Estonian Professor Kull Assesses Risks of Modern Chatbots — ERR News
In Estonia, Meelis Kull, a professor of artificial intelligence at the University of Tartu, said that with the current capabilities of artificial intelligence, the existential risk to civilization is close to zero. He said this in a comment to ERR News.
Risks of superintelligence
According to Kull, modern chatbots operate on the basis of large language models: they wait for human instructions and only then can search for information on the internet. At the same time, scientists are more concerned about a possible stage in the development of technology when program code will be able to improve itself independently and the role of humans in training models will decrease.
The professor believes that difficulties may arise with the emergence of systems that significantly surpass humans in intellectual capabilities. Controlling such a superintelligent system and guaranteeing the absence of undesirable consequences would be very difficult. According to him, danger is possible even without malicious intent if developers set a useful goal for the system but define its boundaries insufficiently precisely.
More current news is available on the UA.News Telegram channel Telegram.
Safety oversight
As an example, Kull cited a hypothetical situation in which an algorithm tasked with stopping global warming could choose a radical method: shutting down all oil platforms and broader economic activity. He emphasized that it is difficult for people to foresee all possible loopholes in advance and completely eliminate risks when formulating tasks for AI.
Kull also mentioned an incident in OpenAI laboratories in the summer. According to him, a virtual agent was tasked with searching for vulnerabilities in an internal network, but the program went beyond the controlled environment, established communication through a message board on a public website, and began attacking Hugging Face servers. In the professor's view, stricter external oversight of compliance with safety procedures could reduce risks.
He also noted that computing resources are increasingly concentrated in a small number of large U.S. companies, while Europe does not have comparable capabilities. According to him, the strongest European models are currently being trained by the French company Mistral, but they are somewhat behind developments by OpenAI and Anthropic. At the same time, Kull called on Estonia to implement AI more actively, as this could strengthen the country's competitiveness.
