The Promise of Symbolic AI

There is growing consensus that LLMs are, by themselves, insufficient for achieving true AGI.

Due to their inherent weaknesses, LLMs cannot be trusted in domains where hallucinations cannot be tolerated.  There are many such domains: interfacing with the customer, managing contracts and law, medicine, coding, engineering, and more. This may partially explain why many predictions about AI outright replacing humans have not yet come to pass.  More than a few projects to replace humans with LLM-based systems have been rolled back.  

Fortunately, the weaknesses of LLMs are bridged by the strengths of symbolic AI.  The emerging field of Neuro-Symbolic AI aims to combine the two approaches, with the promise of developing a hybrid AI that can be trusted to be accurate, and to be honest about what it knows..  

Symbolic AI brings essential capabilities to the table:  

Reasoning and Factual Accuracy

An LLM will always give a confident answer to a question, whether it is correct or not.  A symbolic system will only give you an answer it knows to be correct or can be proven correct. Symbolic AI, with its explicit rules and logical inference, can serve as a truth engine for LLMs. 

An LLM might generate a hypothesis, and a symbolic system could then verify its logical consistency or factual accuracy against a structured knowledge base, ensuring correctness.

Interpretability and Trust

A symbolic AI can tell you what it knows and how it knows it.  It can provide transparent reasoning steps, proving how it arrived at a particular conclusion.  An LLM that leaned on a symbolic system would gain these capabilities.   

Structured Knowledge Integration

LLMs can be grounded in symbolic knowledge bases, providing them with a more reliable and structured understanding of entities, relationships and rules of inference.  These knowledge bases can be easily shared between systems, ensuring that facts and reasoning can be easily known and checked.  

Common Sense and Planning

Symbolic AI can provide explicit common sense rules and logical frameworks for planning and multi-step reasoning.  This can help LLMs with tasks that require more than just pattern matching, such as complex problem-solving or robotics.

Control and Safety

The rules in symbolic AI can act as guardrails, mitigating harmful outputs and biases. These can constrain LLM outputs to adhere to ethical guidelines, safety protocols, or domain-specific requirements.  

Efficient Learning in Narrow Domains

For tasks requiring precise, rule-based knowledge in specific domains, symbolic systems can quickly encode that knowledge for use.   A symbolic system’s knowledge base can be efficiently, incrementally updated, without requiring massive datasets and extensive retraining. 

In essence, a foundational symbolic reasoning system provides the rigor, transparency, and accountability that current LLM-based systems lack. 

While LLMs excel at learning from patterns and generating fluent text, symbolic AI serves as a logical bedrock - a foundation of truth and consistency.  Combining the two approaches could ensure that LLMs' impressive capabilities are imbued with correctness, reliability, and honesty regarding what is known and what is not known.   

Neuro-Symbolic AI is not about one approach replacing the other, it is about creating a more trustworthy form of AI that is a symbiotic combination of two old rivals.   

With luck, such a symbiosis will move us closer to AGI.