Insight Beating AI News: Google DeepMind co-founder Shane Legg, along with economists Alex Imas and Andrew Koh, have published a paper on the economic impact of AGI, attempting to explain a paradox: despite AI's growing capabilities in programming, mathematics, and other fields surpassing human experts, why haven't enterprises and labor markets undergone equally dramatic transformations?

The researchers argue that the issue lies in past overemphasis on individual AI capabilities, overlooking how organizations create value through division of labor and collaboration. They conducted tests involving 100 participants across six domains including mathematics and science, finding that teams paired based on complementary expertise outperformed randomly paired groups by 32%. Humans possess diverse strengths and can compensate for each other’s weaknesses; however, deploying multiple identical AI models often results in a group of agents sharing similar flaws.

The paper further develops an economic model, showing that even if AI surpasses human experts in all professional tasks, as long as coordination bottlenecks persist among intelligent agents, companies will struggle to expand infinitely, and the market may remain dominated by numerous small and medium-sized enterprises.

Conversely, even if individual AIs have not yet reached top expert levels, once agent collaboration crosses a critical threshold, outcomes could drastically change. When adding new agents no longer hampers team performance but instead continues to boost per-capita output, companies may break traditional scale limitations. Large-scale, standardized sectors such as software and finance might then consolidate into a few dominant AI-driven enterprises.

Current AI systems still fall significantly short of this collaborative capability. Previous evaluations like Stanford’s CooperBench test revealed that two programming agents working together achieved an average success rate about 30% lower than a single agent completing the same workload. They repeated tasks, lost information, and failed to properly handle each other’s code modifications.