2017, An 11-layer neural network learned to fold a towel after just 21 training attempts

Originally published on Substack.

Tokyo Big Sight , November 30, 2017, An 11-layer neural network learned to fold a towel after just 21 training attempts.

Provide a caption (optional)

Back then, people were amazed by the breakthrough of a robot mastering a single, specific task.

Fast forward to 2026. Embodied AI companies are raising billions of dollars, while compute, models, and datasets have improved by several orders of magnitude. Yet towel folding remains one of the industry’s favorite demonstration tasks.

From an entrepreneurship and investment perspective, technological progress is not always linear. Often, a decade is spent not answering “Can it be done?” but rather “Can it be done reliably, affordably, and at scale?”

I don’t chase every wave, nor do I mock every bubble.

What truly matters is not whether a robot can fold a towel today. What matters is who, ten years from now, can turn towel folding into a business, technology into value, and a demo into an industry.