Current AI comparative tests are trying to keep up with modern models. Although they are helpful in measuring the performance of the model in specific tasks, it may be difficult to know if models trained in the field of internet data actually solve problems or just remember the answers they have already seen. Because the models are approaching 100% on some references, they also become less effective in revealing significant performance differences. We still invest in new and more difficult research, but on the way to general intelligence we must continue to look for new assessment methods. The newer shift towards dynamic, man -evaluated tests solves these issues of remembering and saturation, but in turn creates new difficulties resulting from the inseparable subjectivity of human preferences.
While we are still evolving and carrying out current AI test tests, we consistently want to test new approaches to models. That's why we present today Kaggle Game Arena: A new, AI public comparative platform, in which AI models compete in strategic games, ensuring a verifiable and dynamic measure of their capabilities.
