OpenAI and DeepMind researchers warn on self-improving AI risks

An article published by qz.com, titled 'OpenAI and DeepMind researchers warn on self-improving AI risks,' reports warnings from researchers at two frontier AI laboratories about the risks of self-improving artificial intelligence. The piece cites Juan Felipe Ceron Uribe, an AI alignment research engineer at OpenAI, and Neel Nanda, a research scientist at DeepMind. Both were quoted from videos, according to the article. The title frames the subject as self-improving AI risks.

In one of the videos, Ceron Uribe said frontier labs are 'racing each other, kind of blindfolded.' In a separate video, Nanda said he believed there was at least a 10% chance AI could lead to human extinction. He described that probability as 'ridiculously high,' according to Reuters. The article attributes the 10% figure and the accompanying description to that quotation as reported by Reuters. No further detail on the videos, including when they were recorded or where they were published, appears in the passage quoted here.

The article presents the two researchers as employees of OpenAI and DeepMind respectively, and characterises their remarks as warnings about risks from self-improving AI. Reuters is identified as the source for Nanda's statement, including the 10% figure and the phrase 'ridiculously high.' In the material quoted, the article does not report a change in policy at either company, a new safety publication, a departure by either researcher, or any regulatory response. It also does not state what measures, if any, the two researchers proposed.