When will AI destroy humanity? This expert has a guess
What's the story
Leading artificial intelligence (AI) expert Daniel Kokotajlo has revised his prediction for when AI systems might become capable of autonomous coding, and accelerate their own development toward superintelligence. The former OpenAI employee had earlier predicted an "intelligence explosion" by 2027, where AI agents would fully automate coding and research. This would lead to a superintelligent system that could potentially threaten humanity.
Timeline adjustment
Revised predictions push back timeline for AI superintelligence
Kokotajlo and his team have now pushed back their prediction for when AI might achieve "fully autonomous coding" to the early 2030s, from the initial estimate of 2027. The updated forecast sets 2034 as the new target year for superintelligence, without any speculation on when this could lead to humanity's potential destruction. Kokotajlo acknowledged in a post on X that things appear to be progressing slower than initially anticipated in the AI 2027 scenario.
Challenges ahead
AI's unpredictable performance and societal inertia
Malcolm Murray, an AI risk management expert, noted that many are pushing their timelines further out as they realize the jagged performance of AI. He emphasized that for a scenario like AI 2027 to occur, AI would need a lot more practical skills useful in real-world complexities. This highlights the enormous inertia in the real world that could delay complete societal change.
Terminology shift
The evolving definition of AGI
Henry Papadatos, Executive Director of French AI nonprofit SaferAI, said the term AGI made sense when AI systems were very narrow. But now that we have systems that are quite general already, the term does not mean as much. This shows how the understanding and definition of artificial general intelligence (AGI) is changing with advancements in AI technology.
Skepticism
AI safety report authors question AGI timelines
The authors of the International AI Safety Report, which includes Murray, have also questioned the timelines for transformative AI. They note that many people are pushing their predictions further out as they observe the unpredictable performance of AI systems.