Sequoia Capital — searchable YouTube transcripts & AI Q&A

Inside the Robot Data Factory: Mecka's Josh Gao on Sensorizing Humans to Train Robots

Josh Gao argues that robotics is data-constrained, unlike LLMs which benefited from internet-scale data, and that Mecca's role is to create, process, and license the diverse data needed to train robots. He explains that scaling data collection involves sensorizing humans with custom hardware like tactile gloves to capture force data, and that the industry is converging on a mix of egocentric, teleoperation, and simulation data. He predicts that useful robots will first deploy in constrained industrial settings, and that the path to general robots will be gradual, unlocking commercial use cases along the way.

David Silver and Lasse Espeholt of Ineffable Intelligence: A Superlearner for the Age of Experience

David Silver and Lasse Espeholt argue that superintelligence will emerge from systems that learn autonomously from experience, not from scaling human-language data. They describe building a 'superlearner' that discovers knowledge for itself in simulated 'crucibles,' citing AlphaZero's rapid superhuman performance after abandoning human data as evidence. They advocate for accelerating research in simulation while pacing real-world deployment, and predict strong evidence of progress within a year and true superintelligence within three.

Luca Ferrari: The Operating System Behind Bending Spoons

Luca Ferrari argues that Bending Spoons' success comes almost entirely from its operating platform, not from picking acquisition targets. The platform integrates acquired businesses deeply, replatforming them on shared technology and moving a core team across products, which allows a small, high-talent team to run a business that others would need 200 people for. He emphasizes that talent density and culture matter more than experience, and he shares that hiring is data-driven, with interviews being the weakest predictor of performance.

Google's Head of AI Infra, Amin Vahdat, on the Architecture of Intelligence

Amin Vahdat argues that AI data centers are increasingly purpose-built and co-designed with hardware, unlike traditional data centers planned for 25-30 years. He emphasizes that the key metric is 'goodput'—the useful work delivered per watt—rather than theoretical flops, because at 100,000-accelerator scale, failures are constant and recovery overhead can dominate. He also notes that most performance gains come from software and model optimizations, not just hardware, and that Google's TPU architecture has remained fundamentally stable across generations.

Databricks’ Ali Ghodsi Never Wanted to Be CEO. Now He’s Among the Best

Ali Ghodsi argues that CEOs should identify the single biggest bottleneck and focus the entire company on it for years, even if it means taking a step back to eventually leap forward. He illustrates this with Databricks' shift from a product-led growth model to an enterprise sales force, and the creation of the lakehouse category, which was initially ridiculed but eventually became a major success. He also emphasizes the importance of being conflict-averse as a CEO and the need to be extremely picky when hiring executives.