Analysis finds that improvements in AI models that’reason’ may slow down soon

The following are some of the ways to get in touch with each other According to an analysis conducted by Epoch AI (a nonprofit AI research institute), the AI industry might not be able achieve massive performance gains from reasoning AI models much longer. The report suggests that progress with reasoning models could slow within a year.

In recent months, reasoning models like OpenAI’s O3 have made significant gains in AI benchmarks. This is especially true for benchmarks that measure math and programming skills. The models are able to apply more computing power to problems which can improve performance. However, they also take longer to complete tasks than conventional models.

The development of reasoning models involves first training a conventional algorithm on a large amount of data. This is followed by a technique known as reinforcement learning which gives the model feedback on its solutions to complex problems. According to Epoch, so far, frontier AI laboratories like OpenAI haven’t applied a huge amount of computing power to reinforcement learning during the training of reasoning models.

That’s changing. OpenAI said it used around 10x as much computing power to train o3 compared to its predecessor, and Epoch speculates the majority of this computing power was devoted towards reinforcement learning. OpenAI researcher Dan Roberts revealed recently that the company’s plans for the future call for Prioritizing reinforcement learning will use much more computing power than the initial model training.

There’s a limit to how much computing power can be used for reinforcement learning.

Josh You explains that performance improvements from standard AI model are currently quadrupling each year, whereas performance improvements from reinforcement learning are tenfolding every 3-5 months. He continues that the progress of reasoning training “will probably converge with overall frontier by 2026.”

Epoch’s analysis is based on a number assumptions and includes public comments by AI company executives. It also makes the argument that scaling reasoning models could prove challenging for reasons other than computing, such as high overhead costs for researchers. You writes

: “If research overhead costs are persistent, reasoning models may not scale as far.” “Rapid computation scaling is potentially an important ingredient in reasoning models progress, so it is worth tracking this closely.” Studies have shown that reasoning model, which can be expensive to run, has serious flaws. For example, they tend to hallucinate a lot more than some conventional models.

www.aiobserver.co