OpenAI brings the operator to Germany
OpenAI has started rolling out the AI agent Operator in the EU. This is based on a computer-using agent.
(Image: Novikov Aleksey/Shutterstock.com)
AI agents should be able to act and thus take over one or a number of tasks for humans. At OpenAI, the general AI agent is called Operator. It is now also coming to Germany as a preview. It has been available in the USA since the end of January. Operator is intended to be helpful in both private and professional environments. Just the day before, OpenAI made a new API available that developers can use to build their own versions of the AI agent.
Videos by heise
The operator is primarily based on the computer-using agent (CUA). This in turn is a model that is based on GPT-4o and combines it with reinforcement learning. The latter means reinforcement learning, which means that a model searches for paths and answers that are rewarded and thus learns what is desired. CUA has also been trained to interact with graphical user interfaces. For example, the agent must be able to recognize text field inputs to use them.
Operator books you a table in a restaurant
OpenAI writes that “it is the beginning of a future in which AI not only provides information, but also performs workflows independently – to support companies and private individuals alike.” The operator has already been tested with companies such as Booking, Expedia and Uber. Since the hype surrounding generative AI, AI providers have been dreaming in their announcements that AI, whether in the form of plugins or as an agent, will be able to book tables in restaurants for people or even an entire vacation. To achieve this, the AI agent must, of course, be equipped with information such as payment details. In such a case, however, the operator asks whether it should use them, for example. The same applies to log in data and CAPTCHAs – in both cases, the human must approve the input in the background.
In the future, the operator's research preview will be available to people with a Pro account aged 18 or over in the EU, Switzerland, Liechtenstein, Iceland, and Norway. Preview means limited capabilities, but also the opportunity to provide feedback. The operator still has its domain, but will soon be integrated into ChatGPT. OpenAI describes all security risks and the measures taken in an Operator System Card.
(emw)
