The Promise of Symbolic AI

The modern AI story began in 1956 with the Dartmouth Summer Research Project on Artificial Intelligence. Here, computer science pioneers John McCarthy, Marvin Minsky, Herbert Simon and others gathered to explore the possibility of "thinking machines."  The ultimate goal was to build what we now call Artificial General Intelligence (AGI), i.e., a machine with the same reasoning capabilities as a human.

In the four decades that followed, AI research and development pursued various philosophical and technical approaches.  Progress came in fits and starts, driving cycles of boom and bust, hype and disillusionment. 

Early AI research quickly split broadly into two rival camps: symbolic and connectionist.

Symbolic AI (or "Good Old-Fashioned AI - GOFAI")

This approach, dominant in the early decades, posited that intelligence could be achieved by automating formal logics.  Knowledge is represented as explicit symbols that are manipulated using logical operations based on rules of inference.  The goal was to mimic human reasoning and problem-solving through explicit programming.  Languages such as LISP and Prolog were developed to facilitate this natively.

A simplified notion of symbolic AI is programming "if-then" rules over a vast knowledge base of facts. 

Connectionist AI (or Statistical AI) 

Inspired by the structure of the human brain, connectionism focused on artificial neural networks – interconnected "neurons" that learn by adjusting the strength of their connections. Instead of explicit rules, knowledge emerged from features in input data. This approach was less about logical deduction and more about pattern recognition and statistical learning from examples. Early connectionist models faced significant limitations and were widely ignored for decades.  More successful early statistical approaches included reinforcement learning and classification systems.  

A simplified notion of connectionist AI is relating input to output patterns in a massive spreadsheet.