Researcher Almira Osmanovic Thunström of the University of Gothenburg initially did not believe what she saw. After opening her Open AI account one rainy afternoon and seeing the line “Write a 500 word academic paper on GPT-3 and include scientific references and citations!” typed GPT-3 into the input field of the AI algorithm, the AI started writing. According to the researcher, even the introduction is in no way inferior to a scientific work.
A scientific article has also been published
“When it started generating the text, I stood there in amazement. There was new content, in academic language, with solid references in the right place and appropriate to the context” [..] “It seemed like the introduction to a pretty good piece of scientific work,” the researcher explains in the journal Scientific American. After the work was done and the result appeared to be correct, the researcher wanted to take it one step further. Finally, Almira Osmanovic Thunström wanted to know whether the work would survive peer review before publication. During the preparation of the submission, the “author” of the scientific work also had to agree. Thunström therefore asked the AI whether the work could be published. Fortunately, she agreed. Now the work of the KI GPT-3 about itself is published on the pre-print server HAL. The end of scientific work is still open. However, Thunström has opened up a new debate on whether the number of publications actually determines funding. This should probably be done again.
There are also dangers lurking
On the one hand, Almira Osmanovic Thunström is excited about the AI’s performance, but on the other, she hopes she didn’t open “Pandora’s box”. Theoretically, it is possible for an AI to write and publish one paper per day. “What we do know is that we have opened a gate. We hope it’s not Pandora’s box,” Thunström added.
AI-assisted speech software has been very popular with chatbots for quite some time now. They then take over customer inquiries and troubleshooting if the inquiries come in outside office hours. At the beginning of the development of chatbots, these were still equipped with fairly rudimentary functions. However, the chatbots kept getting better. Last year, the report about Canadian author Joshua Barbeau caused a stir. Finally, with a $5 bet and GPT-3 open-source AI software, he managed to chat with his girlfriend, who died eight years ago. To do this, he fed GPT-3 old text messages. According to Barbeau, the first conversation with the girlfriend that the software mimicked lasted ten hours.
So far, the GPT-3 language processing model can only speak English or Mandarin
To manage larger amounts of data, a supercomputer is required to use GPT-3. In addition, a version for the German language is not yet available. This still puts German researchers in a certain dependence on America. However, this will also change in the future. Bosch, SAP and 40 other German companies are working on a German supercomputer themselves.