AI: liar, master of geometry and referee support

Researchers from the Anthropic startup checked whether chatbots such as ChatGPT can lie. Not only can adequately trained models learn to deceive humans and other artificial intelligence models, but it also turns out that this process cannot be reversed using the security measures available for advanced AI systems. That’s why scientists warn against a “false sense of security” about the threats associated with artificial intelligence.

The Google Deepmind team presented the AlphaGeometry artificial intelligence system that can solve complex geometric problems at the level of an International Mathematical Contest champion. Scientists reported that the program correctly solved 25 of 30 problems from the previous contests, a result similar to that of contest gold medalists. Currently, the system is limited to specific types of geometry, but researchers believe it could eventually be used in various branches of mathematics.

AI has also entered the world of gymnastics. The Judging Support System used by the International Gymnastics Federation helps you notice nuances that are invisible to the eye. Using a high-resolution camera system, a three-dimensional image of moving athletes is created, allowing for a precise assessment of their movements. Supporters of the technology also hope that it will eliminate the problem of judges’ biases and the currently opaque grading system.