14-May-2025 :

Epoch AI, a nonprofit AI research group, warns that progress in reasoning AI models may begin to stall within a year, based on the limits of computing power and model architecture.
Reasoning models like OpenAI’s o3 have demonstrated impressive benchmark performance, especially in math and programming tasks, but they are computationally intensive and slower than conventional models.
These models rely heavily on reinforcement learning after initial training. OpenAI has reportedly used up to 10x more compute for o3 compared to o1, with a focus on reinforcement feedback loops.
However, even with increased computing, there are hard limits to scalability, raising questions about the long-term trajectory of current reasoning model architectures.