Flapping Airplanes: AI's Future Demands Radically Different Approaches

In recent months, a wave of innovative, research-focused AI laboratories has emerged, with Flapping Airplanes standing out as particularly intriguing. Driven by its youthful and inquisitive founders, Flapping Airplanes is dedicated to pioneering methods for training AI that require significantly less data. This pursuit holds the potential to fundamentally transform the economic landscape and capabilities of AI models. Bolstered by a substantial $180 million in seed funding, the team has ample resources to explore and realize this vision.

Last week, I had the opportunity to converse with the lab’s three co-founders — brothers Ben and Asher Spector, and Aidan Smith. Our discussion centered on the compelling reasons behind launching a new AI lab at this particular juncture, and their recurring interest in the intricate workings of the human brain.

I began by posing the question: "Why now? Established labs like OpenAI and DeepMind have invested heavily in scaling their models, making the competitive landscape seem formidable. What made this feel like an opportune moment to launch a foundation model company?"

Ben Spector responded, "There’s simply so much work yet to be done. The advancements we’ve witnessed over the past five to ten years have been spectacular, and we genuinely appreciate and utilize these tools daily. However, the critical question is whether this represents the full extent of what needs to happen. After careful consideration, our answer was a resounding no; there is significantly more to accomplish. For us, the data efficiency problem emerged as the paramount area to investigate. Current frontier models are trained on the sum totality of human knowledge, yet humans can clearly operate effectively with far less. This indicates a substantial gap that warrants deep understanding."

He elaborated, "Our endeavor is fundamentally a concentrated bet on three core tenets. Firstly, we believe the data efficiency problem is crucial, representing a new and distinct direction where tangible progress can be made. Secondly, we are confident that solving this will yield immense commercial value and ultimately contribute to a better world. Lastly, we posit that the ideal team for this challenge is one that is creative, and even in some respects, inexperienced, enabling them to approach these problems afresh, from the ground up."

Aidan Smith added, "Absolutely. We don't perceive ourselves as being in direct competition with other labs, primarily because we are addressing a fundamentally different set of problems. The human mind learns in a profoundly distinct manner compared to transformer models. This isn't to say one is superior, merely that they are very different, leading to unique trade-offs. Large Language Models possess an incredible capacity for memorization and drawing upon vast knowledge bases, but they struggle to acquire new skills rapidly, demanding 'rivers and rivers of data' for adaptation. Conversely, when we examine the brain, the algorithms it employs are fundamentally divergent from gradient descent and other common AI training techniques today. This conviction drives us to cultivate a new generation of researchers dedicated to tackling these issues and fostering a truly different way of thinking about the AI landscape."

Asher Spector emphasized the dual appeal: "The question of why the intelligent systems we've built are so different from human cognition is scientifically fascinating. Understanding the origin of this difference and how that knowledge can inform better systems is a profound pursuit. Simultaneously, I believe this area is immensely commercially viable and beneficial for the world. Many critical domains, such as robotics or scientific discovery, are inherently data-constrained. Even within enterprise applications, a model that is a million times more data efficient would likely be a million times easier to integrate into the economy. For us, it was incredibly exciting to adopt a fresh perspective on these approaches and envision the possibilities if we could develop a vastly more data-efficient model."

This led to my subsequent question, which also connected to the lab's name, Flapping Airplanes. "In AI, there's a philosophical debate about the extent to which we should strive to replicate human brain functions versus creating a more abstract intelligence that follows an entirely different trajectory. Aidan, given your background at Neuralink, which focuses intently on the human brain, do you see Flapping Airplanes pursuing a more neuromorphic approach to AI?"

Aidan clarified, "I view the brain primarily as an existence proof. It serves as evidence that alternative algorithms exist; there isn't just one orthodoxy. Furthermore, the brain operates under some truly extraordinary constraints. Its underlying hardware is remarkable, taking a millisecond for an action potential to fire—a timeframe during which a computer executes an immense number of operations. Realistically, there's likely an approach that surpasses the brain's efficiency, and is also very different from the transformer architecture. So, while we are deeply inspired by certain aspects of the brain's functioning, we do not feel constrained by it."

Ben added, "To build on that, our name, 'Flapping Airplanes,' is quite illustrative. Consider current systems as large Boeing 787s. We're not attempting to build birds; that would be a step too far. Instead, we aim to construct a type of 'flapping airplane.' From my perspective in computer systems, the fundamental constraints of the brain and silicon are sufficiently distinct that we shouldn't expect the resulting systems to be identical. When the underlying substrate differs so profoundly, leading to genuinely disparate trade-offs regarding compute costs, locality, and data movement, you naturally anticipate these systems will look somewhat different. However, this divergence does not preclude us from drawing inspiration from the brain and leveraging its interesting aspects to enhance our own systems."

It seems there's a newfound liberty for labs to prioritize fundamental research over immediate product development, marking a significant shift for this generation of AI companies. Some are deeply research-focused, while others are 'research-focused for now.' What does this internal dialogue look like within Flapping Airplanes?

Asher responded, "I wish I could provide a definitive timeline. I wish I could say that in three years, we'll have solved the core research problem and outline our commercialization strategy. But I can't. We don't have those answers; we are actively seeking truth. That said, we all possess commercial backgrounds. I've spent considerable time developing technology that generated substantial revenue for companies. Ben has incubated numerous startups with strong commercial foundations. We are genuinely excited by the prospect of commercialization; we believe it benefits the world to take created value and place it in the hands of those who can utilize it. So, we are not opposed to it. However, our immediate imperative is to conduct research, because if we divert our focus by pursuing large enterprise contracts prematurely, we risk distraction and failing to accomplish the truly valuable research."

Aidan concurred, "Indeed, we aim to explore truly radical approaches. Sometimes, these radically different things might initially prove less effective than the current paradigm. We are deliberately exploring a distinct set of trade-offs, with the long-term hope that they will lead to significant differentiation."

Ben further explained, "Companies operate at their peak when they are intensely focused on excelling at one thing. Large corporations can afford to juggle many different initiatives simultaneously. As a startup, however, you must meticulously choose the single most valuable thing you can do and commit to it entirely. For the time being, we generate the most value by dedicating ourselves fully to solving fundamental problems."

He continued, "I am optimistic that reasonably soon, we might achieve sufficient progress to begin engaging with the real world. There is immense learning to be gained from real-world feedback. The world is an incredible teacher, a vast reservoir of truth accessible at will. I believe the most significant enabler of this current environment, driven by recent shifts in economic and financing structures, is the ability for companies to maintain a sharp focus on their core strengths for extended periods. That focused approach is what excites me most, as it will allow us to produce truly differentiated work."

To clarify what I believe you're referencing: the sheer excitement and clear investment opportunity in AI have led investors to commit $180 million in seed funding to a completely new company, staffed by highly intelligent but also very young individuals who haven't just exited major tech companies. What was the experience of engaging with that fundraising process like? Did you anticipate this level of appetite going in, or was it a discovery that allowed you to envision something larger than initially thought?