Livescience: Live Science: Quantum Boost for AI: Less is More?

Summary

Imagine reducing AI uncertainty not by adding massive computing power, but by using quantum technology. Researchers at Multiverse Computing have demonstrated a method to achieve "quantum enhancement" in a production-scale large language model. They focused on a metric called perplexity, or PPL, which measures how well an AI predicts the next word. Lower PPL means a better, more coherent AI. Traditional methods to lower PPL involve massive increases in model size and computational infrastructure. However, this new technique uses small, trainable quantum circuit blocks called Cayley-parameterized unitary adapters, or CUAs. When these CUAs were added to Meta's Llama 3.1 8B model, perplexity decreased by one point four percent, while the model only needed six thousand additional parameters – a tiny increase. Crucially, this quantum-enhanced model also answered questions correctly that the original model got wrong, such as identifying all jovian planets as ringed in an astronomy example. This proof-of-concept suggests a future where quantum computers can help overcome limitations in classical AI scaling and potentially achieve "quantum advantage," where quantum systems outperform classical ones.