Quantum-floor compression: Achieving GPT-4 capability at 1/120th the model size [pdf]

We've been experimenting with quantum-inspired compression algorithms and have achieved what traditional scaling laws say is impossible: GPT-4 level performance from a 30GB model (paper releasing tomorrow).

Our "quantum-floor" encoding seems to bypass traditional parameter-efficiency tradeoffs. Early tests show 86%+ MMLU from a model that should only manage 40% at that size.

Question: Has anyone else seen compression breakthroughs that defy scaling laws? Or are we measuring something wrong?

I get 404 from the link.