Eliezer Yudkowsky on the future of humanity-destroying AI
In this video podcast, Yudkowsky explains and clarifies his pessimistic outlook on how AGI (Artificial General Intelligence) will unravel. Yudkowsky, an AI researcher and author of “Harry Potter and the Methods of Rationality,” is a towering figure in the AI community. According to OpenAI’s Sam Altman, he “has done more to accelerate AGI than anyone else.” Yudkowsky is also famous for saying that unrestricted development of AGI will lead to the end of humanity. In the podcast, he lays out his position in a very newbie-friendly way with concrete examples of why and how AI will destroy the world, which is a welcome change from the usual rationalist heavy prose.
The Waluigi Effect
The main reading for this week in AI circles is a post on a rationalist blog LessWrong (cue heavy prose). The author tries to explain why it’s so easy to make ChatGPT and Bing not just contradict its rules but do the exact opposite of what their creators intended them to do (as we covered in our Bing meltdowns compilation). The pessimistic conclusion seems to be that RLHF models such as GPT-3 are an irreparably inadequate solution to the AI alignment problem (something Yudkowsky pointed out a few times before).
How ChatGPT makes people more productive (a research paper!)
A Twitter thread by one of the authors of a new experiment involving ChatGPT describes how it affects low- and high-performers. The short-term outlook (until AI kills everyone) seems mildly optimistic: ChatGPT “reduces productivity inequality; low-ability workers gain the most from access to ChatGPT.” Interestingly, “a majority of participants paste the task instructions into ChatGPT and directly submit the output. There is limited editing, reprompting, or other human-machine iteration.” People are happy doing what the AI tells them to do.
AI art continues to be controversial
An article from Hyperallergic describes a kerfuffle at Hague’s Mauritshuis museum, which decided to substitute Vermeer’s painting with AI analogues while the originals were on loan to another museum. This prompted a flurry of accusations from European artists and activists, including one from the “European Guild for Artificial Intelligence Regulation.” The critics point out that AI models are built on the works of artists who are not compensated for being part of the training dataset. This is still an unsolved issue as Stability, the company behind Stable Diffusion, goes to court on charges of copyright infringement.