When discussing a topic, it is always important to define the vocabulary we’re using so that a common shared definition is used.
When it comes to AGI, this definition is not yet 100% clear.
Here I share my own definition. I will try to explore alternative properties as this article evolves.
Properties #
General AGI #
- Is able to learn (behavior, facts) from input
- Is able to retrieve (behavior, facts) from memory
- Is able to regenerate (behavior, facts) as output
- Is able to generate predictions
- Can interact with an environment (either virtual or real)
- Does not need to know everything
Humanoid AGI #
- Is able (after some amount of learning) to accomplish some humanoid behavior
- We expect an AGI to be able to reproduce human behaviors,
- Can interact with human beings at an appropriate speed
- Interacting too slowly would make it appear dumb/non-human
- Can interact with the physical environment
Wikipedia’s definition #
Artificial general intelligence (AGI) is the hypothetical ability of an intelligent agent to understand or learn any intellectual task that a human being can.
Arguments for specific requirements #
In this section we explore why certain requirements are necessary as part of the definition of an AGI.
Arguments against specific requirements #
In this section we explore why certain requirements are not necessary as part of the definition of an AGI.
Language based #
An agent does not necessarily need to use an advanced language like writing to be able to accomplish human-like behaviors.
If we claim that communication implies language, then language is a requirement. It however does not mean that language needs to be complex. The simple fact that there is a medium that allows two agents to communicate implicitly means that there is an underlying language. The point here is that spoken and wrriten languages are not a necessity but likely something that emerges from the evolution of simpler languages.