Why Data Is the Real AI Bottleneck: Flapping Airplanes' Ben and Asher Spector
In a Nutshell
Flapping Airplanes founders Ben and Asher Spector argue that data efficiency is AI's key bottleneck, as LLMs succeed in data-rich domains like search and coding but struggle in scarce ones like robotics, trading, and science—humans achieve similar feats with 10,000-100,000x less data. A 1000x more efficient model eases deployment amid rising compute costs and data scarcity, democratizing AI beyond centralized players. Their approach fuses novel GPU systems—bypassing PyTorch limitations for fine-grained, irregular operations—with algorithms to unlock untapped hardware capabilities for data-efficient training.
These notes were generated by AI and may contain inaccuracies.
Ben and Asher Spector, founders of Flapping Airplanes, discuss data efficiency. Third co-founder is Aiden Smith, a Thiel fellow with expertise in brain and machine learning.
Ben spent 3 years in PhD writing low-level GPU systems, helped start incubator Prod. Asher, Ben's older brother, was Stanford PhD, spent time at Cursor and Mercor.
Launched company 3 months ago. Received inbound from aviation industry offering runways, airplane parts, wind tunnels. Disclaimer: not an airplane company, an AI lab.
Two parts: thesis on why data efficiency is important, approach fusing systems and algorithmic work.
LLMs excel at search and coding, trillion-dollar markets. Success due to abundant data: entire internet for search, large fraction for coding. Coding allows easy synthetic data generation.
Can these capabilities be achieved with much less data? Humans become good at coding with 10,000 to 100,000 times less data than current models.
Reasons it matters: future domains with less data include robotics (data generation complicated), trading (limited financial data), scientific discoveries (very little data, unbounded potential), end-to-end toaster supply chain (represents tens of thousands of under-resourced economic tasks).
Sign in to read the full notes
Get access to AI-generated notes, topic timestamps, and more.