DeepMind report weighs AI agents' coordination, not model intelligence

Google DeepMind published a report titled 'The Economics of Intelligence' and a 108-page working paper titled 'Organizing Intelligence' on 29 October, according to aitimes.com. The authors argue that the economic effect of AI will hinge less on how capable a single model is than on how well agents coordinate. The work is a theoretical analysis rather than a prediction built on real economic data.

The report was written by Alex Imas, DeepMind's head of AGI economics, researcher Andrew Koh, co-founder Shane Legg and others. They split AI capability into two parts: 'Competence', meaning the ability to carry out individual tasks, and 'Coordination', meaning agents dividing work among themselves and reaching shared goals.

The researchers lay out four stages of economic change from AI: partial task replacement, 'Supercompetent AI', 'Supercooperative AI' and artificial superintelligence (ASI). They argue that large-scale agent coordination could allow very large firms to operate with few human staff, and could damage the competitiveness of existing companies in software, finance and logistics.

Call for evaluations that test teamwork

The paper cites OpenAI vice president Noam Brown's experience running thousands of AI agents on math problems, where coordination accounted for under 10% of overall performance. The researchers call for AI evaluations that also measure agents' ability to divide work, share information and correct errors, and suggest AI's economic impact may appear first in wages rather than in productivity or employment figures.

In September, DeepMind published 'Artificial Symbiotic Intelligence', which described a future of collective intelligence formed by multiple AI agents combined with humans.