I argue that hallucinations are a natural consequence of the language modeling objective, which focuses on simulating confident behavior even when that behavior is hard to predict, rather than predictable behaviors that take uncertainty into account. I also discuss five strategies for avoiding this mismatch.
What worked, and what didn't work, for my ICLR 2017 paper "Learning Graphical State Transitions".
An elegant (but hard to read) smartwatch watchface.
Debugging a hand-written processor with ModelSim, QEMU, and GDB.
Building generative models of jazz with inductive biases.
Visualizing some geometric algorithms.
A tiny time-reversal puzzle game.
Exploring the hidden activations in my music generative model.
Writeup for my first major machine learning project.
Adding color to an open-source watchface.