AI's Wild Future: Flapping Airplanes & Radical Innovation

The artificial intelligence landscape is buzzing with new research labs, and Flapping Airplanes is quickly becoming one of the most intriguing. Founded by the young and inquisitive brothers Ben and Asher Spector, and Aidan Smith, the lab is laser-focused on developing AI models that require significantly less data for training. This approach has the potential to revolutionize the economics and capabilities of AI – and with a substantial $180 million in seed funding, they have ample resources to explore this groundbreaking path. This article delves into the vision of Flapping Airplanes, their unique approach, and the potential impact on the future of AI.

Why Now? A New Wave of AI Research

Last week, I had the opportunity to speak with the three co-founders of Flapping Airplanes about the current moment in AI and their decision to embark on this ambitious venture. The question of timing is crucial, especially considering the massive investments made by established labs like OpenAI and DeepMind. The competition appears daunting, so why launch a foundation model company now?

Ben Spector explained, “There’s just so much to do. The advances we’ve seen in the last five to ten years have been spectacular. We love the tools and use them daily. But is this the complete picture? We believe there’s much more to explore. We identified the data efficiency problem as a key area for innovation. Current frontier models are trained on a vast collection of human knowledge, yet humans learn effectively with far less data. This gap is significant and warrants investigation.”

He continued, “Our strategy is a concentrated bet on three core principles: the importance of data efficiency, the potential for commercial value and positive global impact, and the necessity of a creative, even inexperienced, team to approach these problems from the ground up.”

Beyond Scaling: A Different Approach to AI

Aidan Smith emphasized that Flapping Airplanes isn’t directly competing with existing labs. “We’re tackling a different set of problems. The human mind learns in a fundamentally different way than transformers. That’s not to say better, just different. LLMs excel at memorization and leveraging vast knowledge, but struggle with rapid skill acquisition. They require enormous datasets to adapt. The brain, however, utilizes algorithms drastically different from gradient descent and other common AI training techniques. That’s why we’re building a new generation of researchers to rethink the AI space.”

Asher Spector added, “This is a scientifically fascinating question: why are the intelligent systems we’ve built so different from the human brain? Where does this difference originate? How can we leverage this understanding to create better systems? Beyond the scientific intrigue, we believe this approach is commercially viable and beneficial for the world. Many critical areas, like robotics and scientific discovery, are highly data-constrained. A model a million times more data-efficient could be a million times easier to integrate into the economy.”

Flapping Airplanes: Inspiration from Biology, Not Replication

The name "Flapping Airplanes" itself hints at a core philosophical question in AI: how much should we strive to recreate the human brain versus creating a completely novel intelligence? Given Aidan Smith’s background at Neuralink, a company focused on the human brain, is Flapping Airplanes pursuing a neuromorphic view of AI?

Aidan clarified, “I view the brain as proof of concept. It demonstrates that alternative algorithms exist. There isn’t a single, definitive approach. The brain operates under significant constraints – the speed of an action potential, for example, is incredibly slow compared to computer processing speeds. Realistically, there’s likely an approach far superior to the brain, and very different from transformers. We’re inspired by the brain, but not bound by it.”

Ben elaborated, “Think of current systems as large Boeing 787s. We’re not trying to build birds – that’s too far. We’re aiming for a flapping airplane. The constraints of the brain and silicon are fundamentally different. We shouldn’t expect these systems to look alike. Different substrates and trade-offs regarding compute cost, data locality, and data movement will inevitably lead to different architectures. However, that doesn’t mean we should ignore the brain’s insights and incorporate elements we find valuable.”

A Shift Towards Research-Focused AI Labs

There’s a growing sense that new AI labs have more freedom to focus on research rather than immediate product development. This represents a significant shift from previous generations of labs. Some are research-focused, while others are “research-focused for now.” How does this dynamic play out within Flapping Airplanes?

Asher responded, “I wish I could provide a timeline. I can’t predict when we’ll solve the research problem and commercialize our findings. We’re driven by a pursuit of truth. However, we all have commercial backgrounds. I’ve spent time developing technologies that generated substantial revenue for companies. Ben has incubated successful startups. We’re excited about commercialization and believe it’s beneficial to put valuable tools in the hands of those who can use them. But we must prioritize research. If we immediately sign enterprise contracts, we’ll become distracted and lose focus on the fundamental research that truly matters.”

Aidan added, “We want to explore radically different ideas, and sometimes those ideas simply don’t work. We’re exploring various trade-offs, hoping they’ll yield long-term benefits.”

Ben concluded, “Companies excel when they’re focused on doing one thing exceptionally well. Large companies can handle multiple initiatives, but startups must prioritize. We create the most value by focusing on solving fundamental problems for now. I’m optimistic that we’ll make enough progress soon to begin real-world testing and gather valuable feedback. The world is a constant source of truth, and real-world interaction is crucial. The recent changes in economics and financing allow companies to focus on their strengths for longer periods, which I believe will lead to truly differentiated work.”

Securing Funding: A Thirst for Research

The $180 million in seed funding secured by Flapping Airplanes is particularly noteworthy. It demonstrates a growing investor appetite for research-driven AI companies, even those led by young, relatively inexperienced founders. How was the fundraising process?

Ben explained, “The market has been hot for months, so large rounds were already happening. However, you never know how investors will respond to your specific ideas. Fundraising is a feedback loop – we learned and refined our approach throughout the process. We were surprised by how well our message resonated. It was clear to us, but you never know if others will share your vision. We’re fortunate to have found investors who believed in our approach and were excited to support us.”

Aidan added, “There’s a growing thirst for the age of research. We’re increasingly positioned as the player to pursue radical ideas and push the boundaries of AI.”

Compute Costs and the Future of Scaling

Foundation models typically require immense computational resources. How will compute costs impact Flapping Airplanes’ runway, especially given their focus on data efficiency rather than massive scale?

Ben addressed this concern, “Paradoxically, deep, fundamental research is often cheaper than incremental work. Incremental improvements require scaling up to determine their effectiveness, which can be expensive. Radical new ideas, on the other hand, often fail quickly, saving computational resources. This doesn’t mean scale is irrelevant. Scale is a valuable tool, and we’ll leverage it when appropriate. However, our approach allows us to test many ideas at a small scale before committing to large-scale deployments.”

Asher added, “You should be able to use all the internet, but you shouldn’t *need* to. We find it perplexing that training AI requires access to the entire internet.

What Becomes Possible with Data-Efficient AI?

If Flapping Airplanes succeeds in training AI models more efficiently, what new possibilities will emerge? Will we see improved out-of-distribution generalization, or models that learn tasks more quickly with less experience?

Asher outlined three hypotheses: “First, there’s a spectrum between statistical pattern recognition and true understanding. Current models fall somewhere in between. Data efficiency might force models to develop deeper understandings, potentially sacrificing factual knowledge for improved reasoning. Second, it could significantly reduce the cost of teaching models new capabilities, allowing them to adapt with just a few examples. Third, it could unlock new verticals for AI, such as robotics and scientific discovery, where data is scarce.”

Ben expanded on this, “AI has the potential to be a deflationary technology, automating tasks and reducing costs. However, the most exciting vision is one where AI enables new scientific and technological breakthroughs that humans couldn’t achieve on their own. We want to build models that are creative and capable of generating novel insights, going beyond simply automating existing processes.”

AGI and the Future of Intelligence

Does Flapping Airplanes’ approach align with the pursuit of Artificial General Intelligence (AGI)?

Asher cautiously responded, “I don’t know what AGI truly means. Capabilities are advancing rapidly, and there’s tremendous economic value being created. I don’t believe we’re close to a ‘God-in-a-box’ scenario or a singularity where humans become obsolete. I agree with Ben – there’s a lot of work to do.”

Ben emphasized, “We’re not trying to replicate the brain; we’re trying to be different. These systems will inevitably have different trade-offs. Having more fundamental technologies that address diverse domains will allow AI to diffuse more effectively throughout the world.”

The Importance of a Creative Team

Flapping Airplanes has distinguished itself with its hiring approach, prioritizing creativity and unconventional backgrounds. What qualities do they look for in potential hires?

Aidan explained, “We look for people who dazzle us with new ideas and think differently. Our team is exceptionally creative, and I’m fortunate to collaborate with them on radical solutions to complex AI problems.”

Ben added, “The number one signal I look for is whether someone teaches me something new. If they can do that, they’re likely to contribute valuable insights to our research. We value creativity and a willingness to challenge the status quo.”

Looking Ahead: A Weird and Wonderful Future

Asher concluded, “We should expect the future to be really weird, and the architectures to be even weirder. We’re looking for 1000x wins in data efficiency, not incremental change. We should anticipate alien and unknowable capabilities at the limit.”

Ben agreed, “We want to put these capabilities in forms that are accessible and understandable. But fundamentally, we’re building systems that will be different from anything we’ve seen before.”

Engaging with Flapping Airplanes

Are there ways for people to get involved with Flapping Airplanes? Asher shared, “You can reach out to us at Hi@flappingairplanes.com. We also have disagree@flappingairplanes.com for those who want to challenge our ideas. We welcome constructive criticism. We’re also actively seeking exceptional people who want to change the field. If you’re interested, please reach out.”

Ben added, “You don’t need two PhDs. We’re looking for people who think differently.”