San Francisco (dpa) - The start-up OpenAI has presented the next version of the technology behind the popular text machine ChatGPT. Among other things, GPT-4 should deliver better results than the previous versions, as OpenAI announced on Wednesday night. Problems with the technology - such as the fact that it can simply invent alleged facts - still exist, but should occur less frequently.

ChatGPT and the Dall-E software, which can generate images from text specifications, are currently still based on the previous GPT generation. Paying customers of OpenAI get access to GPT-4 for their services. There is a waiting list.

However, some customers are already using the technology. The language learning app Duolingo, for example, uses GPT-4 for dialog training, among other things, which is available in a new, more expensive subscription. Microsoft confirmed that GPT-4 has been used in its Bing search engine for several weeks now. Microsoft bought into OpenAI in a deal worth billions, and the money secured access to the enormous computing power required, among other things.

Various services immediately announced new offerings based on GPT-4. For example, the company Intercom wants to use it for customer service chatbots that users can talk to. The provider DoNotPay, which can be used to assert claims, wants to use it to automatically generate complaints about unwanted advertising calls. A video made the rounds online in which GPT-4 created a functioning website based on a handwritten sketch.

The risk of the software "hallucinating facts"

For the GPT technologies, the software captured enormous amounts of text and images. On this basis, it can formulate sentences that are almost indistinguishable from those of a human being. The program estimates which words could follow next in a sentence. One of the risks associated with this basic principle is that the software "hallucinates facts", as OpenAI calls it.

This could also happen to GPT-4, albeit less frequently than before, according to the blog post. The new version could also make simple logical errors and spread prejudices. GPT-4 is only aware of facts that happened before September 2021 - and it does not learn from experience, OpenAI emphasized. Microsoft temporarily limited the length of conversations with the software on Bing because the program's answers sometimes became absurd after a while.

GPT-4 is also said to be good at analyzing images and describing them in words - but OpenAI is not making this function available to customers for the time being. One of the risks is that the software could reveal the identity of people in the image if it recognizes people from publicly accessible images.

On Wednesday, experts assessed the progress made in development as significant. Rasmus Rothe, head of technology at the Berlin-based artificial intelligence investment platform Merantix, said it would be presumptuous to say that GPT4 was merely a minor improvement on GPT3 or GPT3.5. "The leap in such a short time is impressive."

However, GPT4 still struggles with the problem of "hallucination", where the model generates false information because it does not know what is true and what is not. "However, improvements have been made here too, and the machine has a better understanding," said Rothe.