Gradient descent is taking away your Android developer job
summary
A provocative Andrej Karpathy tweet claims gradient descent can write code better than you and the talk shows concrete experiments: a neural net implemented with Kotlin DL in a Kotlin Jupyter notebook learns to approximate the Apollo 11 sine routine and, after about 1,000 epochs, matches the polynomial’s output (so it works but at compute cost). Practical tooling listed includes TensorFlow Lite and TensorFlow Hub for on‑device models, plus JVM libraries KMath and Krangl for data work.
AI-written summary · about
chapters (9)
- 0:00Introduction
- 1:08Software to zero era
- 10:44Neural networks basics
- 16:02Gradient descent explained
- 20:16Learning resources
- 21:00Kotlin for data science
- 23:53Machine learning on Android
- 24:57Live Kotlin demo
- 27:47Future of coding
related talks
46mGradient Descent: The Ultimate Optimizer
39mMachine Learning on Android Demystified
35mDerivatives - Important Concept. Simple to grasp in Kotlin
44mKotlin strives for Deep Learning
47mDeep Learning: a practical introduction from an Android dev perspective
41mOptimization Algorithms for Deep Learning
27mHow to Trick a Neural Network
19mGetting started with automatic differentiation