Files
Abstract
Innovation research has oscillated for decades between “technology-push” and “market-pull” explanations, yet empirical tests that isolate their joint dynamics remain scarce. We combine historical synthesis with a new methodological contribution: autonomous populations of large-language-model (LLM) agents that reason, converse and innovate in silico. After tracing how modern growth theory evolved from supply-side linear models to demand-sensitive, networked systems, we deploy four agent-based experiments. Survival-scarcity simulations show that moderate resource pressure spurs early, high-quality collective inventions, whereas abundance breeds complacency and extreme inequality multiplies but degrades innovations. Network-topology experiments reveal that fully connected societies exploit ideas quickly but converge prematurely, ring lattices preserve diversity yet diffuse slowly, and emergent small-world structures balance both, especially when agents display heterogeneous “engineer”, “artist” and “scientist” personas in a more complex simulation set up. In career-long academic ecosystems, we replicate 20 years of scholarship under three incentive regimes. Publish-or-perish rules maximize paper counts but generate roughly eight-times fewer breakthroughs, truncate researcher careers and accentuate Matthew-effect citation inequality. Five-year HHMI-style support delivers nearly an order-of-magnitude more breakthroughs, sextuples new paradigms, preserves topic diversity and keeps over 90% of scholars active; a dual-tier system lands in between. Across experiments, innovation thrives in a Goldilocks zone: sufficient urgency to provoke action, structural diversity to explore alternatives, and incentive horizons long enough to reward risk. Our results align with decades of organizational and science-policy scholarship, yet are produced by agents drawing only on self-consistent language priors—demonstrating that LLM societies can serve as reproducible, high-throughput “wind tunnels” for social-scientific theory. Limitations and for this emerging methodology are also discussed.