Artificial General Intelligence

Artificial General Intelligence is effectively the modern term for what we used to call “AI” before machine learning and the marketeers took over. It is a form of AI which is completely general purpose - it can be taught or learn how to deal with an ever increasing range of problems of increasing complexity - and ultimately perform as well, or better, than a human at almost any task (although imagination and emotion may represent a different axis - or maybe not). It should be noted that there is no corollary between AGI and “virtual human” - the AGI could be very computer like (the the HUHN lens), or very human like, or somewhere in between.

A couple of useful ideas to bear in mind when thinking about AGI are below, along with some useful links. The message at the moment (2019) though is that we are nowhere near to creating AGI.


Ben Goertzel has identified 4 research approaches towards AGI:

  • Symbolic – such as Cyc, SOAR and ACT-R discussed in Chapter 6.
  • Emergentist – a sub-symbolic approach akin to the low-level neuron/synapse model of the brain from which other properties and capabilities emerge. The IBM Blue Brain project (Markham, 2006) is an example of the more computational neuroscience orientated approach.
  • Hybrid – a hybrid of the above two approaches using elements of each in combination, e.g. Goertzel’s CogPrime described in Chapter 6, although lacking the elegance and simplicity of a single approach.
  • Universalist – a more theoretical approach based on creating the ‘ideal’ program given enough computing power to iteratively evolve it. Hutter’s AIXI system (Hutter, 2004) is an example (although it bears a frightening resemblance to Douglas Adam’s ‘nice hot cup of tea’ approach to creating the infinite improbability drive).

Ben's CogPrime Overview gives a useful "where are we now" assessment of AGI development based around a similar component model to our own.


Here are a set of tests for an AGI which have been proposed for various people (mainly Muehlhauser). We are still a long way from even the first step.

  • the Turing Test as implemented in a ‘Gold’ Loebner Prize type competition (see Chapter 5, involving both natural language conversation and audio-visual presentation and understanding);
  • the coffee test – just going into a typical (virtual?) house and making a cup of coffee;
  • the On-Line Student Test - an AGI taking an on-line/MOOC style course as though an ordinary student;
  • the robot college student test – enrolling and taking classes just like any other student; and
  • the employment test – being able to perform an ‘economically important’ job;
  • the Artificial Scientist Test - being able to perform a scientific  research job;
  • the Nobel Prize test - winning the Nobel prize!