DeepMind’s 145-page paper on AGI safety may not convince skeptics

Google DeepMind published one on Wednesday exhaustive paper About the security approach of AGI, roughly defined as AI who can accomplish any task that a person can do.

Agi is a bit of a controversial subject in the AI ​​field, with no -sayers Sugguing that it is little more than a pipe dream. Others, including large AI Laboratories such as Anthropic, warn that it is around the corner and can lead to catastrophic damage if no steps are taken to implement suitable guarantees.

DeepMind’s 145-page document, which was co-author of DeepMind-MEDE founder Shane Legg, predicts that Agi could arrive by 2030 and that it can lead to what the authors call ‘serious damage’. The paper does not define this in concrete terms, but gives the alarming example of ‘existential risks’ that ‘permanently destroy humanity’.

‘[We anticipate] The development of an exceptional AGI before the end of the current decade, “the authors wrote.” An exceptional AGI is a system with a capacity that corresponds to a wide range of non-physical tasks, including metacognitive tasks such as learning new skills. “

From the bat, the newspaper DeepMind’s treatment of AGI risk reduction contrasts with that of Anthropic and OpenAi. Anthropic, says it, puts less emphasis on ‘robust training, monitoring and security’, while OpenAi is overly bullish about ‘automating’ a form of AI safety research known as Alignment Research.

The newspaper also raises doubts about the viability of super intelligent AI AI that can make better jobs than any person. (OpenAi recently claimed that the goal from Agi to Superintelligence is ranging.) In the absence of ‘important architectural innovation’, the DeepMind authors are not convinced that super intelligent Systems will soon arise – like never.

However, the paper finds it plausible that the current paradigms make “recursive AI improvement” possible: a positive feedback loop where AI conducts its own AI research to create more advanced AI systems. And this can be incredibly dangerous, the authors claim.

At a high level, the article proposes and argues for the development of techniques to block the access of poor actors to hypothetical AGI, to improve the understanding of the actions of AI systems and the environments in which AI can act. It acknowledges that many of the techniques are budding and have ‘open research problems’, but warns against ignoring the safety challenges on the horizon.

“The transforming nature of AGI has the potential for both incredible benefits and serious damage,” the authors write. “As a result, it is crucial to build AGI in a responsible manner for Frontier AI developers to proactively intend to limit serious damage.”

However, some experts do not agree with the field of the newspaper.

Heidy Khlaaf, chief AI scientist at the non-profit AI Now Institute, told WAN that she thinks the concept of AGI is too badly defined to “be rigorous scientifically evaluated”. Another AI researcher, Matthew Guzdial, a university teacher at the University of Alberta, said he does not believe that recursive AI improvement is currently realistic.

‘[Recursive improvement] Is the basis for the arguments of the intelligence singularity, “Guzdial told WAN,” but we have never seen any evidence that it works. ‘

Sandra Wachter, a researcher who studies technology and regulation at Oxford, claims that a more realistic care is that AI strengthens itself with ‘incorrect output’.

“With the proliferation of generative AI outputs on the internet and the gradual replacement of authentic data, models now learn from their own outputs that are full of Mistruths of Hallucinations,” she told WAN. “At the moment, chatbots are mainly used for search and truth friendly purposes. That means that we are constantly running the risk of getting recognition and to believe them because they are presented in very convincing ways.”

No matter how extensive it is, the DeepMind article probably seems to regulate the debates about how realistic AGI is – and the areas of AI safety in the most urgent need for attention.