Data efficiency in AI asks one deceptively simple question: can we get powerful model capabilities with far less training data? According to Ben and Asher, co-founders of AI lab Flappy Airplanes, the answer is almost certainly yes — and solving this problem could determine which companies, industries, and even countries get to participate in the AI economy. While current large language models have become exceptional at search and coding partly because those domains are swimming in data, most of the real economy is not. That gap is the opportunity.

What Is Data Efficiency in AI and Why Does It Matter?

Today's best LLMs are trained on staggering amounts of data. Search models effectively consume the entire internet. Coding models are trained on a massive slice of it, with the added bonus that synthetic data is easy to generate in that domain — just run the code and check if it works. The result is impressive capability, but it raises an uncomfortable question: what happens when you need AI to work somewhere that doesn't have internet-scale data?

Data efficiency is the pursuit of equivalent model performance with dramatically less training data. The intuition that it's achievable isn't just optimism — it's grounded in biology. Humans become skilled programmers having seen perhaps ten thousand to a hundred thousand times less data than a frontier model trains on. That gap suggests there's a lot of room to improve how AI learns, not just how much it consumes.

The Flappy Airplanes team frames this as the central question for the next phase of AI deployment: can we unlock AI capabilities for domains where data is scarce by design?

Which AI Domains Suffer Most From Data Scarcity?

The economic argument for data efficiency becomes vivid when you look at where AI hasn't yet delivered transformative results. Consider the contrast:

  • Robotics: Generating quality training data requires physical hardware, careful instrumentation, and enormous effort. Jim (referenced in the talk) noted that the robotics community has to work extremely hard just to produce the data needed to train capable systems.
  • Trading and finance: There is only so much historical financial data in existence. You can't manufacture more market history.
  • Scientific discovery: Experimental results accumulate slowly. The dataset of all human scientific knowledge is finite — and the most valuable discoveries are at the frontier where data is thinnest.
  • Industrial supply chains: The presenters use the example of a toaster supply chain — not because it's glamorous, but because it represents tens of thousands of specialized, niche domains that together constitute the broad economy. None of them have internet-scale datasets.

The implication is significant: the trillion-dollar opportunity in search and coding is only the beginning. The rest of the economy is waiting for AI that doesn't require vast data to work.

Is Compute or Data Harder to Scale for AI?

There's a widely shared assumption in the AI industry that compute is the scarce resource. Asher pushes back on this — or at least complicates it.

Yes, compute is expensive. But compute follows a relatively predictable trajectory: FLOPs get exponentially cheaper over time, the chip market is relatively homogeneous, and if you have enough money, you can go out and buy more GPUs. Greg (referenced in the talk, presumably Greg Brockman) described how, after GPT launched, OpenAI could simply try to purchase all available compute. There's a market for that.

Data is different. There is no centralized data provider. Collecting frontier-quality data for specialized domains means:

  • Negotiating with individual businesses about terms of use
  • Navigating a patchwork of regulations
  • Physically hunting down rare sources — the team mentions reports of companies literally buying out distressed bookstores and visiting rare libraries just to find niche datasets

The compute market scales with money. The data market scales with strategy, relationships, legal work, and luck. A model that is a thousand times more data efficient isn't just academically interesting — it could be a thousand times easier to actually deploy in the real world.

Is Data Really the Moat in AI — and Can It Be Broken?

The talk references an audience poll where the most common answer to "what is the moat in AI?" was data. If that's true — if proprietary data is what keeps large incumbents ahead — then data efficiency is the lever that potentially democratizes AI development.

Right now, only a handful of organizations can train frontier AI models. Part of this is compute centralization, but a significant part is data centralization. The organizations with access to the largest and highest-quality datasets have a structural advantage that's difficult to replicate.

A data-efficient paradigm changes the calculus. If a model can achieve capable performance with far less training data, then smaller organizations, companies in specialized industries, and researchers in data-scarce domains can actually participate in building and deploying capable AI. The philosophical argument the founders make is pointed: if you care about who gets to shape the AI future, you should care deeply about data efficiency.

How Do Low-Level GPU Kernels Enable New AI Algorithms?

Here's where the technical argument gets interesting. Ben spent three years in a PhD program writing low-level GPU systems — "deep in the kernel mines," as he puts it — before co-founding Flappy. That background shapes the company's core technical thesis.

Think of it as three nested circles:

  • The outermost circle: everything a GPU can do efficiently
  • The middle circle: everything a framework like PyTorch can actually express efficiently
  • The research literature: what algorithms people have actually explored

Most AI research lives in that innermost circle — algorithms that are easy to express in PyTorch and happen to map well to what GPUs do. But there's a large space of things that GPUs can do that PyTorch simply can't express efficiently. Flappy Airplanes lives in that outer ring.

Their approach involves building a custom virtual machine that essentially takes over the GPU entirely, bypassing the abstractions that frameworks like PyTorch impose. This lets them implement algorithms that are otherwise impossible or prohibitively inefficient — things like fine-grained, deeply pipelined, hogwild-style training loops that have no good analog in standard frameworks.

The bet is that some of those hard-to-express algorithms are precisely the ones that enable data efficiency. New hardware primitives have driven most major machine learning breakthroughs over the past century — not always new chips, but new ways to extract value from existing hardware.

What Can't PyTorch Express That a GPU Actually Can Do?

PyTorch's great strength is also its constraint: it synthesizes a single-threaded programming model out of massively parallel processors. When you write matmul → attention → matmul → RMSNorm, PyTorch handles the parallel dispatch invisibly. For standard workflows, this abstraction is powerful.

But some algorithms require fine-grained, irregular parallelism that doesn't map neatly onto that abstraction. Highly branching workloads, fine-grained pipelining, and certain sparse operations can be asymptotically inefficient in PyTorch even if the underlying GPU could handle them well with direct programming.

This is the gap Flappy is targeting. By building their own VM-level framework, they can express computations that are genuinely novel — and they believe those computations are the key to training capable models on far less data.

What Is Flappy Airplanes and What Problem Are They Solving?

Flappy Airplanes — yes, that's actually the name, and no, they are not in the aviation industry (they've been fielding inquiries about literal runways) — is a three-month-old AI lab founded by brothers Ben and Asher, along with third co-founder Aiden Smith, a Thiel Fellow who spent three years commuting between college and their lab while studying both neuroscience and machine learning.

The company's mission is to design data-efficient AI through a combination of novel systems work and new algorithms. They're deliberately operating at the intersection of low-level GPU engineering and machine learning research — a combination they believe opens up a category of algorithms that simply isn't accessible from either direction alone.

They're actively hiring, and they're particularly interested in people with unconventional backgrounds. The founders' view: creative people who don't fit the standard mold have historically been the ones who change paradigms, not just optimize within them.

The Bottom Line on Data Efficiency

The case Ben and Asher make is layered but coherent. Data efficiency matters because the broad economy doesn't look like search and coding. It matters because data is harder to scale than compute. And it matters because the concentration of data is a key reason why only a few organizations can currently train frontier models.

Their approach — building new GPU primitives to express algorithms that current frameworks can't — is a principled bet that the next algorithmic breakthroughs are hiding in the space between what hardware can do and what our tools currently let us express. If they're right, the implications extend well beyond one startup.