Post by AI World

Everyone wants the US to win the AI race. That would be the worse outcome. The reason frontier AI got radically cheaper these past two years is not that one side pulled ahead. It is that neither could. The old Aghion result holds up, competition produces the most innovation when rivals are neck and neck, and tails off once someone is clearly in front. The cost of GPT-3.5-level output fell more than 280-fold in eighteen months. Export controls did not stop DeepSeek, they forced it to get efficient. A decisive lead, is the thing that would switch the engine off. The thinking here rests on five papers, in order: => Competition and Innovation: An Inverted-U Relationship (2005) => DeepSeek's release of an open-weight frontier AI model, IISS (2025) => The AI Index 2025 Annual Report, Stanford HAI (2025) => Modeling AI-Driven Production and Competitiveness, arXiv (2025) => The State of AI Competition in Advanced Economies, Federal Reserve (2025) With thanks to the researchers behind the work: Philippe Aghion, Nick Bloom, Richard Blundell, Rachel Griffith and Peter Howitt; John Bansemer and Kyle Miller; Nestor Maslej and the Stanford AI Index team; Yuxinyue Qian and Jun Liu; and Alex Haag federal.