Why adding ontologies to LLMs won't yield machine intelligence

1 points by veqq

veqq

Long interested in symbolic AI, neuro-symbolic instantly jumped out at me some time ago. Jobst Landgrebe disagrees. Why? To summarize:

He defines practical intelligence (in humans and animals) as ability to adapt to new (irregular) situations, suddenly, appropriately, without training
AI always has deterministic (no error) or implicit stochastic models (with an error, the sum of deviation), both requiring regularities/modeling regular relationships. Modern AI is mostly learning off of input and output sets (e.g. whether email is spam or not), scope-restricted and error-prone (and though necessary, reducing dimensionality causes many more errors)
So people resort to ontologies to reduce data requirements in training, for "certifiable" or "explainable" AI and general error reduction
chart of 3 ways to combine symbolic and neuro

Basically, he asserts a definition of intelligence and says it's not possible, axiomatically. Everything falls apart when asserting there are only 3 ways of working and further he says ontologies must be static - but I've been using evolving ontologies (with fuzzy logic) for a few years.

madhadron

I've been using evolving ontologies (with fuzzy logic) for a few years.

That sounds really cool. Do you have a suggestion of an entry point for the field?
- ocramz
  
  I would love to hear more too. Problem with ontologies has always been ingestion and reconciliation: how do you "graft" a new term to a preexisting graph, such that the graph doesn't explode combinatorially , and that new queries keep giving meaningful/minimal results ?