Drink Me: (Ab)Using a LLM to Compress Text

An exploration of text compression using a large language model trained to predict text sequences. The author tests the ability to extract and reproduce copyrighted text, achieving significant compression on known texts like 'Alice's Adventures in Wonderland'.

• LLMs may allow text compression.
• Chapter 1 of 'Alice' was compressed 8%.
• The full text achieved 15% size.
• Decompression was successful.
• Consistent results on GPU & CPU.