ControlAI Risks Page

Extinction Risks from AGI | ControlAI 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

 

 
 
 
 
 
 
 
 
 The Race to AGI: Humanity's Greatest Threat

 Extinction Risks from AGI

 The Current State of AI Development

 AI progress has exploded in the last 10 years, reaching near-human-level capabilities in writing, coding, art, math, and many more fields of human activity. This progress has been driven by deep learning: modern AIs are grown by feeding them massive amounts of data and letting them evolve in response, not built piece by piece by humans.

 Researchers and engineers don't need to understand AIs to create them – indeed, experts consistently fail to anticipate how quickly new skills will be unlocked, or how existing AIs work. Progress is bottlenecked only by resources (such as AI chips, electrical power, data) instead of scientific insights. As tech companies and frontier AI companies collaborate to unlock ever more resources, the path leads to increasingly intelligent yet opaque AIs.

 The Path to Superintelligence

 How far can this trend of smarter AI go? Looking at how humanity has historically increased its intelligence, three strategies emerge: tools, groups, and methods. AIs can leverage these same strategies; indeed each is already being used in current AI R&D.

 Although skeptics claim that various components of intelligence cannot be automated, no existing scientific theory of intelligence backs these opinions. Thus we should assume the trend towards smarter AIs will continue, eventually leading to Artificial General Intelligence (AGI), AIs able to do the same intellectual tasks as humans.

 The Emergence of AGI and Beyond

 Since AGI would be able to do anything a human can, it would be able to do AI research and improve its own intelligence. AGI companies and AI researchers are already pushing hard in this direction. Since software is cheaper and more efficient than brains, AGI would improve far faster than any human, eventually reaching artificial superintelligence (ASI) surpassing humanity's collective intelligence.

 As it continues to scale, ASI would unlock abilities to shape matter and energy that would look godlike compared to human engineering. Even without malicious intent, these godlike-AIs would by default wipe out humanity as collateral damage while pursuing their own goals, in the same way ants are just collateral damage for contractors building a house.

 The Challenge of AI Alignment

 Godlike-AIs lead to catastrophe because of the incredible difficulty of aligning AI's goals with those of humanity. Alignment is the harder version of problems humanity already struggles with: making companies and governments beneficial for what citizens care about.

 Solving alignment requires massive progress on questions like finding what we value, reconciling contradictions between values, and predicting consequences to avoid side-effects. This would require decades of research and trill

... (truncated, 5 KB total)