Live experiments and interactive results.
Ten frontier open models scored on compression (bits per byte) of provably human-written text and code published after their training cutoffs — parliamentary speech, AI-ban open-source commits, and undisclosed canary slices. Includes the GPT-OSS anomaly and a code-model regression.
Models with knowledge cutoffs from Sep 2021 to Feb 2026 forecast valuations of OpenAI, Anthropic, NVIDIA, Alphabet and Meta — versus what actually happened. Plus: why point estimates mislead, hindsight calibration, and what tail shapes LLMs actually believe in.