A discussion has erupted in the technology industry following a statement by Evan Hubinger, an artificial intelligence safety specialist at Anthropic. He believes there is over a 10% chance that humanity could be destroyed by the development of AI within the next decade.
Evan Hubinger, who holds the position of Head of Alignment Science at Anthropic, expressed his concern about the future of AI in his post. In his opinion, the greatest danger may arise when AI significantly surpasses human intelligence and begins to improve itself autonomously.
This process is called recursive self-improvement. If AI can improve itself without human assistance and create more powerful versions on its own, its capabilities could grow very rapidly. Hubinger noted that although Anthropic is working on AI safety, a reliable solution for complete control over a future superintelligent system is still lacking.
His statement came amid the departure of Anthropic researcher Jacob Coxon, who also voiced concerns about the rapid race among AI companies. Coxon pointed out that companies are actively developing more powerful and self-improving systems, but safety issues remain unresolved.
It is important to note that Hubinger does not claim that existing models, such as ChatGPT, will destroy people in the coming decade. His concern is directed at the potential of future superintelligent systems. The main question is whether safety is keeping pace with the pace of AI development striving to become smarter.
Hubinger's 10% estimate reflects this growing anxiety, which is now being discussed not only externally but also within the AI companies themselves.