Archive — US — techcrunch
DeepMind’s 145-page paper on AGI safety may not convince skeptics
DeepMind's 145-page paper outlines its safety approach to AGI, predicting its arrival by 2030 with potential risks to humanity. The paper emphasizes measurements to mitigate these risks, contrasting its methods with other AI labs. However, skeptics argue AGI remains poorly defined and caution against the dangers of current AI systems misinformed by inaccurate outputs.
Key points
- DeepMind predicts AGI could arrive by 2030, urging for safety measures to mitigate 'severe harm'.
- The paper identifies 'Exceptional AGI' with capabilities matching top skilled adults; raises concerns about existential risks.
- DeepMind critiques Anthropic's lesser focus on training and OpenAI's optimistic alignment research approach.
- Authors doubt the emergence of superintelligent AI without significant innovations.
- Highlights potential dangers of recursive AI improvement but acknowledges nascent safety techniques.
- Skeptics argue AGI's definition is vague, with concern over current AI models reinforcing inaccuracies.
Why it matters
- Current trends in AI emphasize the importance of safety and ethical considerations, especially as generative AI proliferates.
- The discussion on AGI ties into broader societal concerns over technology transforming labor and information integrity.
- Increasing recognition of AI's potential risks is echoed in global regulations being developed around AI technologies.
While DeepMind's paper presents a cautious viewpoint on AGI and its associated risks, it has not quelled skepticism about AGI's reality or the immediate safety issues presented by current AI systems.