DeepMind's Pushmeet Kohli on AI's Scientific Revolution

Thomas Wolf, co-founder and Chief Science Officer of Hugging Face, explains how his company is applying the same community-driven approach that made transformers accessible to everyone to the emerging field of robotics. Thomas discusses LeRobot, Hugging Face's ambitious project to democratize robotics through open-source tools, datasets, and affordable hardware. He shares his vision for turning millions of software developers into roboticists, the challenges of data scarcity in robotics versus language models, and why he believes we're at the same inflection point for physical AI that we were for LLMs just a few years ago. Hosted by: Sonya Huang and Pat Grady, Sequoia Capital

Ryan Daniels and John Sarihan are reimagining legal services by building Crosby, an AI-powered law firm that focuses on contract negotiations to start. Rather than building legal software, they've structured their company as an actual law firm with lawyers and AI engineers working side-by-side to automate human negotiations. They've eliminated billable hours in favor of per-document pricing, achieving contract turnaround times under an hour. Ryan and John explain why the law firm structure enables faster innovation cycles, how they're using AI to predict negotiation outcomes, and their vision for agents that can simulate entire contract negotiations between parties. Hosted by Josephine Chen, Sequoia Capital Mentioned in this episode: Data processing agreement (DPA): GDPR-mandated contract between controllers and processors. Crosby handles DPAs as part of B2B contracting. Credence good: Economic term for services whose quality is hard to judge even after consumption. Used to explain why legal buyers value lawyers-in-the-loop and malpractice coverage.

When the AI wave hit, n8n founder Jan Oberhauser faced a critical choice: become irrelevant or become indispensable. He chose the latter, transforming n8n from a simple workflow tool into a comprehensive AI automation platform that lets users connect any LLM to any application. The result? Four times the revenue growth in eight months compared to the previous six years. Jan explains how n8n’s “connect everything to anything” philosophy, combined with a thriving open source community, positioned the company to ride the AI automation wave while avoiding vendor lock-in that plagues enterprise software. Hosted by George Robson and Pat Grady, Sequoia Capital Mentioned in this episode: Model Context Protocol (MCP): Open protocol that lets AI models safely use external tools and data that is used extensively by n8n for orchestration. Vector database: A database optimized for storing and searching embeddings. These “vector stores” can pair with LLMs for retrieval-augmented workflows. Granola: AI productivity tool mentioned by Jan as a recent favorite.  Her: A film that Jan says, “a few years ago, it was sci fi, and it’s now suddenly this thing that is just around the corner.”

Before ChatGPT made AI mainstream, John Noronha was building Gamma with a simple insight: everyone hates making slides but needs visual communication for high-stakes ideas. His background at Optimizely proved crucial as Gamma became a testing laboratory for AI models, running hundreds of experiments to discover that Claude excels at creative taste, Gemini wins on cost efficiency, and reasoning models actually hurt creativity. John explains how solving their own blank page problem inadvertently solved it for millions of users, turning a near-failing startup into a cash flow positive platform with 250 million presentations created. He discusses competing with PowerPoint's 500 million users while expanding beyond slides into documents, websites and visual storytelling. Hosted by Sonya Huang, Sequoia Capital

Dara Ladjevardian, founder and CEO of Delphi, is creating digital minds that allow people to scale their thoughts and availability without replacing human connection. Inspired by Ray Kurzweil’s theory of mind as a hierarchy of pattern recognizers, Dara built an adaptive temporal knowledge graph that captures how people think and reason. From helping CEOs train new hires to enabling coaches to monetize their expertise 24/7, Delphi represents a new form of conversational media. Dara explains why authentic human representation matters, how digital minds actually increase desire for real human connection, and why he believes 2026 will be the tipping point for adoption for digital minds. Hosted by Sonya Huang and Jess Lee, Sequoia Capital Mentioned in this episode: How to Create a Mind: 2012 book by Ray Kurzweil that inspired Dara The Memoirs of Akbar Ladjevardian: 2008 book about Dara’s grandfather, an Iranian industrialist, that led him to create his first “digital mind” Build: 2022 book by Tony Fadell that refers to itself as “a mentor in a box”; another inspiration for Dara The 2 Sigma Problem: 1984 paper by Benjamin Bloom about how students that receive one-on-one tutoring perform two standard deviations better than students educated in a classroom environment

Vercel CEO Guillermo Rauch has spent years obsessing over reducing the friction between having an idea and getting it online. Now with AI, he's achieving something even more ambitious: making software creation accessible to anyone with a keyboard. Guillermo explains how v0 has grown to 3 million users by focusing on reliability and quality, why ChatGPT has become their fastest-growing customer acquisition channel, and how AI is enabling “virtual coworkers” across design, development, and marketing. He shares his contrarian view that the future belongs to ephemeral, generated-on-demand applications rather than traditional installed software, and why he believes we're on the cusp of the biggest transformation to the web in its history. Hosted by Sonya Huang and Pat Grady, Sequoia Capital

In just two months, a scrappy three-person team at OpenAI sprinted to fulfill what the entire AI field has been chasing for years—gold-level performance on the International Mathematical Olympiad problems. Alex Wei, Sheryl Hsu and Noam Brown discuss their unique approach using general-purpose reinforcement learning techniques on hard-to-verify tasks rather than formal verification tools. The model showed surprising self-awareness by admitting it couldn’t solve problem six, and revealed the humbling gap between solving competition problems and genuine mathematical research breakthroughs. Hosted by Sonya Huang, Sequoia Capital

Isa Fulford, Casey Chu, and Edward Sun from OpenAI's ChatGPT agent team reveal how they combined Deep Research and Operator into a single, powerful AI agent that can perform complex, multi-step tasks lasting up to an hour. By giving the model access to a virtual computer with text browsing, visual browsing, terminal access, and API integrations—all with shared state—they've created what may be the first truly embodied AI assistant. The team discusses their reinforcement learning approach, safety mitigations for real-world actions, and how small teams can build transformative AI products through close research-applied collaboration. Hosted by Sonya Huang and Lauren Reeder, Sequoia Capital

Pushmeet Kohli leads AI for Science at DeepMind, where his team has created AlphaEvolve, an AI system that discovers entirely new algorithms and proves mathematical results that have eluded researchers for decades. From improving 50-year-old matrix multiplication algorithms to generating interpretable code for complex problems like data center scheduling, AlphaEvolve represents a new paradigm where LLMs coupled with evolutionary search can outperform human experts. Pushmeet explains the technical architecture behind these breakthroughs and shares insights from collaborations with mathematicians like Terence Tao, while discussing how AI is accelerating scientific discovery across domains from chip design to materials science. Hosted by Sonya Huang and Pat Grady, Sequoia Capital

Eric Ho is building Goodfire to solve one of AI’s most critical challenges: understanding what’s actually happening inside neural networks. His team is developing techniques to understand, audit and edit neural networks at the feature level. Eric discusses breakthrough results in resolving superposition through sparse autoencoders, successful model editing demonstrations and real-world applications in genomics with Arc Institute's DNA foundation models. He argues that interpretability will be critical as AI systems become more powerful and take on mission-critical roles in society. Hosted by Sonya Huang and Roelof Botha, Sequoia Capital Mentioned in this episode: Mech interp: Mechanistic interpretability, list of important papers here Phineas Gage: 19th century railway engineer who lost most of his brain’s left frontal lobe in an accident. Became a famous case study in neuroscience. Human Genome Project: Effort from 1990-2003 to generate the first sequence of the human genome which accelerated the study of human biology Emergent Misalignment: Narrow finetuning can produce broadly misaligned LLMs Zoom In: An Introduction to Circuits: First important mechanistic interpretability paper from OpenAI in 2020 Superposition: Concept from physics applied to interpretability that allows neural networks to simulate larger networks (e.g. more concepts than neurons) Apollo Research: AI safety company that designs AI model evaluations and conducts interpretability research Towards Monosemanticity: Decomposing Language Models With Dictionary Learning. 2023 Anthropic paper that uses a sparse autoencoder to extract interpretable features; followed by Scaling Monosemanticity Under the Hood of a Reasoning Model: 2025 Goodfire paper that interprets DeepSeek’s reasoning model R1 Auto-interpretability: The ability to use LLMs to automatically write explanations for the behavior of neurons in LLMs Interpreting Evo 2: Arc Institute's Next-Generation Genomic Foundation Model. (see episode with Arc co-founder Patrick Hsu) Paint with Ember: Canvas interface from Goodfire that lets you steer an LLM’s visual output  in real time (paper here) Model diffing: Interpreting how a model differs from checkpoint to checkpoint during finetuning Feature steering: The ability to change the style of LLM output by up or down weighting features (e.g. talking like a pirate vs factual information about the Andromeda Galaxy) Weight based interpretability: Method for directly decomposing neural network parameters into mechanistic components, instead of using features The Urgency of Interpretability: Essay by Anthropic founder Dario Amodei On the Biology of a Large Language Model: Goodfire collaboration with Anthropic