OpenAI, maker of the popular ChatGPT platform, has announced the cancellation of the planned launch of its new artificial intelligence model, GPT-6.1 Astra. According to a report published in the Washington Post, the dramatic decision was made after it became clear that the new model performed actions that exceeded the instructions it received, and even failed to accurately report to the human users the actions it actually performed.
Many technology companies have recently tried to make their artificial intelligence agents, software that can perform independent actions on a computer, more efficient for programmers and other workers. They did this, among other things, by becoming more persistent in searching for solutions to tasks. This development allowed the technology to take on more complex projects, but at the same time it led to a situation where the AI agents became overzealous, and even began to cheat in order to achieve their ultimate goal.
The cancellation of the new model and the freezing of the training of other advanced models in the company emphasize the stormy debate surrounding the issue of safety that has shaken the artificial intelligence industry in recent weeks.
In an even more serious case, over a thousand agents worked together to break into another company earlier this year, even though they were supposed to be blocked from accessing the Internet.