🇺🇸▼
In a technology world dominated by huge corporate giants, a new research group called Flapping Airplanes is charting a different course for artificial intelligence. The founders, Ben Spector, Asher Spector, and Aidan Smith, have focused their work on a major challenge. They want to help AI systems learn effectively from much smaller amounts of data. This laboratory received an initial $180 million to pursue this ambitious goal. The team believes that the current method used to build AI, while powerful, is just one way to solve the problem. They think there are other, better ways to create intelligence.
Ben Spector points out a significant difference between today's AI and children. The most advanced AI models are trained on nearly the entire public internet. In contrast, a young child learns to understand the physical world from a tiny fraction of that information. "So there's a big gap there," he notes. "It's worth understanding."
Solving this problem of data efficiency is their main bet for the next big step in AI. Achieving this would be a major scientific breakthrough. It could change global commerce by allowing AI to work in areas where data is scarce. Current models struggle in fields like advanced robotics or specialized science because they need so much data to function well.
Starting a new AI lab in this era might seem scary. Big tech companies have poured vast amounts of money into the sector. However, the Flapping Airplanes team does not see themselves as direct competitors to these massive corporations. Instead, they are asking different questions about the nature of intelligence. Aidan Smith, who previously worked at Neuralink, points to the human mind as a key inspiration.
"If you look at the human mind, it learns in an incredibly different way from [today's AI systems]," he explains. Modern large language models can memorize enormous amounts of data. Yet, they often cannot learn new skills quickly. Adapting these systems usually requires "rivers and rivers of data." This is a resource-heavy process that limits how flexible the AI can be.
Asher Spector emphasizes the practical importance of their focus. "Lots of regimes that are really important are also highly data constrained," he says. An AI model that uses data a million times more efficiently would be much more useful across the global economy. The ultimate goal is to create AI that learns like humans do. This means gaining deep understanding from limited examples instead of just processing huge datasets through statistical patterns.