Skip to content
hackertubeai

Computerphile

Glitch Tokens - Computerphile

About this video

Rob Miles explains "glitch" tokens, odd strings in large language model vocabularies that make the model produce gibberish. He covers how they likely arise from tokenizer training data, such as Reddit usernames, that never appeared in the model's own training.

Keep your curiosity going.

A few more ideas from the same corner of the archive.

The spelled-out intro to neural networks and backpropagation: building micrograd

Andrej Karpathy

But what is a neural network? | Deep learning chapter 1

3Blue1Brown

A Better Default Colormap for Matplotlib | SciPy 2015 | Nathaniel Smith and Stéfan van der Walt

Enthought

How Self-Driving Cars will Destroy Cities (and what to do about it)

Not Just Bikes