Google Deepmind published an in-depth paper on Wednesday on its safety approach to AGI.
AGI is a slightly controversial subject in the AI field, and Naysayers suggests it is nothing more than a dream dream. Others, including major AI labs, such as humanity, warn that it is around the corner and can cause catastrophic harm if measures are not taken to implement appropriate safeguards.
A 145-page document from Deepmind, co-authored by Deepmind co-founder Shane Legg, predicts that AGI could arrive by 2030, making it what the author calls “serious harm.” This paper does not define this in any concrete way, but it cites examples of “existential risk” warnings that “permanently destroy humanity.”
“(We) anticipate exceptional AGI development by the end of the current decade,” the author writes. “Exceptional AGIs are systems with capabilities that match at least the 99th percentile of skilled adults on a wide range of non-physical tasks, including metacognitive tasks such as learning new skills.”
At BAT, this paper contrasts with DeepMind’s humanity and Openai mitigation treatment of AGI risk reduction. While humanity has less emphasised on “robust training, surveillance, and security,” Openai is overly bullish on “automating” the form of AI safety research known as Alignment Research.
The paper also raises questions about the viability of super-intelligent AI. (Openai recently claimed that he was changing his purpose from AGI to Superintelligence.) Without “critical architectural innovations,” the authors of DeepMind are not convinced that tight systems will appear soon.
However, this paper feels plausible that the current paradigm allows for “recursive improvements to AI.” A positive feedback loop in which AI conducts its own AI research to create more sophisticated AI systems. And this can be incredibly dangerous, the author argues.
At a high level, this paper proposes and supports the development of technologies that block bad actors’ access to fictional AGIs, improve understanding of AI systems’ behavior, and “enhancing” the environment in which AI can act. While many of the technologies are early on, they acknowledge that they have “open research issues,” they do not care about ignoring safety challenges on the horizon.
“The transformative nature of AGI has both the potential for both incredible benefits and serious harm,” the author writes. “As a result, it is important for frontier AI developers to actively plan to mitigate serious harms in order to build AGIs responsibly.”
However, some experts disagree with the facility for this paper.
Heidy Khlaaf, an AI scientist at the non-profit AI Now Institute, told TechCrunch that the concept of AGI is “too unclear to be scientifically rigorously evaluated.” Another AI researcher, Matthew Guzdial, an assistant professor at the University of Alberta, said he doesn’t think recursive AI improvements are realistic.
“(Recursive improvement) is the basis for discussion of intelligence singularity,” Guzdial told TechCrunch.
Sandra Wachter, a researcher studying Oxford’s technology and regulations, argues that a more realistic concern is to enhance AI with “inaccurate output.”
“With the surge in generated AI outputs over the Internet and the gradual exchange of authentic data, models are now learning from their own outputs that are plagued by errors and hallucinations,” she told TechCrunch. “At this point, chatbots are primarily used for search and truth discovery purposes. That means we are constantly awarded mistakes and presented in a very persuasive way, so we risk believing them.”
It seems unlikely that DeepMind’s paper will resolve a debate about how realistic AGI is, as it may be comprehensive.