DeepMind new research: AI does not need to surpass all human experts to spawn giant oligopolistic corporations.

Beating AI News Flash: 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: as AI's capabilities in programming, mathematics, and other areas continue to surpass human experts, why haven't corporate organizations and the job market undergone similarly dramatic changes?

The researchers argue that the problem lies in the past focus on individual AI capabilities, neglecting the value created by enterprises through division of labor and collaboration. They had 100 people take tests covering six domains including mathematics and science, and found that those paired by complementary expertise performed 32% better than randomly paired individuals. Humans each have their strengths and can compensate for each other's weaknesses; replicating multiple identical AI models, however, often only yields a group of employees with similar weaknesses.

The paper further builds an economic model and finds that even if AI surpasses human experts in all professional tasks, as long as bottlenecks remain in collaboration between agents, enterprises will find it difficult to expand indefinitely, and the market may still consist of a large number of small and medium-sized enterprises.

Conversely, even if individual AI has not yet reached the level of top experts, as long as collaboration between agents breaks through a critical threshold, the situation could be completely different. When adding new Agents no longer drags down the team but instead continues to raise per-capita output, enterprises may break through traditional scale limits. Large-scale standardized markets such as software and finance may therefore concentrate toward a few giant AI enterprises.

Current AI still has a significant gap to reach such collaborative capabilities. A CooperBench test previously released by Stanford and other institutions found that having two programming Agents collaborate yielded an average success rate about 30% lower than a single Agent completing the same workload. They duplicate work, lose information, and even fail to correctly handle each other's code modifications.