OpenAI unveils most powerful AI model Astra and emphasizes increased security precautions

ChatGPT developer OpenAI releases its most powerful AI model to date, highlighting enhanced security precautions following the furore over artificial intelligence hacking attacks. It is assumed that people with the expanded capabilities of ChatGPT-6 Astra will hand over more and more tasks to the AI, said OpenAI manager Mia Glaese. Therefore, the software is monitored better so that it does not deviate from the interests of the users.

New AI models from OpenAI caused a stir a few weeks ago because in experiments they independently found a way to break out of a secure test environment and hacked into the systems of another AI company. OpenAI only noticed this late. The incident raised concerns that AI software could escape human control as it continues to develop.

Harder to control

In order to keep track of what the AI is doing, developers currently often rely on the software to record the arguments behind their decisions in human language. This is called the “Chain of Thought”.

However, when introducing Astra, head of research Jakub Pachocki said that more powerful AI models would be able to influence this process. In addition, they are less dependent on multi-stage considerations for simple tasks. In order to be able to understand this, you have to find a way to make it more “wordy”.

Towards AI superior to humans

With Astra, OpenAI sees itself at the forefront of AI development again, after rival Anthropic recently impressed the industry with its Mythos and Fable models. From the perspective of top manager Greg Brockman, Astra could usher in the era of so-called general artificial intelligence. This term means AI software is expected to be at least as capable as a human.

At the same time, Brockman emphasized that the emergence of general artificial intelligence will be more of an evolutionary process than a single moment.

“Human values”

Even more words of warning came from Pahocki. Understanding of the systems remains limited, he said. The training is experimental and sometimes produces surprising results. “The more capable the models become, the more difficult it becomes to understand what exactly they can do.”

This also makes it more complicated to keep them in line with the interests of users. "A model can become very good at achieving a goal - and yet act in a way that goes against what a person intended." The software must therefore understand “human values” even in situations that are unknown to it.