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

Summary

Ben and Asher Spector, co-founders of Flapping Airplanes, argue that AI progress has been concentrated in data-rich tasks like search and coding, while most of the economy is data-poor. They believe the next wave of AI value depends on making models far more data-efficient, since compute is easier to scale than data. They describe their lab's approach of building new GPU-level primitives and systems to enable algorithms that current frameworks like PyTorch cannot easily express. The main market implication is that data efficiency could broaden AI competition and deployment into domains such as robotics, trading, and scientific discovery.

• LLMs have excelled in search and coding because those tasks are extremely data-rich.
• Most of the economy, including robotics, trading, and scientific discovery, is data-poor.
• Compute gets cheaper and is more homogeneous than data, making data the key bottleneck.
• Data efficiency could make frontier AI easier to deploy and allow more companies to compete.
• Flapping Airplanes is building new GPU primitives and systems to enable data-efficient algorithms.
• The speakers view data efficiency as central to the broad deployment of AI beyond search and coding.