Symbolic AI / GOFAI

History of AI — Symbolic AI / GOFAI
History of AI — The Birth & Growth of Real AI
Course 2 · Chapter 4 · Symbolic AI / GOFAI

🔣 Symbolic AI / GOFAI

Logic Theorist, GPS, and ELIZA look quite different on the surface — a theorem prover, a puzzle solver, a chatbot. This chapter steps back to name the single underlying paradigm all three actually shared, the confident philosophical claim behind it, and the specific weakness in that claim that would eventually define the rest of this course's Real AI era.

🏷️ A Name Coined in Hindsight

The term "GOFAI" — Good Old-Fashioned AI — wasn't used by the researchers building this era's programs at the time. Philosopher John Haugeland coined it in 1985, retrospectively, specifically to distinguish this earlier symbol-manipulation approach from the different techniques (including the statistical, learning-based methods covered starting in Course 3) that were beginning to compete with it by the mid-1980s. Naming a paradigm "old-fashioned" is itself a signal that something newer had already started to take its place.

🧠 The Physical Symbol System Hypothesis

Newell and Simon — already met twice in this course, for the Logic Theorist and GPS — formalized the core assumption behind the entire era in a 1976 paper, as the physical symbol system hypothesis:

The Claim
"A physical symbol system has the necessary and sufficient means for general intelligent action."

In plain terms: a system that manipulates discrete symbols according to formal rules — nothing more, nothing categorically different — is both enough to produce genuine intelligence, and required for it. Not one possible route to intelligence among several; the necessary foundation for any intelligence at all, artificial or otherwise.

This is a genuinely strong philosophical claim, and it's worth being clear that it was, for two decades, close to the field's unquestioned default assumption — not a hedge, not one hypothesis among competing schools of thought, but the working foundation nearly everything covered so far in this course was built on top of.

🧊 SHRDLU (1970) — GOFAI's High-Water Mark

Terry Winograd's SHRDLU, built at MIT, is frequently cited as the clearest and most impressive demonstration of what symbolic AI could achieve. SHRDLU could hold a genuine conversation about a simulated "blocks world" — colored blocks, pyramids, and boxes on a virtual tabletop — understanding and correctly executing instructions like "find a block which is taller than the one you are holding and put it into the box," tracking pronoun references across a conversation, and even explaining its own past actions when asked why it had done something.

Impressive — Within an Extremely Narrow World

SHRDLU's apparent understanding worked because its entire universe of discourse — every object, every possible relationship, every valid action — was small, closed, and fully specified in advance. Language understanding, logical reasoning about the block world's state, and action planning were all tightly integrated, which is exactly why it worked so well: there was no open-ended real-world ambiguity for the symbolic representation to fail to capture.

⚠ Where the Hypothesis Started to Break

The Knowledge Acquisition Bottleneck

Every fact, every rule, every piece of world knowledge a symbolic system uses has to be manually identified and hand-encoded by a human — there's no mechanism for the system to learn new knowledge from raw experience the way Course 3's statistical approaches eventually would. This became known as the knowledge acquisition bottleneck: human common-sense knowledge is vast, constantly context-dependent, and full of exceptions, and hand-encoding even a meaningful fraction of it turned out to be a staggering undertaking. The Cyc project, begun in 1984 with the explicit goal of hand-encoding millions of pieces of everyday common-sense knowledge as formal symbolic rules, is still ongoing today — a genuinely sobering illustration of just how large the task actually is, decades after it began.

Closely related: symbolic systems tend to be brittle — SHRDLU inside its blocks world was remarkably capable, but the same architecture couldn't be meaningfully extended to, say, understanding a conversation about cooking, or sports, or anything outside its narrowly pre-specified domain, without essentially starting the hand-encoding process over from scratch. A system built entirely on explicit rules tends to fail completely, often nonsensically, the moment it steps outside the boundaries someone thought to encode in advance.

📜 Strengths and Weaknesses, Side by Side

StrengthCorresponding Weakness
Reasoning is transparent — you can read the exact rule that produced a conclusionEvery rule has to be written by hand, one at a time
Excellent for closed, well-defined domains (theorem proving, SHRDLU's blocks world)Brittle and often useless the moment the domain's boundary is crossed
Grounded in a clear, principled philosophical hypothesisThe hypothesis itself doesn't scale to the genuine scope of real-world common sense
🔮 The Pivot That's Coming

The knowledge acquisition bottleneck is the single biggest reason the field eventually shifted away from hand-encoded symbols toward the statistical, learn-from-data approach that dominates modern AI — the entire subject of Course 3. Instead of a person writing down every rule in advance, a statistical system infers patterns directly from large amounts of real-world data. That pivot doesn't happen for decades yet in this course's timeline — but the specific problem driving it is fully visible right here, in the 1970s and 80s, well before anyone had the data or computing power to actually attempt the alternative.

🤔 Questions to Sit With

Reflection 1

The physical symbol system hypothesis claims symbol manipulation is both necessary and sufficient for intelligence. Based on everything covered in this course so far, do you find the "necessary" half or the "sufficient" half of that claim more convincing — or less?

Reflection 2

The Cyc project has been hand-encoding common-sense knowledge since 1984 and is still not finished. What does that timeline tell you about the actual scale of "common sense," compared to how casually the phrase gets used in everyday conversation?

Reflection 3

SHRDLU could explain its own reasoning in a way that's genuinely harder for many modern statistical AI systems to do. Is transparent, explainable reasoning worth trading away for broader real-world capability — or should both be achievable at once?

🎯 What's Next

Next chapter: The First AI Winter — the Lighthill Report, the funding collapse that followed, and how the knowledge acquisition bottleneck this chapter covered helped bring it about.