Deep Neural Nets: 33 Years Ago and 33 Years From Now

Source: https://karpathy.github.io/2022/03/14/lecun1989/ Author: Andrej Karpathy Date: 2022

Summary

Karpathy recreates LeCun’s 1989 handwritten digit network with modern tools (PyTorch, auto-grad), comparing what was hard vs. trivially easy 33 years later. Uses this as a lens to predict what 2056 might look like for today’s networks.

Key Claims

  • LeCun 1989: trained a neural network in Lisp with manually-computed gradients, took days
  • Same experiment with PyTorch 2022: minutes, 10 lines of code
  • Insight: the algorithms haven’t changed much — the tooling and scale have
  • Prediction: 2056 will look back at 2022’s manual prompt engineering and custom architectures as similarly primitive
  • The “bitter lesson” applied: compute + simple algorithms win over clever human engineering

Entities

Concepts