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

One analysis EPOCH AI, the Non-Profit AI Research Institute, indicates that the AI ​​industry cannot be gained by a long time-based models. During the year, according to the findings of the report, the progress of the thinking models may slow down.

Openai’s Thought Models O3 Prices measuring significant gains, math and programming skills in AI Benchmarks in recent months. Models can apply more computing to the problems that improve their work, which has been missing so that they take longer than ordinary models to complete the tasks.

Provisional models, first training, develop a conventional model for a larger amount of information, then apply a technique called “Review” in solving difficult problems, by applying a technique of strengthening learning.

To date, the border EU laboratories have not been applied to a large calculation power for the strengthening learning stage of the Model Training of Model Training according to EPOCH.

This is changing. Openai said that the majority of this calculation is dedicated to the learning of the majority of this calculation, more than 10,000 computing approaching O3 and Epoch. And Roberts from Openai researcher, recently announced the call of the company’s future plans Prioritize the learning of reinforcement More for initial model education to use more computing power.

However, how much computing for the learning of reinforcement in one period is still closed as it can be applied.

Analyst, an analyst, the author of the EPOCH and analyst, explains that performance earnings from the standard AI model training are used every year in four times each year. The progress of training training “will probably unite in the general border by 2026,” he said.

Epoch’s analysis is a number of assumptions and participates in public statements of the AI ​​managers. However, it also proves that the scale models of medical models can be caused by the calculation, including high surface expenses for research.

“If there are a continuous spending costs required for the research, justification models cannot scale as expected,” he said. “Fast calculation scale is a very important ingredient to make the model progress, so it is worth watching closely.”

Any signs that the models of models will reach a kind of range in the near future will concern the AI ​​industry that invests large sources developed by such models. Studies have already shown the main models incredibly expensive to escapehave serious imperfections such as an inclination More hallucinat than certain ordinary models.