Improvements in ‘reasoning’ AI models may slow down soon, analysis finds

AI reasoning models may face performance limits in the near future as the industry shifts toward more intensive use of reinforcement learning, according to Epoch AI's analysis. The organization highlights that progress from standard model training is quadrupling annually, while reinforcement learning gains are growing tenfold every 3-5 months. Josh You at Epoch AI notes that scaling reasoning models will likely align with broader industry advancements by 2026, but also emphasizes the potential challenges of high overhead costs and other systemic limitations beyond computing power.