Impossible Biomimetic Synthesis: AI Reasoning in Engineering Design

We designed a hierarchical, constrained multi-agent synthesis framework to study how a language model reasons when tasked with engineering design. The system instructed a language model to synthesize a hydraulic rock breaker with a pistol shrimp. Given the deep incompatibility between these systems, any coherent synthesis of the two was extremely unlikely. This deliberate adversarial framing however created conditions for observing where model reasoning succeeds and where it degrades. The framework decomposed mechanical systems across six levels of abstraction, enforced provenance tracking against source specifications, and validated outputs against 22 physics and structural constraints. The system produced seven discrete designs of varying technical complexity. We observed design decisions reflecting genuine synthesis and creative problem-solving, but also spurious rationalizations and difficulty sustaining multi-step inference and managing long-range dependencies. Overall, our results suggest that language models can effectively navigate constrained design spaces, although nothing in our findings indicates they can perform the radical restructuring and reconceptualization underlying paradigm-shattering breakthroughs. AI's most productive contribution to engineering may be in accelerating the systematic exploration and incremental optimization that support the engineering process, while the breakout insights that drive profound advances still exceed the capabilities of current architectures.