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

Analysis Suggests AI Industry May Face Challenges with Reasoning Models in the Near Future

The AI industry may soon encounter limitations in achieving significant performance gains from reasoning AI models, according to a recent analysis by Epoch AI, a nonprofit AI research institute. The report indicates that progress in reasoning models could potentially slow down within a year, signaling a potential shift in the landscape of AI development.

Reasoning models, such as OpenAI’s o3, have demonstrated notable improvements on AI benchmarks in recent months, particularly in tasks related to math and programming skills. These models have the capability to leverage more computing power to enhance their performance, albeit at the cost of longer processing times compared to traditional models.

The development of reasoning models typically involves training a conventional model on extensive datasets, followed by the application of reinforcement learning techniques to provide the model with feedback on solving complex problems effectively.

While AI research labs like OpenAI have not historically allocated substantial computing resources to the reinforcement learning phase of reasoning model training, this approach is changing. OpenAI recently disclosed that they utilized approximately ten times more computing power to train o3 compared to its predecessor, o1, with a significant portion of this computing power dedicated to reinforcement learning. Additionally, OpenAI’s future plans emphasize prioritizing reinforcement learning with even greater computing power allocation than initial model training.

Josh You, an analyst at Epoch and the author of the analysis, highlights that while performance gains from standard AI model training are quadrupling annually, gains from reinforcement learning are increasing tenfold every 3-5 months. This trend suggests that advancements in reasoning model training may align with industry norms by 2026.

Epoch’s analysis underscores potential challenges in scaling reasoning models beyond computing power, citing high overhead costs for research as a key consideration. The report suggests that if there are persistent overhead costs associated with research, reasoning models may not achieve the scalability anticipated. The analysis emphasizes the importance of closely monitoring rapid compute scaling as a critical factor in the progress of reasoning models.

Concerns about the potential limitations of reasoning models in the near future are likely to raise alarms within the AI industry, which has made substantial investments in these models. Studies have already revealed that reasoning models, despite their significant computational requirements, exhibit certain drawbacks such as a propensity to generate erroneous outputs more frequently than traditional models.