Google DeepMind released on Wednesday Detailed paper Regarding its security approach to AGI, it is roughly defined as an AI that can accomplish any task that humans can accomplish.
AGI has a somewhat controversial topic in the AI field Opponents Implications it is more than just a daydream. Others, including AI major laboratories, etc.warning it around the corner that if steps are not taken to implement appropriate safeguards, it can cause disastrous harm.
DeepMind’s 145-page file, co-authored by DeepMind co-founder Shane Legg, predicts that AGI can arrive by 2030 and could lead to what the author calls “severe harm.” Instead of defining this specifically, the paper gives examples of the dangerous “existent risks” of “permanent destruction of humanity.”
“[We anticipate] “The development of a special AGI before the end of the current decade is an outstanding AGI with a system that matches the abilities of at least the 99th skilled adult, which includes a wide range of non-physical tasks, including metacognitive tasks such as learning new skills,” the authors wrote.
In the case of bats, paper compares DeepMind with the treatment of reducing AGI risk reduction by Anthropic and Openai. It says that anthropomorphism less emphasizes “strong training, monitoring and security”, while Openai is too optimistic about the form of “automated” AI security research, called A Alignment Research.
The paper also raises doubts about the viability of super-intelligent AI – AI can do better than anyone else. (Openai Recent claims Its goal is to change from AGI to super smart. ) Without “important architectural innovations”, DeepMind authors do not believe that superintelligent systems will soon appear (if any).
However, this article does find that the current paradigm will enable “recursive AI improvements”: a positive feedback loop, within which AI conducts its own AI research to create more complex AI systems. This can be very dangerous, the author asserts.
At a very high level, this article proposes and advocates to develop technologies to prevent bad actors from accessing hypothetical AGI, improve understanding of AI system actions, and “harden” the environment that AI can take. It acknowledges that many technologies are new and there are “open research questions”, but warns not to ignore the possible security challenges.
“The transformative nature of AGI has incredible benefits and serious harm,” the author writes. “So, to build AGI responsibly, it is crucial for Frontier AI developers to proactively plan to mitigate serious harm.”
However, some experts disagree with the premise of the paper.
Heidy Khlaaf, the current chief AI scientist at the Institute for Nonprofit AI, told TechCrunch that she believes the concept of AGI is too defined to be “rigorous scientific evaluation.” Another AI researcher, Matthew Guzdial, an assistant professor at the University of Alberta, said he does not think the improvement of recursive AI is realistic at the moment.
“[Recursive improvement] “But we have never seen any valid evidence,” Guzdial told TechCrunch.
Sandra Wachter, a researcher who studies Oxford’s technology and regulations, believes that the more realistic focus is that AI strengthens itself with “inaccurate output.”
“As the proliferation of generative AI outputs on the internet and gradually replacing real data, models are now learning from their own outputs, which are full of misunderstandings or hallucinations,” she told TechCrunch. “At this point, chatbots are primarily used for search and truth-finding purposes. This means we are constantly at risk of being abandoned because they are presented in very convincing ways because they are presented in very convincing ways.”
Comprehensively, DeepMind’s paper seems unlikely to resolve the debate on the reality of AGI, and in the field of AI security that most urgently needs attention.
