This Sequoia-backed lab thinks the brain is 'the floor, not the ceiling' for AI

What It Covers

Flapping Airplanes, a Sequoia-backed AI startup, pursues data-efficient foundation models inspired by brain learning mechanisms. The three founders explain their $180 million seed round, research-first approach targeting thousand-fold improvements in data efficiency, and hiring strategy focused on young, creative researchers willing to challenge AI orthodoxy rather than incrementally improve existing transformer architectures.

Key Questions Answered

• •Data Efficiency Gap: Current frontier models train on the sum totality of human knowledge, while humans learn with vastly less data. Flapping Airplanes targets thousand-fold improvements in data efficiency by studying why biological intelligence requires orders of magnitude less information than transformers. This approach unlocks domains like robotics and scientific discovery where data remains constrained and expensive to generate.
• •Research Cost Paradox: Radical research proves cheaper than incremental improvements because crazy ideas fail quickly at small scale, while incremental work requires expensive scaling runs to validate. Many interventions appearing promising at small scale fail at large scale, forcing incremental researchers up the costly scaling ladder. Fundamental research allows rapid iteration and failure detection without massive compute expenditure before attempting large-scale validation.
• •Commercialization Timeline: The team prioritizes deep research over immediate product launches, acknowledging they cannot provide specific timelines for solving fundamental problems. They maintain commercial backgrounds and plan to commercialize discoveries but recognize that signing enterprise contracts early would distract from valuable research. Focus remains their competitive advantage, enabled by investor willingness to fund longer research horizons in the current AI funding environment.
• •Intelligence Spectrum Theory: Models exist on a spectrum between statistical pattern matching and deep understanding, with current systems somewhere in the middle. Training on less data may force models toward genuine reasoning rather than memorization, potentially creating systems that know fewer facts but reason better. This shift could enable AI to generate novel scientific insights and medical advances rather than simply automating existing human work.
• •Hiring for Creativity: The team recruits exceptionally young researchers, including those still in high school or college, specifically seeking candidates unpolluted by thousands of existing papers. The primary signal involves whether candidates teach interviewers something new during conversations. This approach stems from experience showing young people compete effectively at the highest industry levels when given permission and support to contribute fundamentally new ideas.