Are We Any Nearer to AGI?

 

What should we expect from AGI?

We should expect something that starts out with:

·        An understanding of complex text (legislation, specifications, instructions) with an ability to actively link multiple documents.

·        Ability to read and make sense of photographs and diagrams

·        Understanding of reality and moving pictures (videos)

·        Understanding of simple mathematics, physics and chemistry, with an understanding of how to increase the complexity to match the problem.

In other words, the equivalent of a first year university student, with the addition of an ability to link many things – an ability that far exceeds a human’s severely limited ability (about four concepts). We have the problem that the person or persons will have insufficient “bandwidth” to describe the problem in the first place, and will need to be led

So are we getting any closer to AGI?

Unfortunately, no.

Generative AI

Generative AI has largely taken over AI, and this is antithetical to the first goal, understanding text. Generative AI can search a very large dataset (all the text on the internet), and may find an article written by a human that links the words used in the prompt (it may find many articles – if so, it takes the most popular one).  When it works well, it seems to work effortlessly (it is not really AI, instead it is a lookup service on a huge dataset. The problem comes if nothing that matches the prompt can be found (the problem is new or emerging), or you wish to combine pieces of information into a larger synthesis. Generative AI does not understand the meaning of a single word, so synthesising a more complex entity is impossible for it.

Neurosymbolics

Neurosymbolis is often touted as a stepping stone to AGI. This seems unlikely, as it throws away information about the text. English has a logic of its own, where a group of words has a special meaning – “a house of cards”, “an imaginary flat surface of infinite extent” (a “plane”, where “imaginary” covers “infinite extent”), “electronic funds transfer instruction”.

Neurosymbolics throws away the connections between words, instead replacing them with logical operators that know nothing about connections among English words. Needless to say, throwing information away is not the best way to proceed to understand something. It also assumes that the words and symbols completely express the problem, when many problems are only partially known, and much hypothetical reasoning will be required to discover the full outline of the problem (or the solution is given in several hypothetical solutions).

Reading of Photographs and Diagrams

There will be many AGI problems where the problem cannot be described in text alone, or failure to read a photo or diagram leads to an inconsistency. An example is a photo of a loaded container ship. The accompanying text talks about a level of stacking (of containers on top of each other of seven or nine. The photo shows a level of stacking of eleven (inconsistencies will be rife in descriptions that use different media forms).

The problem of inconsistencies is very likely to occur with large and complex projects – typically where tens or hundreds of billions of dollars can be wasted due to unfamiliarity (the US F-35 project as an example).