Paper Summary: AGI Preschool: A Framework for Evaluating Early-Stage Human-like AGIs

This is the paper review of the following paper.

Goertzel, Ben, and Vladimir Bugaj. “AGI preschool: a framework for evaluating early-stage human-like AGIs.” In Proceedings of the 2nd Conference on Artificiel General Intelligence (2009). Atlantis Press, 2009.

How we can evaluate the artificial general intelligence(AGI) is an open challenge in the community as suggested in this paper. We can start with the imitation test sometimes called Turing test. However, it is too difficult. We are struggling to find the basic principle of learning but the Turing test requires an adult level of knowledge. We can use all the knowledge in the world to trick the bot. Also because the agent in the Turing test is imitating the humans, there are many tricks that can fool humans into believing. For example, Eliza imitates an psychotherapist who keeps asking a question, and the recent winner from Loebner Prize mimics a russian teenager who learns English to fool the judges.

Other tests such as College Student test is adopted to block these loopholes. However, one major limitation of this approach is that they don’t deal about how these agents acquire these skills.

Ben Goertzel and Stephan Vladimir Bugaj suggests an alternative framework for AGI testing resembling preschool. It has similar curriculum for human including linguistic, social, logical-mathematical, nonverval communication, spatial-visual, object manipulation, social skills.

It is an advanced framework compared to other scenarios that it includes how they can acquire those skills. But in my opinion, it is still too challenging. We need to focus more earlier period, especially from inception to 24 months.

To solve the challenge of social learning they propose to use VR technology where a human participant controls avatar. And they suggest to use the physics engine to simulate realistic object manipulation.

After they pass the AGI preschool, the roadmap to more capable AGI is apparent. That they can build AGI grade school, secondary school, and so on.

My main criticism is that they are taking verbal and non-verbal communication as given. For example, to test AGI they suggest following tests.

  • Write a set of instructions; speak on a subject
  • analyse how a machine works
  • review a musical work
  • create a mime to explain something
  • coach or counsel another

Actually, language acquisition is the main challenge that needs to be tackled even before AGI preschool.

Below is my mindmap for the related papers to artificial general intelligence.