[ALT] Symbolic AI & Expert Systems (GOFAI)
Good Old-Fashioned AI (GOFAI) — the paradigm that intelligence consists of manipulating symbolic representations using explicit logical rules — was the dominant AI approach from the 1950s through the 1980s and generated two "AI winters" when it failed to scale to real-world complexity.
The Fork
What won: Statistical machine learning and deep learning — training systems from data using gradient descent, without explicit rules. This approach, largely dismissed as "not real AI" by GOFAI proponents, now underpins every major AI system.
What was abandoned (or severely marginalised): Symbolic AI — logic-based reasoning, expert systems, knowledge representation, semantic nets, production rules. Commercial expert systems (XCON/R1 for DEC, MYCIN for medical diagnosis, PROSPECTOR for geology) achieved genuine commercial value in narrow domains in the 1980s before becoming maintenance nightmares that couldn't generalise.
The Technical Division
GOFAI's central bet: human intelligence is fundamentally about symbol manipulation (Newell & Simon's "Physical Symbol System Hypothesis," 1976 Turing Award lecture). Feed the computer enough symbolic rules and it will be intelligent. The failures: (1) the "frame problem" — it's impossible to enumerate all the things that don't change when an action occurs; (2) brittleness — systems failed on inputs outside their explicit rule base; (3) knowledge acquisition bottleneck — getting experts to articulate their tacit knowledge as rules proved far more expensive than building the systems.
What GOFAI Got Right
GOFAI is not dead. It thrives in: theorem provers (Lean4, Coq, Isabelle — verifying mathematics and software), constraint solvers (SAT/SMT solvers used in chip design), knowledge graphs (Google Knowledge Graph, Wikidata), planning systems (NASA spacecraft scheduling), and formal verification. The Cyc project (1984–present) has accumulated 25M+ symbolic logical assertions about the world and remains the largest explicit knowledge base.
The Hybrid Future
The current frontier is neurosymbolic AI: combining the statistical pattern-matching of neural networks with the logical reasoning and compositionality of symbolic systems. DeepMind's AlphaGeometry (2024) solved IMO geometry problems by combining a neural model with a symbolic geometry engine. Chain-of-thought prompting in LLMs is a form of implicit symbolic reasoning. The fork may be healing.
Discovery Character
Surprise level: High — the AI community's conviction that symbolic manipulation was the essence of intelligence was near-universal until the statistical learning revolution of the late 1980s–1990s. The idea that a system trained purely on data patterns could surpass rule-based systems was heretical in the 1970s GOFAI world.
Mode of overthrow: Systematic empirical defeat — data-driven methods consistently outperformed symbolic methods on benchmark after benchmark throughout the 1990s–2000s, until by 2010 the evidence was overwhelming.