Why Data Is the Real AI Bottleneck: Flapping Airplanes' Ben and Asher Spector

At AI Ascent 2026, Ben and Asher Spector, brothers and co-founders of the AI lab Flapping Airplanes, argue that the future of AI belongs to whoever can build models that need a lot less data. They make the case that the trillion-dollar wins in AI so far (search, coding) happen to be the most data-rich problems on earth, and that almost everything else in the economy (robotics, trading, scientific discovery, the long tail of every industry) is data-poor. They explain why compute is genuinely easier to scale than data, why a thousand-fold improvement in data efficiency would democratize who can train frontier models, and how their approach—building new primitives for interacting with GPU hardware—opens up algorithmic territory that current frameworks like PyTorch can't easily reach. 00:00 Introduction 01:14 Not an Airplane Company 01:44 Talk Outline and Thesis 01:57 Why Data Efficiency Matters 02:46 The Long Tail of Low Data 03:28 Compute Scales Easier Than Data 04:20 Data Moats and Competition 05:16 Systems Plus Algorithms Approach 06:25 GPU Primitives and Closing