A number of exciting new research-focused AI labs have emerged recently, and Flapping Airplanes is one of the most interesting. Propelled by its young and curious founders, the lab is focused on finding less data-hungry ways to train AI. This research is a potential game-changer for the economics and capabilities of AI models. With $180 million in seed funding, they will have considerable runway to figure it out.
I recently spoke with the lab’s three co-founders—brothers Ben and Asher Spector, and Aidan Smith—about why this is an exciting moment to start a new AI lab and why they keep returning to ideas about the human brain.
I began by asking why now. Established labs like OpenAI and DeepMind have invested so much in scaling their models, making the competition seem daunting. Why did this feel like a good moment to launch a foundation model company?
Ben explained that there is simply so much more to do. The advances over the last five to ten years have been spectacular, and they use the tools every day. However, they questioned whether this represents the entire universe of what needs to happen. After careful thought, their answer was no. They identified the data efficiency problem as the key issue to explore. Current frontier models are trained on the sum totality of human knowledge, while humans can obviously make do with far less. That significant gap is worth understanding.
Their work is a concentrated bet on three things. First, it is a bet that the data efficiency problem is the important direction to pursue—a new and different path where genuine progress can be made. Second, it is a bet that solving this will be commercially valuable and make the world a better place. Third, it is a bet that the right team to tackle it is a creative and, in some ways, inexperienced team that can re-examine these problems from the ground up.
Aidan agreed, noting they do not see themselves as competing with other labs because they are looking at a very different set of problems. The human mind learns in an incredibly different way from transformers. That is not to say it is better, just very different. Large language models have an incredible ability to memorize and draw on a great breadth of knowledge, but they cannot pick up new skills quickly. It takes rivers of data to adapt. When you look inside the brain, the algorithms it uses are fundamentally different from gradient descent and other techniques used to train AI today. That is why they are building a new guard of researchers to address these problems and think differently about the AI space.
Asher added that the question is scientifically fascinating: why are the intelligent systems we have built so different from what humans do? Where does this difference come from? How can we use knowledge of that difference to make better systems? He also believes it is commercially viable and good for the world. Many important regimes, like robotics or scientific discovery, are highly data-constrained. Even in enterprise applications, a model that is a million times more data efficient is probably a million times easier to integrate into the economy. For them, it was exciting to take a fresh perspective and consider what could be done with a vastly more data-efficient model.
This led to my next question, which ties into the lab’s name, Flapping Airplanes. There is a philosophical question in AI about how much we should try to recreate the human brain versus creating a more abstract intelligence that takes a completely different path. Aidan comes from Neuralink, which is all about the human brain. I asked if they see themselves pursuing a more neuromorphic view of AI.
Aidan said he sees the brain as an existence proof—evidence that there are other algorithms out there beyond the current orthodoxy. The brain operates under crazy constraints; for instance, it takes a millisecond to fire an action potential, while a computer can perform countless operations in that time. Realistically, there is probably an approach that is much better than the brain and also very different from the transformer. They are inspired by some things the brain does but are not tied down by it.
Ben added that this idea is captured in their name. Think of current systems as big Boeing 787s. They are not trying to build birds, which would be a step too far. They are trying to build a flapping airplane. From a computer systems perspective, the constraints of the brain and silicon are so different that we should not expect the resulting systems to look the same. When the substrate is different and you have genuinely different trade-offs regarding the cost of compute and data movement, you expect the systems to look different. However, just because they will look different does not mean we should not take inspiration from the brain to improve our own systems.
It feels like there is now more freedom for labs to focus on research rather than just developing products. This seems like a big difference for this generation of labs. Some are very research-focused, while others are “research-focused for now.” I asked what that conversation looks like within Flapping Airplanes.
Asher said he wishes he could provide a timeline, but they do not know the answers. They are looking for truth. That said, they have commercial backgrounds. He spent time developing technology for companies that generated reasonable revenue, and Ben has incubated startups. They are excited to commercialize because they believe it is good for the world to put the value they create into the hands of people who can use it. They are not opposed to commercialization, but they need to start with research. If they began by signing big enterprise contracts, they would get distracted and not do the valuable research.
Aidan added that they want to try radically different things, acknowledging that sometimes radical ideas are just worse than the current paradigm. They are exploring a set of different trade-offs, hoping they will prove superior in the long run.
Ben stated that companies are at their best when they are really focused on doing one thing well. Startups must pick the most valuable thing to do and commit to it fully. They create the most value by being all-in on solving fundamental problems for the time being. He is optimistic that they might make enough progress reasonably soon to start engaging with the real world, where you learn a lot from feedback. The recent change in the economics and financing of these structures enables companies to focus on what they are good at for longer periods, which is what he is most excited about.
To spell out what I was referring to: there is so much excitement and clear opportunity for investors that they are willing to give $180 million in seed funding to a completely new company full of very smart but very young people who did not just cash out of a previous venture. I asked how engaging with that process was. Did they know going in that there was this appetite, or did they discover they could make this a bigger thing than they thought?
Ben said it was a mixture. The market had been hot for many months, so it was no secret that large rounds were coming together. However, you never quite know how the fundraising environment will respond to your specific ideas. Even during their fundraise, they learned a lot and refined their opinions on what to prioritize and the right timelines for commercialization. They were somewhat surprised by how well their message resonated because it was very clear to them, but you never know if others will agree or think you are crazy. They were extremely fortunate to find amazing investors with whom their message resonated deeply, which was surprising and wonderful.
Aidan noted that a thirst for an age of research has been in the water for a while. More and more, they find themselves positioned as the player to pursue that age of research and try radical ideas.
For scale-driven companies, there is an enormous cost of entry for foundation models, as building at that scale is incredibly compute-intensive. Research is a bit in the middle; presumably, you are building foundation models, but if you are doing it with less data and are not so scale-oriented, maybe you get a bit of a break. I asked how much they expect compute costs to limit their runway.
Ben explained that one advantage of doing deep, fundamental research is that, somewhat paradoxically, it is much cheaper to try really crazy, radical ideas than to do incremental work. Incremental work often requires going far up the scaling ladder to see if an intervention holds, which is very expensive. In contrast, a crazy new idea about a new architecture or optimizer will probably fail on the first run, so you do not have to scale it up. That does not mean scale is irrelevant for them; it is an important tool in the toolbox. Being able to scale up their ideas is relevant, but they can try many ideas at a very small scale before considering large scale.
Asher added that you should be able to use all the internet, but you should not need to. They find it perplexing that you need all the internet to achieve human-level intelligence.
So, what becomes possible if you can train more efficiently on data? Presumably, the model will be more powerful and intelligent. But do they have specific ideas about where that leads? Are we looking at more out-of-distribution generalization, or models that get better at a particular task with less experience?
Asher said they are doing science, so he does not know the answer, but he can offer three hypotheses. First, there is a broad spectrum between looking for statistical patterns and having deep understanding. Current models live somewhere on that spectrum. It is possible that training models on less data forces them to develop incredibly deep understandings of everything they have seen, making them more intelligent in interesting ways. They may know fewer facts but get better at reasoning.
A second hypothesis is that it is currently very expensive, both operationally and monetarily, to teach models new capabilities because they need so much data. An output of their work could be vastly more efficient post-training, where only a couple of examples could put a model into a new domain.
A third possibility is that this unlocks new verticals for AI, like certain types of robotics or scientific discovery, where limited data, not hardware, is the constraint. The fact that you can tele-operate robots proves the hardware is sufficiently good.
Ben added that when considering AI’s impact on the world, one view is that it is a deflationary technology meant to automate jobs and make work cheaper. While that will happen, it is not the most exciting vision. The most exciting vision is AI enabling new science and technologies that humans are not smart enough to conceive. For that, the axis between true generalization and memorization is extremely important. Models need to be on the creativity side of the spectrum to have the deep insights that lead to advances in medicine and science. He is mission-oriented around the question of whether AI can do things humans fundamentally could not do before, which is more than just replacing jobs.
Does that put them in a particular camp on the AGI conversation or the out-of-distribution generalization conversation?
Asher said he does not exactly know what AGI means. Capabilities are advancing quickly, and tremendous economic value is being created, but he does not think we are very close to a “God-in-a-box.” He does not believe there will be a singularity within two months or even two years where humans become completely obsolete. He agrees with Ben that it is a really big world with a lot of work to do, and they are excited to contribute.
The idea about the brain and the neuromorphic aspect feels relevant. They are saying the relevant comparison for large language models is the human brain more than the Mechanical Turk or deterministic computers that came before.
Aidan emphasized that the brain is not the ceiling but the floor. He sees no evidence that the brain is not a knowable system following physical laws. It is under many constraints, so we should expect to create capabilities that are much more interesting, different, and potentially better than the brain in the long run. They are excited to contribute to that future, whether it is AGI or otherwise.
Asher added that the brain is a relevant comparison because it helps us understand how big the space is. It is easy to see all the progress and think we have the answer and are almost done. But with more perspective, there is a lot we do not know.
Ben clarified that they are not trying to be better per se, but different. All these systems will have different trade-offs; you gain an advantage somewhere, and it costs you somewhere else. The world has many domains with different trade-offs, so having more systems and fundamental technologies to address them will likely help AI diffuse more effectively and rapidly through the world.
One way they have distinguished themselves is through their hiring approach, bringing in very young people, in some cases still in college or high school. What clicks for them when talking to someone that makes them think they want that person working on these research problems?
Aidan said it is when someone dazzles you with new ideas and thinks about things in a way established researchers cannot because they have not been influenced by the context of thousands of papers. The number one thing they look for is creativity. Their team is exceptionally creative, and he feels lucky to discuss radical solutions to big AI problems with them every day.
Ben said the number one signal he looks for is whether someone teaches him something new when he spends time with them. If they do, the odds are good they will teach the team something new about their work. Creative, new ideas are the priority in research. Part of his background involved starting an incubator called Prod during his undergrad and PhD, which worked with companies that turned out well. That experience showed that young people can absolutely compete at the highest echelons of industry. A big part of the unlock is realizing you can go do this stuff and contribute at the highest level.
They also recognize the value of experience and have hired people who have worked on large-scale systems. Their mission has resonated with experienced folks as well. Their key requirement is people who are not afraid to change the paradigm and imagine a new system of how things might work.
One thing I have been puzzling about is how different the resulting AI systems will be. It is easy to imagine something like Claude Opus that works twenty percent better and can do twenty percent more things. But if it is completely new, it is hard to think about where that goes or what the end result looks like.
Asher said he does not know if I have ever talked to the GPT-4 base model, but it had many strange emerging capabilities. For example, you could give it a snippet of an unwritten blog post and ask who wrote it, and it could identify the author. Models are smart in ways we cannot fathom, and future models will be smarter in even stranger ways. We should expect the future to be really weird and the architectures to be even weirder. They are looking for thousand-fold wins in data efficiency, not incremental change, so we should expect unknowable, alien changes and capabilities at the limit.
Ben broadly agreed but was slightly more tempered in how these things will eventually be experienced by the world. Just as the GPT-4 base model was tempered by OpenAI, you want to put things in forms where consumers are not staring into the abyss. That is important. However, he broadly agrees that their research agenda is about building capabilities fundamentally different from what can be done now.
Finally, I asked if there are ways people can engage with Flapping Airplanes now, or if they should just stay tuned for when the research and models come out.
Asher said they have an email for general inquiries and another specifically for disagreements. They have had some cool conversations where people send long essays about why what they are doing is impossible, and they are happy to engage with them.
Ben added that no one has convinced them yet.
Asher also mentioned they are looking for exceptional people who want to change the field and the world, so interested individuals should reach out.
Ben reiterated that an unorthodox background is okay; you do not need two PhDs. They are really looking for folks who think differently.