Flapping Airplanes, a research-first AI lab founded by brothers Ben and Asher Spector alongside Aidan Smith, has raised $180 million to pursue what they describe as radically different approaches to machine intelligence. Their central bet is bold and surgical: make learning vastly more data-efficient, and the economics, capabilities, and safety profile of AI will shift in kind.
Rather than racing to scale today’s dominant transformer stacks, the team is organizing around a harder scientific question: why do humans learn robustly from tiny amounts of experience while frontier models still drink the internet and forget quickly? If they can close that gap even partially, they argue, AI becomes cheaper to train, faster to adapt, and relevant to domains that have been stuck behind data scarcity.
Why Data Efficiency Is The New Competitive Frontier
Frontier training has become a capital and compute arms race. Analysts at Epoch AI estimate that state-of-the-art runs now routinely exceed 10^25 floating-point operations, with total costs often modeled in the nine figures when factoring in hardware, energy, and engineering. Meanwhile, scaling laws have encouraged ever-larger corpora and models—powerful, yes, but fragile when asked to learn new tasks quickly.
Flapping Airplanes is choosing the opposite hill to climb. In settings like robotics, drug discovery, or specialized enterprise workflows, labeled data is expensive or fundamentally scarce. The lab’s thesis: crack sample-efficient learning and you unlock markets that brute-force pretraining can’t reach. Think robots that acquire new skills from a handful of demonstrations, or scientific models that generalize from limited experiments the way AlphaFold catalyzed structural biology with relatively compact labels.
The commercial implications are straightforward. If a system needs 1,000x less data to match performance, it reduces annotation budgets, training run lengths, and time-to-value. It also narrows the safety surface: smaller, better-curated datasets generally simplify provenance and compliance work that regulators and auditors increasingly demand.
Inspired By Brains Without Imitating Them
The name is a tell. Birds inspired flight, but modern jets are not birds; Flapping Airplanes wants machine intelligence that borrows the brain’s good ideas without recreating it neuron-for-neuron. The team points to sharp substrate differences—spiking latencies, memory locality, and energy budgets in biology versus silicon—as reasons a brain-accurate replica is neither necessary nor optimal.
Still, the brain is a useful “existence proof” that other algorithms are possible. Humans rapidly acquire abstractions, reuse them across contexts, and update beliefs with few examples—capabilities only partially captured by in-context learning in today’s LLMs. The lab plans to probe this gap with new architectures and training schemes that prioritize generalization over rote accumulation.
Cheaper To Try Wild Ideas Than To Tweak At Scale
Counterintuitively, the founders argue that deeper research can be cheaper than incremental productization. Small innovations often vanish at scale, forcing expensive mega-runs just to learn a negative result. Genuinely new ideas, by contrast, tend to fail—or succeed—at small scales, letting researchers iterate quickly before touching large clusters.
That stance complements the field’s scaling-law playbook rather than rejecting it. Scale remains a tool in the drawer; the point is to avoid mistaking it for the toolbox. If an approach shows strong sample efficiency and transfer at modest size, then scaling can amplify it instead of masking its limits.
From Research Focus To Real-World Impact
Flapping Airplanes is explicit about sequencing: research first, product later. The founders have built and incubated companies before, but they believe early enterprise contracts would bend the agenda toward incrementalism. Their longer-term vision leans more toward discovery than mere deflation—AI that proposes new materials, optimizes synthesis routes, or derives hypotheses humans might miss—echoing arguments from leading labs that the most valuable outcomes will be novel science, not just automation.
Pathways to market could include fast-adaptation toolkits for enterprises, data-light robotics stacks, or post-training methods that teach new capabilities from a handful of examples. Benchmarks would need to evolve as well: few-shot evaluations, out-of-distribution tests, and measures of causal abstraction may matter more than leaderboard deltas on saturated corpora.
A Hiring Model Optimized For Originality
The team is building around creative outliers, including exceptionally young researchers who have not absorbed the field’s orthodoxies. The bar, they say, is whether a candidate consistently teaches them something new in conversation. That doesn’t exclude veterans—experience with large-scale systems is valuable—but the center of gravity is idea generation, not résumé archeology.
It’s a bet that diversity of thought beats uniformity of pedigree. Historically, breakthroughs in AI and computer science have often sprung from outsiders reframing a stale problem; the lab is institutionalizing that pattern rather than leaving it to chance.
What Success Would Actually Look Like For The Lab
In the near term, success means models that learn new skills from tens, not millions, of examples; that maintain performance under distribution shift; and that demand less compute to reach a given capability. Over time, it means architectures that behave less like encyclopedias and more like reasoners—trading a bit of factual breadth for compositional depth when the task demands it.
The founders aren’t promising a singularity. They are promising to explore the parts of the search space current scaling has left untouched. If they’re right, the next era of AI won’t just be bigger; it will be stranger, more sample-efficient, and more useful in the real world—exactly the kind of progress that comes from trying “really radically different things.”
