Lossless text compression via arithmetic coding over LLM probability distributions, with lightweight error correction for GPU non-determinism.
Detect steganographic data exfiltration by verifying that LLM-generated text honestly came from a specific model. Reduces exfiltratable information by ~100x. Live Demo runs inside an Nvidia H200 Trusted Execution Environment.
Live demo for "Question-Asking" Compression - a novel technique we invented in our paper which enables the compression of complex text in just 10 bits through the power of question asking! Check out the demo for a live demo of Haiku asking Opus questions.
I scraped a bunch of public datasets and found credentials - some of which are valid! password is `demo` (any username).
A webapp that lets you paste in any privacy policy and grades it for you, highlighting key concerns and scoring overall privacy practices.