Analysis by Epoch AI, the nonprofit AI institute, shows that the AI ​​industry may not be able to get longer from inference AI models. According to the report's findings, progress in the inference model will slow down within one year.

Inference models such as OpenAI's O3 have seen considerable growth in AI benchmarks in recent months, especially benchmarks for measuring mathematical and programming skills. These models can apply more computing applications to problems that can improve their performance, with the disadvantage of them being longer than traditional models to complete tasks.

The inference model is developed by first training a normal-scale model of large amounts of data, and then applying a technique called reinforcement learning that effectively provides “feedback” for its solutions to difficult problems.

According to Epoch, so far, Frontier AI labs like OpenAI have not applied a lot of computing power to the reinforcement learning phase of inference model training.

This is changing. Openai said that it applies to training O3 to about more computations than its predecessor, O1 and the era training O3, and speculates that most computations are committed to strengthening learning. OpenAI researcher Dan Roberts recently revealed that the company's future plans require priority to reinforcement learning to use more computing power, even more than the initial model training.

However, how much calculation can be used for reinforcement learning in each period.

Epoch analyst Josh You, who is also the author of the analysis, explained that performance training for standard AI model training is currently improving twice a year, while enhanced learning performance increases tenfold every 3-5 months. He continued that progress in reasoning training will "may be integrated with the overall boundary by 2026."

Epoch's analysis makes many assumptions and draws in part on public comments from AI company executives. But this also makes scaling inference models likely to prove challenging, including high overhead research costs in addition to computation.

“If the research requires ongoing overhead, the inference model may not scale as expected,” you write. “Fast calculation of scaling can be a very important element of the progress of the inference model, so it is worth keeping a close track of this.”

Any sign that a reasoning model may reach some kind of limit in the near future may worry the AI ​​industry, which has invested a lot of resources to develop these types of models. Studies have shown that running inference models that can be very expensive have serious flaws, such as hallucinations that are more hallucinating than some traditional models.