode centers on Chai-2’s jump in antibody hit rate from roughly 0.1% to 15% and the company’s plan to turn discovery into a computer-aided design loop.
Building the Automated AGI Lab: Core Automation's Jerry Tworek and Rohan Anil
Core Automation’s Jerry Tworek and Rohan Anil argue that transformers have reached an architectural ceiling and that the next frontier is continual learning, test-time adaptation, and tighter end-to-end optimization. They also describe an automation-first research lab strategy, starting with kernel generation and GPU performance as the key bottleneck.
Factory's Matan Grinberg: The Coming ‘Dark Factory’ Where Software Builds Itself
Matan Grinberg says Factory bet early on autonomous coding agents, endured years of weak product-market fit, and eventually rebuilt around a model-agnostic router and CLI that developers actually wanted. He argues the future of software is a “dark factory” of asynchronous agents, with open models taking most token share and enterprises optimizing token spend like headcount.
Anthropic's Katelyn Lesse & Angela Jiang: Building an Ecosystem, not a Walled Garden
Anthropic’s platform team is building a layered developer stack for both internal products and external builders, with a roadmap that moves from knowledge and execution to coordination strategies. The guests argue for an open ecosystem built on standards like MCP and interoperable sandboxes, while keeping harnesses tuned to Claude rather than treating models as freely swappable commodities.
Inside Zipline's Autonomous System: 140M Miles, Zero Incidents
Zipline’s founders argue the company is not a drone business but an autonomous logistics infrastructure layer, with the aircraft only a small part of a much larger system. The episode traces how that system scaled from Rwanda blood deliveries to millions of deliveries, zero safety incidents, and a path toward delivery economics that can beat cars.
Why Hardware-Software Co-Design Is AI's Real 100x: Dylan Patel of SemiAnalysis
Dylan Patel argues that AI’s biggest gains come from hardware-software-model co-design, not raw chip speed alone. He also makes the case that inference is becoming a giant market, and that benchmarking, supply chains, and data-center economics will determine who wins.
Memory and Continual Learning: Engram's Dan Biderman and Jessy Lin
Engram’s Dan Biderman and Jessy Lin argue that AI’s real bottleneck is memory and continual learning, not raw intelligence. Their thesis is to bake company-specific knowledge into model weights so teams get faster, cheaper, more context-aware models that improve over time.
Simulating Humans at Scale: Simile's Joon Sung Park
Joon Sung Park argues Simile is building a simulation layer for human society, not a superhuman reasoning model, by training on real behavioral data, interviews, surveys, and RCTs. He says the company can already use these models for concept testing, earnings-call simulation, and broader social-science questions, with a long-term vision of modeling large-scale societal dynamics.
Google DeepMind's Logan Kilpatrick: Why the Model Eats the Harness
Logan Kilpatrick argues Google’s next platform shift is from model-centric AI to agent harnesses that the model will increasingly absorb. He also says coding is already a form of narrow superintelligence, while Google’s multimodal Omni effort aims to collapse many separate media systems into one model.