Open AI is currently keeping the technology world in suspense with its chatbot ChatGPT, which can answer questions of all kinds with the help of artificial intelligence (AI). The system is based on the GPT-3.5 language model. For months, however, rumors have been circulating in the industry about a successor version GPT-4, which is said to be much more powerful.

Roland Lindner

Business correspondent in New York.

  • Follow I follow

This system was officially unveiled by Open AI on Tuesday, the company described it as its "latest milestone." It turned out that GPT-4 has been in use for weeks. The software company Microsoft, a close partner of Open AI, has now confirmed that the new version of its search engine Bing presented in February runs with GPT-4. The company had previously only said that it was using a more advanced variant of ChatGPT.

The most noticeable difference between GPT-4 and GPT-3.5 is that the new system can handle not only text, but also images. Both text and images can be entered into the program, with the response consisting only of text. As an example, a blog post mentioned a photo of the inside of a refrigerator. The system can not only identify the content, but also analyze what can be cooked with it. However, this image function is not yet widely available, there is currently only a waiting list for it.

Premium subscription for $20

However, GPT-4's usual text feature will now be made available via ChatGPT to users of the paid premium subscription, which costs $20 a month. At the beginning, the volume of use of the new system is to be limited.

Open AI describes GPT-4 as a significant improvement over GPT-3.5. While the new program is "less powerful than humans in many real-world scenarios," it demonstrates performance "on a human level" in various professional and academic fields. For example, it had passed a state examination in law and was among the top 10 percent. GPT-3.5, on the other hand, was at the worst 10 percent. In simple conversations, the difference between GPT-4 and GPT-3. 5 be "subtle". However, there was a noticeable deviation from a certain complexity of the task. Then GPT-4 is more reliable and also more creative.

As Open AI admits, GPT-4 still has "similar limitations" to previous language models. It is prone to "hallucination", so it provides wrong answers, albeit less than previous versions. And like GPT-3.5, the new model is not up-to-date and mainly only uses data up to 2021. It also does not learn "from its experience".