Early Programs
💻 Early Programs
🧩 The General Problem Solver (1957)
Fresh from the Logic Theorist's success at Dartmouth (Chapter 2), Newell, Simon, and J.C. Shaw set an even more ambitious goal: a single program capable of solving any problem that could be formally represented — not just logic proofs, but puzzles, formal reasoning tasks, anything reducible to a clear starting state and a goal state. They called it the General Problem Solver, or GPS.
GPS worked by comparing its current state to the desired goal state, identifying the biggest difference between them, and searching for an action that would reduce that difference — repeating the process until the gap closed entirely. It could genuinely solve problems like the Tower of Hanoi this way, treating wildly different puzzle types as instances of the same underlying search structure.
GPS's fundamental limitation set the tone for an enormous amount of AI research to follow: it worked well on cleanly formalized problems with well-defined states and rules, and essentially couldn't function on anything resembling the genuine messiness of the real world — ambiguous goals, incomplete information, situations that resist being reduced to a tidy "current state vs. goal state" comparison. This gap — between dazzling performance on a clean, formal toy problem and near-total failure on anything genuinely real-world — recurs constantly throughout the rest of this course, right up to the deep learning era in Course 3.
🛋️ ELIZA (1966)
Joseph Weizenbaum, at MIT, built something far simpler than GPS in ambition, and far more consequential in what it accidentally revealed. ELIZA simulated conversation with a Rogerian psychotherapist — a real therapeutic style built around reflecting a patient's own statements back to them as open-ended questions, which turned out to be an almost perfect match for what simple text pattern-matching could actually pull off.
ELIZA: In what way?
User: They're always bugging us about something or other.
ELIZA: Can you think of a specific example?
User: Well, my boyfriend made me come here.
ELIZA: Your boyfriend made you come here?
Underneath, ELIZA had no model of meaning at all — it matched keywords against a script (Weizenbaum's most famous script was named DOCTOR), applied simple transformation rules (swapping "my" for "your," "I" for "you"), and reflected fragments of the user's own input back as a question, falling back to generic prompts when nothing matched. No understanding, no memory of the conversation beyond simple pattern rules, no actual model of what a "boyfriend" or "bugging" meant.
😮 The "ELIZA Effect"
Weizenbaum reported being genuinely disturbed by how readily people attributed real understanding and empathy to ELIZA — most famously, his own secretary reportedly asked him to leave the room so she could have a private conversation with the program, despite knowing exactly how it worked and having watched it being built. This phenomenon was later named the "ELIZA effect": the tendency to unconsciously treat a computer's behavior as though it reflects genuine human-like understanding, even when the observer knows perfectly well the underlying mechanism is simple and mechanistic.
ELIZA is, in a very real sense, people unconsciously running their own private Turing Test (Chapter 1) — and the machine passing it, for many users, despite the mechanism being about as simple as a mechanism can get. This is the clearest confirmation yet, in this course's actual timeline, of the Mechanical Turk pattern from Course 1: a sufficiently well-targeted convincing performance gets believed, almost regardless of how sophisticated the thing producing it actually is.
🔄 From AI Success Story to AI's Sharpest Critic
What happened next is genuinely striking: Weizenbaum, the creator of one of the most famous early AI "successes," became one of the field's most prominent and outspoken skeptics as a direct result of what he'd witnessed. His 1976 book Computer Power and Human Reason argued forcefully against trusting computers with tasks requiring genuine human judgment, wisdom, or compassion — precisely because he'd seen firsthand how easily people could be fooled into over-trusting a system with no genuine understanding at all.
📜 Two Programs, Two Very Different Legacies
| General Problem Solver (1957) | ELIZA (1966) | |
|---|---|---|
| Goal | Solve any formally-representable problem | Simulate a specific style of conversation |
| Underlying approach | Means-ends analysis, state-space search | Keyword pattern-matching and reflection |
| Where it succeeded | Clean, formal toy problems (Tower of Hanoi) | Convincing many real users of genuine understanding |
| Lasting legacy | Exposed the toy-problem-vs-reality gap | Named a whole psychological phenomenon (the ELIZA effect) |
🤔 Questions to Sit With
ELIZA's specific choice of a Rogerian therapist persona (which naturally involves reflecting statements back as questions) was arguably a big part of why it worked as well as it did. Do you think ELIZA would have been nearly as convincing simulating a different kind of conversation — say, a friend giving advice?
Weizenbaum's secretary knew exactly how ELIZA worked and still wanted a private conversation with it. Have you ever caught yourself experiencing something like the ELIZA effect — responding emotionally to a system you know, intellectually, has no real understanding?
GPS aimed to be maximally general and struggled everywhere except clean formal problems; ELIZA aimed to be narrowly specific and accidentally fooled real people. What does the contrast between these two outcomes suggest about the relationship between a system's scope of ambition and how convincing it actually turns out to be?
🎯 What's Next
Next chapter: Symbolic AI / GOFAI — the broader "Good Old-Fashioned AI" paradigm that Logic Theorist, GPS, and most of this era's research shared, and the core assumption behind it that would eventually prove to be its greatest weakness.