OpenAI introduces improved version GPT-4

GPT-4 is intended to better solve a core problem of AI-supported so-called large language models. So far, the algorithms have produced impressive text, but they can incorporate false statements and invent sources. Experts speak of “hallucinating”. GPT-4 is said to work much more reliably than the previous system.

OpenAI’s GPT-4 should be able to answer more factually

OpenAI said on Tuesday that the tool was “40 percent more likely to produce more factual answers than GPT-3.5, based on our internal evaluations.”

The release of the new version was widely discussed on Twitter on Tuesday. “GPT-4 is a bigger step than GPT-3 represented,” wrote OpenAI developer Jason Wei. GPT-4 achieves human qualities in many tasks and will trigger “a societal revolution where AI reaches every industry, starting with the technology industry”.

Referring to the results of several access tests, Jim Fan, AI researcher at chip manufacturer Nvidia, explained: “The truth is that GPT-4 can already enroll at Stanford as a student.” The judgment of the AI ​​exceeds the usual scales. “Exponential growth is a scary thing, isn’t it.”

“Air up” when hallucinating

The new AI version was also a direct topic at the important technology fair “South by Southwest” (SXSW) in Texas. “We are very excited that another AI milestone has been reached with the release of the new GPT 4 version,” said the head of the US photo app Prequel, Timur Khabirov, the Handelsblatt.

The new capabilities “could be game-changers for generative image processing,” according to Khabirov, “particularly the ability to convert images to text, which will take AI applications in new directions.”

Founded in 2017, Prequel uses AI to create alienated, artistic images and videos based on real-world templates. Celebrities like Katy Perry and Bella Hadid, but also fashion groups like Gucci are already using the application.

“ChatGPT is now smarter than most lawyers in the US and can compete at the top of a biology Olympiad,” said Christian Byza, head of the AI ​​learning platform Learn.xyz. The launch of GPT-4 is “sensational news” for his company, according to Byza. “GPT-4 is significantly smarter” is one of the top ten percent of participants in the US bar exam.

GPT-4 is also more eloquent, can better deal with irony and jokes and summarize sources. This makes generated texts more entertaining and of higher quality. In addition, GPT-4 can work not only with text but also with images, said Byza. The own courses based on GPT-4, for example history seminars, would be significantly better. The downside: GPT-4 is 40 percent better at “not making things up”, but there is “still a lot of room for improvement”.

“GPT-4 still has many known limitations that we are working to address, such as social bias, hallucinations, and contradictory statements,” explained developer OpenAI.

OpenAI in dispute with authors over ChatGPT

GPT-4 does not yet solve another problem with large language models either: The application was trained on the basis of thousands upon thousands of existing texts and images from the Internet. Many of these have been copyrighted by their creators. Therefore, several lawsuits are already pending against OpenAI, including one from the Getty picture agency.

The photo database Shutterstock is now breaking new ground. The provider has included AI functions in its image database, customers can have suitable images created by entering a few keywords. However, the artists and photographers on which the AI ​​systems were trained will be compensated. “Our easy-to-use generative platform will transform the way people tell their stories,” Shutterstock CEO Paul Hennessy told Handelsblatt.

Prequel boss Khabirov sees it similarly: “In the future we will work even more closely with photographers and other creative people. We say very clearly: We don’t want to steal your photos!” On the contrary: In the future, photographers could enter into partnerships with prequels and thus make the templates of a photo shoot once done available to other creative people for AI editing. “Photographers can earn a lot more from a shoot this way,” Khabirov believes.

Offer for Plus subscribers

GPT-4 is said to be available for OpenAIs Plus subscribers. Developers can sign up today to build applications with it. The new functions were presented to them as part of a live stream on Wednesday.

OpenAI released GPT-3 in 2020. Together with the successor version 3.5, the program was used to develop the chatbot ChatGPT and the image generation software Dall-E. Both products attracted a great deal of public attention when they were prominently presented last autumn and caused a hype about the possible fields of application of AI.

Also read about ChatGPT, OpenAI and artificial intelligence:

Since then, there has been speculation as to whether the next model could be even more powerful and possibly take on additional tasks.

According to OpenAI, the major bank Morgan Stanley is already using GPT-4 to manage data, while the payment service provider Stripe is testing whether the program can help fight fraud. The language learning app Duolingo is also one of the customers. At the SXSW in Austin, the Salesforce subsidiary Slack presented the integration of ChatGPT into their messaging applications at a large booth downtown.

A prominent partner is Microsoft. The Windows group had invested a billion dollars in fresh capital in OpenAI and made its cloud infrastructure available to Azure. On Thursday, Microsoft wants to provide information about which AI functions are to be used in its Office product family.

Google, among others, has come under pressure. Search engine giant-backed start-up Anthropic, founded by ex-OpenAI executives, announced on Tuesday the release of its chatbot Claude for business users.

With material from Bloomberg.

More: Creativity from the machine – This man will change your life

First publication: 03/14/2023, 11:06 p.m.