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

Company Identity & Founders

• Organization: Flapping Airplanes (an AI lab, explicitly disclaiming any connection to the aviation industry despite name confusion).
• Timeline: Company launched three months ago.
• Co-founders:
  • Ben: Former PhD researcher in GPU kernel systems; co-founded the incubator "Prod."
  • Asher: Older brother of Ben; former Stanford PhD; previous experience at Cursor and Mercor.
  • Aidan Smith: Thiel Fellow; spent three years commuting between college and Neuralink, bridging neuroscience and machine learning.

Core Thesis: The Economic Necessity of Data Efficiency

• Current Market Gap: While LLMs dominate data-rich domains (search, coding) with trillion-dollar potential, these sectors rely on abundant internet-scale data.
• Target Domains: Future growth lies in data-scarce fields where current models struggle, including:
  • Robotics (high cost/complexity of data generation).
  • Trading (finite financial historical data).
  • Scientific discovery (limited data, unbounded potential).
  • End-to-end supply chains (representative of tens of thousands of broad economic tasks).
• Human Benchmark: Humans achieve coding proficiency with 10,000x to 100,000x less data than current frontier models, proving data efficiency is theoretically possible.

Strategic Drivers for Data Efficiency

• Scalability Disparity: Compute resources (flops) are becoming exponentially cheaper and more homogenous than data acquisition.
• Data Collection Friction:
  • Acquiring high-quality data requires navigating complex regulations and negotiating individual business terms.
  • No centralized "data purveyor" exists; competitors must acquire niche data (e.g., buying distressed bookstores).
• Competitive Landscape: High data requirements currently limit AI participation to a few centralized entities; data efficiency is a prerequisite for broader market competition and democratization.

Technical Approach: System-Level Innovation

• Hardware Abstraction Gap: Current frameworks (e.g., PyTorch) restrict algorithmic expression by synthesizing single-threaded models over parallel hardware, hiding efficient GPU primitives.
• Custom Architecture:
  • Building a custom virtual machine that takes full control of the GPU hardware.
  • Aims to execute "fine-grained" operations (e.g., deeply pipelined, "hog-wild" training loops) that are asymptotically inefficient in standard frameworks.
• Philosophy: Progress in ML over the last century stems from discovering new primitives to interact with existing hardware rather than just building new chips.
• IP Strategy: Specific algorithms are proprietary core IP; the company focuses on system co-optimization to enable these algorithms.

Forward-Looking Statements & Recruitment

• Deployment Impact: Achieving 1,000x data efficiency would theoretically reduce deployment complexity by a factor of 1,000.
• Hiring Focus: Seeking candidates with creative backgrounds and unconventional skills to drive paradigm shifts in system design.
• Timeline: No specific release dates mentioned, emphasizing immediate engagement for those interested in the hardware/software co-design space.