Thousand Variables, Fifteen Runs

A preprint posted yesterday shows GPT-6 Astra redesigning a magnesium stent across 1,650 CAD parameters with a budget of just 15 finite-element simulations, reducing peak strain by 32%. A specialized industrial optimizer managed 10% on the same budget. The result makes a specific claim about why: LLMs carry engineering priors that let them make globally coordinated changes classical gradient-free methods can't.

Read more →