The Promise of Symbolic AI

Despite their impressive abilities, LLMs have significant limitations:

LLMs Lack True Understanding Of Meaning

LLMs are pattern matchers. They can generate text or sounds or videos that seem intelligent but don’t necessarily reflect a deep understanding of concepts or facts. LLMs don't have a true mental model of the world.

LLMs Can Be Wrong

LLMs sometimes generate factually incorrect, nonsensical, or made-up information. We know these as hallucinations, but in essence LLMs are always hallucinating - they just hallucinate the right things most of the time. This is because they are designed to optimize statistically likely word orderings for plausible-sounding text.  They are not designed to be necessarily correct. 

LLMs Can’t Actually Reason

While LLMs can display implicit common sense, they can fail at simple logical inferences or everyday reasoning that humans find trivial.  To the extent that they seem to reason, they are largely performing probabilistic mimicry of examples from their training data.  

We Don’t Know What They Know

The black box nature of connectionist networks - and LLMs by extension - makes it difficult to understand how they work.  An LLM’s outputs over a set of inputs can’t be reasoned about through examination of its architectural weights and connections. One must actually run the inputs through the system.  This hinders predictability and, ultimately, trust.  It also makes LLMs nearly impossible to debug. 

This is often called the interpretability problem.  

Bias Amplification

LLMs reflect and repeat biases that are present in their training data, which can lead to unfair or discriminatory outputs.   

Limited Long-Term Memory/Context

While LLM context windows are growing, they don't maintain continuous memory across extended conversations.  LLMs require explicit external mechanisms to overcome this limitation.