A Google DeepMind study analyzes a decentralized market in which multiple AI agents bid to answer users’ prompts. The researchers report that first-price auctions achieved 99.2% of the performance of an omniscient allocator on the EmbedLLM dataset. This efficiency is attributed to competition among models with similar capabilities and to the limited correlation between how much each agent values a given prompt. The paper also establishes a theoretical price of anarchy of approximately 86.5%, compared with about 63% for first-price auctions in the general case. The findings nevertheless remain limited to one dataset, a typical equilibrium and a competitive structure that could weaken if the model market became concentrated around a few dominant players.
Source: Read the original article
