In a recent analysis from Epoch AI, a nonprofit research institute, it appears that the impressive gains in reasoning AI might be hitting a plateau. The report forecasts a noticeable slowdown in these models within the next year, prompting us to take a closer look at what this means for the industry.
Reasoning AI models — including OpenAI’s o3 — have been excelling in tests for mathematical problem-solving and programming. These systems harness enormous computing power to fine-tune their capabilities. Unlike traditional models that improve more predictably, these advanced systems take considerably longer to reach peak performance.
The process generally starts with training a conventional AI model on vast datasets before shifting to reinforcement learning techniques. Reinforcement learning provides iterative feedback on problem-solving, yet Epoch points out that leaders like OpenAI have not fully capitalised on this approach.
For instance, OpenAI disclosed that training o3 required roughly ten times the computing power used for its predecessor, o1. Epoch suspects that most of this extra capacity was dedicated to reinforcement learning, a trend reinforced by comments from OpenAI’s Dan Roberts, who signals an upcoming focus on boosting this phase of training.
However, there is a natural cap on how much computing power can improve reinforcement learning outcomes. Analyst Josh You notes that while standard AI model performance is set to quadruple each year, breakthroughs via reinforcement learning have been occurring every few months. He predicts that by 2026, the progress in reasoning models will likely align with the broader pace of AI development.
Epoch’s analysis blends industry assumptions with insights from executives, highlighting challenges beyond just raw computing power. Rising research costs, for example, could limit the scale of reasoning models—a crucial consideration given the substantial investments made so far. This reality is compounded by issues such as higher operational expenses and a tendency for these models to occasionally generate inaccurate information.
If you’ve ever wrestled with the promise of rapid tech progress, these findings serve as a good reminder: while the rush to scale up AI is real, pushing past certain limits might require new strategies.
