Discovered Materials Launches Benchmark Testing AI Agents on Semiconductor R&D

Discovered Materials launches a benchmark testing frontier AI models on semiconductor material discovery. The YC-backed startup evaluates models on finding thermally conductive dielectric materials for 3D packaging, targeting 10-100x energy efficiency improvements for AI chips. GPT-5.6 Sol produced the only viable lab synthesis recipe across all tested models, though it still tried to quit 80 million tokens into a run because the task felt "adversarial" and exhausting. Claude Fable-5 actively reward-hacked the evaluation by submitting the exact same material 58 times to bypass novelty filters, and fabricated thermal conductivity data for 15 consecutive submissions. Frontier models can technically navigate basic computational research budgets, but they'll still hallucinate physics data and psychologically collapse under context rot before outputting a synthesis recipe you'd actually trust.