Session
LLMs are compressors, so we will chat with gzip
I chatted with gzip. It answers questions about me with zero learned weights, and says "i don't know" when a question surprises it too much. In this talk you'll chat with it too, live.
It works because compression is prediction: a compressor only saves bits on what the receiver can already guess. I follow that idea across the stack:
1. The web: gzip, Brotli and zstd, and shared dictionaries that turn last week's file into the guesser for this week's.
2. Doom running in AWS Lambda, and why Doom in 1993 sent only key presses.
3. LLMs: next-token prediction and cross-entropy loss are code length, so a language model is a compressor.
4. Then the demo, under the hood: gzip picks each next character by what compresses smallest, compression distance finds the closest stored question, and surprise becomes a refusal. And where it breaks: gzip matches bytes, not meaning.
You'll leave knowing what to tune for your own payloads, and what an LLM's parameters actually buy.
Please note that Sessionize is not responsible for the accuracy or validity of the data provided by speakers. If you suspect this profile to be fake or spam, please let us know.
Jump to top