AI Models Will Tap the Whole Economy in 2 Years | Aidan Smith

Summary · 3 min read

Aidan Smith is cofounder of Flapping Airplanes and previously worked at Neuralink. Flapping Airplanes is building data-efficient AI models by redesigning GPU primitives and model architecture from scratch; the episode describes the company as raising $1 billion.

Smith argues that data-efficient AI could extend models into real-world tasks that current systems handle poorly, opening up what he calls the other $30 trillion of the economy. He says Flapping Airplanes is building custom GPU primitives and a vertically integrated stack to develop algorithms suited to the hardware, rather than copying the brain literally. He expects broader AI use to increase compute demand and says the lab takes safety seriously through observability and scaling science. Smith sees room to coexist with frontier labs by tackling data-limited problems, and says the company may commercialize fast models when it is ready.

Minute by minute

1. 0:31

   Models should learn with less data. Smith says Flapping Airplanes wants to build models that are less rigid and more like natural intelligence. He says the team draws inspiration from the brain, including his prior experience at Neuralink, and asks why models do not learn as efficiently from data as humans do.

2. 2:00

   Data efficiency targets real-world value. Smith says current AI touches only a tiny fraction of the economy, while many valuable tasks require models to learn or adapt from limited data. He says the lab aims to reach the other $30 trillion and sees an opportunity to start from scratch rather than compete on the same approaches as larger labs.

3. 2:57

   GPU primitives shape algorithms. Smith says the brain is an inspiration for useful properties, not a blueprint to copy neuron for neuron, because the best algorithm depends on its hardware. He compares the brain’s 150-millisecond cross-brain signal time with GPU operations measured in microseconds, and argues that new GPU primitives can enable different algorithms.

4. 6:01

   Brain emulation faces two barriers. Smith says whole brain emulation first requires data that current scanners cannot capture with enough temporal detail; even a Neuralink device would cover only a localized area. He also says faithfully running a brain model is extremely difficult because the brain’s physical dynamics do not map cleanly to today’s computing substrates.

5. 9:15

   A custom stack connects research and deployment. Smith says Flapping Airplanes built its software stack from scratch and does not use Torch or NVIDIA’s normal CUDA runtime. He says close integration between systems, pre-training and post-training lets research models run with a production inference stack; he would not commit to a model release date.

6. 11:05

   Compute demand could grow sharply. Smith says tapping the 90% of the economy AI does not reach today could create much more compute demand, even if models run 1,000x more tokens. He says the lab takes safety seriously and is working on observability and scaling science so it can scale responsibly.

7. 13:32

   Frontier labs leave room for new approaches. Smith says Flapping Airplanes can coexist with frontier labs, which he respects, because he believes there is still important research to discover. He sees an opening in areas with limited data and says the lab need not follow the same path as companies scaling coding models.

8. 14:56

   Model skills vary by task. Smith says models are strong on some tasks but struggle when work becomes novel, where they can make rookie mistakes. He argues that people should use models where they are strongest while people retain work such as idea generation and deploying systems in the world.

9. 17:22

   Sparse data exposes model gaps. Smith points to DNA, robotics and trading as areas with little training data. He relays cofounder Ben’s estimate that models cover under 1% of the world’s data and knowledge, including bespoke operational know-how, and says models still produce generic interview questions.

10. 18:51

    Training can narrow model behavior. Smith says reinforcement learning from verifiable rewards can collapse model entropy, raising the challenge of maintaining diversity as models scale. He also says text pre-training can help models learn from a small number of image or video examples, supporting transfer and analogical reasoning.

11. 20:22

    Ownership attracts unconventional hires. Smith says the lab appeals to people who want visible responsibility and whose work can materially affect a project. He cites a Clash of Clans world champion as an example of an unconventional hire and says credentials such as math olympiad medals alone do not guarantee machine-learning ability.

12. 23:01

    Fast models may shape product strategy. Smith says Flapping Airplanes may offer models that match some frontier or open-source models at different points on the throughput and interactivity curve. He says the lab could sell very fast inference, but does not want to get stuck in smaller markets without a path to its longer-term ambition.

Published Oct 22, 2026.