# AI-GAs: AI-generating algorithms, an alternate paradigm for producing general artificial intelligence

**Authors:** Jeff Clune  
**arXiv:** [1905.10985](https://arxiv.org/abs/1905.10985)  
**Date:** 2019-05-27  
**Summary page:** https://www.summarizepaper.com/en/paper/1905.10985v2/

## Key points

- Jeff Clune's paper discusses the creation of general artificial intelligence (AI)
- The current dominant approach in machine learning is the "manual AI approach"
- Clune proposes an alternate path involving creating an AI-generating algorithm (AI-GA) that automatically learns how to produce general AI
- Three pillars are essential for this approach: meta-learning architectures, meta-learning the learning algorithms themselves, and generating effective learning environments
- Pursuing AI-GAs should be considered a new grand challenge of computer science research because it may be the fastest path to general AI and it is scientifically interesting to understand how a simple algorithm can produce general intelligence similar to how Darwinian evolution produced human intelligence on Earth.

## Summary

In his paper titled "AI-GAs: AI-generating algorithms, an alternate paradigm for producing general artificial intelligence," Jeff Clune discusses the creation of general artificial intelligence (AI), which refers to AI that is as smart or even smarter than humans. The current dominant approach in the machine learning community is the "manual AI approach," where researchers attempt to discover each piece required for intelligence with the assumption that some future group will combine them into a complex thinking machine. However, Clune proposes another path that may be more successful at producing general AI based on the trend in machine learning that hand-designed solutions are eventually replaced by more effective learned solutions. This new approach involves creating an AI-generating algorithm (AI-GA) that automatically learns how to produce general AI. Three pillars are essential for this approach: meta-learning architectures, meta-learning the learning algorithms themselves, and generating effective learning environments. Clune argues that either approach could produce general AI first and both are scientifically worthwhile irrespective of which is faster. However, because both approaches are promising and yet the ML community is currently committed to the manual approach, he suggests increasing research investment in the AI-GA approach. To encourage such research, he describes promising work in each of the three pillars and discusses safety and ethical considerations specific to AI-GAs. Clune believes pursuing AI-GAs should be considered a new grand challenge of computer science research because it may be the fastest path to general AI and it is scientifically interesting to understand how a simple algorithm can produce general intelligence similar to how Darwinian evolution produced human intelligence on Earth. Overall, Clune's paper provides an exciting alternative path towards achieving general artificial intelligence through automated learning algorithms rather than relying solely on manual design.