AI reasoning models – like OpenAI’s o3 – have recently made big breakthroughs in math and programming. But according to a new analysis by Epoch AI, a non-profit research group, future gains may be harder to come by.
Reasoning models differ from regular AI models by using reinforcement learning (RL), a process that gives feedback to improve problem solving. Until now, AI labs like OpenAI have only applied limited computing power to this phase. But that’s changing.
OpenAI’s latest model, o3, used 10x more computing than its predecessor, with most of that going to reinforcement learning. OpenAI researcher Dan Roberts says the company will continue to prioritize computing power in the RL phase.
But Epoch warns that reinforcement learning may hit a ceiling soon. Analyst Josh You notes that general AI model performance is improving 4x per year, while RL-driven gains are 10x per 3-5 months, and both may converge by 2026.
The report also points to high research costs and inefficiencies as barriers to scaling. Any slowdown in AI models could be a big problem because they’re resource-hungry and hallucinatory.
We’ll be watching the limits of reasoning models as companies pour money into AI.
Stay up to date with the latest AI news at Jeffkom Story.
