Exclusive: Paying for frontier AI models buys 4-month head start at 5x the cost
Ars previewed Mozilla’s report on how cheap open models caught up on capability.
TL;DR
- The performance gap between frontier AI models from US tech companies and open-weights models from Chinese companies has narrowed to just 4.4 months.
- Many organizations are shifting to cheaper open models for routine work due to their cost-effectiveness and closing performance gap.
- Frontier models remain valuable for expert professional work, high-intensity retrieval, and long context tasks.
- Open-weights models allow for greater accessibility and customization but often lack proprietary training data and code.
- Closed frontier models offer convenience, compliance, and support, which are beneficial for organizations lacking the staff to manage open models.
- The cost of closed frontier models can be up to five times higher per task compared to open models.
- Chinese companies are leading in the development of open-weights AI models, mirroring strategies seen with Android.
- There is a call for US and European labs to compete in the open AI space to prevent concentration of power.
- The development of open AI models could benefit from a coalition model similar to that of open-source software, involving public compute programs, foundations, and companies.