Tag
AI Futures researchers update their timelines forecasts, arguing METR's estimate that AI is driving roughly 1.7x–17x acceleration of AI progress puts the intelligence explosion on track for 2027, with Superhuman AI Researcher likely arriving in 2027 and a 35–65% chance of Top-Expert-Dominating AI before 2028.
Toby Ord 的文章探讨 AI 模型通过更多 token 扩展的效果是否与人类通过更多时间扩展的效果相当,分析了 METR 的对比数据,并讨论了这一问题对 AI 自我改进与智能爆炸前景的影响。
Geoffrey Hinton co-authors a paper warning that AI is now participating in the R&D of the next generation of AI, potentially triggering an intelligence explosion once a self-research and self-improvement loop forms — with AI progress from one year compressed to about five weeks. The paper is co-signed by top researchers from OpenAI, Anthropic, and Microsoft.
Geoffrey Hinton notes that while the concept of an intelligence explosion from recursive self-improvement is not new, many leading researchers now believe it may arrive imminently, and he points to a new paper making this case.
Miles Brundage reflects on his intellectual mistake of not adequately considering intelligence explosion scenarios, which influenced his optimism about AI alignment and skepticism towards the value of slowdowns.
This article assesses evidence that automating AI R&D could trigger an intelligence explosion, analyzing potential impacts and proposing policy responses to address extreme risks and benefits.
The article examines whether AI recursive self-improvement can overcome diminishing returns, arguing that current data suggests the self-improvement loop isn't strong enough to cause a runaway intelligence explosion without significant advancement.
Ramez Naam explores the current state and expectations of the intelligence explosion in AI, questioning why it hasn't occurred yet.
The article argues that data bottlenecks will not prevent a fast intelligence explosion in AI, as software progress can overcome data limitations through improved algorithms and sample-efficient learning.
Explores the subjective time compression an ASI would experience due to vastly faster processing speeds compared to human brains, suggesting one month of human time equals 820,000 years of ASI thought, raising alignment implications.