Gaudí’s upside-down church offers a vivid explanation of backpropagation
theteknosaur · x · 2026-07-26
This long article uses Gaudí’s upside-down church model to build an equation-free explanation of how neural networks learn.
- It starts with the Sagrada Família and the Colònia Güell church model, where strings and weights were hung upside down to let gravity find the natural shape of the structure.
- The author maps that physical process to backpropagation and gradient descent: one point moves, its neighbors react, and the whole system settles into equilibrium.
- The key analogy is that the network, like the hanging model, does not solve everything at once; it iteratively adjusts until the forces balance.
- The piece emphasizes that Gaudí was using a physical simulation of forces decades before modern computing, which makes the analogy especially intuitive for language models learning from mistakes.
More from AGI Musings
- Sam Altman says the modern corporation, not technology, was the Industrial Revolution’s key invention — garrytan · 2026-07-26
- Why entity mapping works for Google but not for a model’s parametric memory — steveplunkett · 2026-07-26
- The real AI jobs question is which work still needs human judgment and ethics — AryHHAry · 2026-07-26
- AI agents could overwhelm public infrastructure, says Yacine MTB — yacineMTB · 2026-07-26
- A Tim Bresnahan talk says coordination frictions may slow AI’s scientific diffusion — JMateosGarcia · 2026-07-26
- A study argues LLMs are creating the same property-rights conflicts as past frontier technologies — JMateosGarcia · 2026-07-26