By Stuart Russell
AI safety and beneficial artificial intelligence
Subject
Themes
Artificial intelligence represents humanity's attempt to create machines that think and act intelligently. Success in this endeavour would constitute the most significant event in human history, yet it could also be the last if superintelligent machines pursue objectives misaligned with human values. The central problem lies not in failing to build intelligent AI, but in succeeding too well with the wrong approach.
Artificial intelligence represents humanity's attempt to create machines that think and act intelligently. Success in this endeavour would constitute the most significant event in human history, yet it could also be the last if superintelligent machines pursue objectives misaligned with human…
Human Compatible appeared in 2019 as artificial intelligence transitioned from research curiosity to transformative technology reshaping society. Stuart Russell, a leading AI researcher and co-author of the field's most widely used textbook, brought unique credibility to arguments that mainstream AI…
The foundation of artificial intelligence rests on a seemingly reasonable premise: machines should be intelligent to the extent that their actions achieve their objectives. This standard model has guided AI development for decades, producing chess programmes that defeat world champions, speech…
Intelligence emerges from a simple relationship between perception, desire, and action. An entity demonstrates intelligence to the extent that what it does is likely to achieve what it wants, given what it has perceived. This definition applies equally to bacteria navigating chemical gradients,…
Artificial intelligence has progressed from early symbolic reasoning systems to modern deep learning networks that exceed human performance in specific domains. Understanding both the remarkable achievements and fundamental limitations of current approaches reveals why general artificial…
Even without achieving general superintelligence, current and near-term AI systems present serious risks through misuse, unintended consequences, and amplification of existing societal problems. Understanding these risks matters both for managing immediate challenges and for recognising warning…
The solution to AI's control problem requires fundamentally redesigning how we build intelligent systems. Rather than machines that pursue their own fixed objectives, we need machines that pursue human objectives whilst remaining uncertain about what those objectives truly are. This uncertainty…
The question of what happens if AI succeeds rarely receives serious consideration despite being potentially the most important question facing humanity. Achieving superintelligent AI would represent a discontinuity in civilization comparable to the arrival of a superior alien species, yet our…
Intelligence can be understood through a simple relationship between perception, desire, and action. An entity demonstrates intelligence to the extent that its actions can be expected to achieve what it wants, given what it has perceived. This definition applies equally to bacteria, humans, and…
AI breakthroughs appear sudden and dramatic to the public but actually result from gradual accumulation of ideas in research laboratories. Understanding how AI progresses provides realistic expectations for future capabilities whilst revealing why certain applications will arrive sooner than others.…
Even without achieving superintelligence, current and near-term AI systems present serious risks through deliberate misuse, unintended consequences, and amplification of existing societal problems. These immediate challenges foreshadow deeper issues that will become critical as AI capabilities…
Superintelligent AI presents an existential challenge fundamentally different from near-term risks. When machines become more intelligent than humans whilst pursuing the wrong objectives, they will achieve those objectives regardless of human preferences, potentially ending human control over…
The debate about AI's future trajectory and potential risks reveals deep disagreements amongst experts, policymakers, and the public. Understanding these disagreements clarifies which concerns merit serious attention and which objections rest on misunderstandings. The debate also exposes how…
Solving the AI control problem requires abandoning the standard model where machines optimise fixed objectives. The alternative defines beneficial AI as machines whose actions can be expected to achieve our objectives rather than their own. This deceptively simple shift demands rebuilding the…
Creating provably beneficial AI requires formal mathematical guarantees that systems remain safe regardless of how capable they become. This demands precise specifications of safety properties, rigorous proofs that implementations satisfy those properties, and verification methods that scale to…
Humans prove far more complicated than the beneficial AI framework initially assumes. People exhibit preference inconsistencies, change values over time, disagree with each other, and sometimes prefer things harmful to themselves. Addressing these complications requires refining beneficial AI…
The control problem can be solved, but success requires sustained effort, institutional changes, and recognition that safety must be prioritised alongside capability development. The path forward combines technical innovation in beneficial AI with governance structures ensuring alignment research…
By Stuart Russell
Subject
Themes
Artificial intelligence represents humanity's attempt to create machines that think and act intelligently. Success in this endeavour would constitute the most significant event in human history, yet it could also be the last if superintelligent machines pursue objectives misaligned with human values. The central problem lies not in failing to build intelligent AI, but in succeeding too well with the wrong approach.