How far away are we from Artificial General Intelligence?

The Long Road to Artificial General Intelligence

Despite great shows of capabilities with Large Language Models (LLMs), I claim that we are still on a long way to Artificial General Intelligence (AGI); maybe not in the sense of time, but we have to figure the following three things out:

  • Artificial Neural Networks (ANNs) are of a black-box nature
  • ANNs have not yet shown great transferability of competences
  • ANNs tend to overestimate themselves
  • Introspection is limited

Here is a link to the poster.

Neuro-Symbolic

Apparently, LLMs lack higher logical thinking. Humans can very accurately calculate -- if they concentrate enough -- difficult computations by applying equivalent transformations on equations. LLMs cannot, as they lack the relevant concepts or fail to apply them at the right time. Neuro-Symbolic AI aims to let ANNs access Symbolic AI (which can perform logical deduction and calculations accurately). But this Symbolic AI is programmed by humans! So, the LLM should be able to add new concepts - but on what basis?

Problem Decomposition

Other attempts said: we could have the ANN learn the concepts and apply them. This was e.g. tested for classifiers. Somebody has yet to show how we could do this for mathematical problem-solving.

The results were promising - the trained ANNs were hierarchically structured, where each neuron corresponds to a concept or sub-concept; meaning, every neuron activation corresponds to some human-interpretable representation.

Emergence

Heidegger pronounced the difference between intentionality and correct use as such: We can open a door using a door handle many times to leave a room. It is only when the door may not be opened that we notice its existence and our intent. We want to leave a room.

Currently, this is one of the main differences I see between mankind and ai-kind. We can be aware, even if we are not in most situations. Still, emergence happened once, why not twice?

Acknowledgements

Special thanks to Professor Peter Dittrich for guiding me through confused phases! He immensely helped me to structure my thoughts and to find the right words to express them.

Special thanks to Johannes; for your feedback on the first draft of this poster and scouting for interesting articles. You laid the foundation for this poster!

Thanks to Maurice, Max and Samuel for your feedback!

Finally, many thanks to Professor Breuer for the opportunity to participate in the AI Summer School 2023!