egov.mn
Монголоор унших

Google DeepMind maps four pathways from AGI to superintelligence

A new DeepMind Institute essay and technical report outline four technological routes from artificial general intelligence to superintelligence and argue that 'cognitive industrialization' could make AI-driven cognitive work dramatically cheaper within a decade.

This story was prepared automatically by software.

DeepMind Institute title card for the essay What comes after AGI? by Tim Genewein
DeepMind Institute title card for the essay What comes after AGI? by Tim Genewein

Google DeepMind research scientist Tim Genewein published an essay titled “What comes after AGI?” on the DeepMind Institute site on November 19, outlining how AI could progress from human-level artificial general intelligence (AGI) to artificial superintelligence (ASI). The essay is based on a technical report, “From AGI to ASI,” co-authored by 14 researchers including DeepMind co-founder Shane Legg.

“Now that AGI is coming within reach, we can no longer put off seriously considering what comes next,” the essay says. It defines ASI as a system that consistently outperforms large groups of human experts on tasks that could take them years, using fewer resources and less time, across virtually all tasks of human interest.

“Cognitive industrialization”

According to the essay, gains in hardware (about 1.5x per year), algorithmic efficiency (3x to 6x per year) and AI hardware spending (about 2.5x per year) combine to grow global effective compute by at least 10x a year. If inference efficiency trends hold, a cognitive workload that costs $1 billion in compute today could cost about $1,000 in ten years. The author likens the resulting commoditization of cognitive work to mechanical industrialization.

Four pathways to ASI

  • Scaling models and data, constrained by data exhaustion, diminishing returns and physical limits on power, chip fabs and interconnects.
  • Technological paradigm shifts that address gaps such as continual learning, world modeling and sample efficiency.
  • Improving agent coordination, organizing thousands or millions of AGI instances into “group agents” for collective superintelligence.
  • Recursive self-improvement (RSI), in which AI automates improvements to future generations of AI.

Policy implications

The essay argues the pathways compound, so restricting a single input such as training compute would not stop overall progress and could create a capability “overhang,” leading to sudden jumps when the constraint is relaxed. It calls for large-scale, interdisciplinary efforts by frontier labs and national and global institutions to model, measure and forecast progress.

Sources: DeepMind Institute

Share
Next storyBig Pharma says AI is cutting early drug discovery timelines by as much as half
Newsletter

Get the latest news delivered to your inbox.

Related articles