Why This Matters
Artificial intelligence isn't just a technology topic. It's central to cognitive science because AI systems serve as computational models of the mind. Every milestone you'll study represents a hypothesis about how thinking works: Can reasoning be reduced to symbol manipulation? Does understanding require embodiment? When a machine solves problems, is it "thinking" in any meaningful sense? You're being tested on your ability to connect these engineering achievements to deeper questions about cognition, representation, and intelligence.
These milestones also trace the evolution of competing theories in cognitive science, from the symbolic AI paradigm (rules and logic) to connectionist approaches (neural networks and learning). Understanding why certain approaches succeeded or failed tells us something profound about the architecture of human cognition. Don't just memorize dates and names. Know what each milestone reveals about the nature of thought, and be ready to argue whether machines can truly "understand" or merely simulate understanding.
Symbolic AI and the Classical Paradigm
The earliest AI systems operated on the assumption that intelligence is symbol manipulation: thinking means applying logical rules to abstract representations. This symbolic AI approach dominated the field for decades and directly tested the hypothesis that human cognition works like a formal reasoning system.
Turing Test (1950)
- Proposed the "imitation game" as a behavioral criterion for intelligence. If a machine's responses are indistinguishable from a human's, it exhibits intelligent behavior.
- Sidestepped the consciousness question by focusing on observable behavior rather than internal states. This aligned with functionalist approaches in cognitive science, where what matters is what a system does, not what it's made of.
- Remains controversial because critics argue it tests linguistic mimicry, not genuine understanding. Searle's Chinese Room argument is the most famous challenge here: a system could pass the Turing Test by following rules without comprehending anything.
Logic Theorist (1956)
- First program to prove mathematical theorems. Developed by Newell and Simon, it proved 38 of the first 52 theorems in Principia Mathematica.
- Demonstrated heuristic search, using shortcuts rather than exhaustive calculation to navigate problem spaces. This mimicked how humans don't consider every possibility but instead use strategies to narrow the search.
- Launched the symbolic AI paradigm at the 1956 Dartmouth Conference, establishing the field's foundational assumption that thinking is computation.
General Problem Solver (1959)
- Aimed to be domain-general. Newell and Simon designed it to solve any problem that could be represented as a goal state and a set of operators for transforming states.
- Introduced means-ends analysis, a heuristic strategy that identifies the difference between the current state and the goal state, then selects an action to reduce that difference. This was directly modeled on verbal protocols of humans solving problems (thinking aloud while working through a task).
- Revealed limitations of pure symbolism. GPS struggled with problems requiring perceptual or embodied knowledge, foreshadowing later critiques that not all cognition reduces to rule-following.
Compare: Logic Theorist vs. General Problem Solver. Both used symbolic reasoning and heuristics, but Logic Theorist was domain-specific (math proofs) while GPS attempted domain-general intelligence. This distinction matters for questions asking about the generality of cognitive architectures.
Natural Language and the Understanding Question
These systems tackled one of the hardest problems in AI: human language. They raised a fundamental question about whether processing language is the same as understanding it, a debate that remains central to cognitive science.
ELIZA (1966)
- Simulated a Rogerian therapist using simple pattern matching and substitution rules. For example, if you typed "I feel sad," ELIZA might respond "Why do you feel sad?" It had no actual comprehension of meaning.
- Triggered the "ELIZA effect." Users attributed understanding and empathy to the program despite its shallow processing. This revealed a deep human tendency to anthropomorphize, to project mental states onto systems that don't have them.
- Became a touchstone for skeptics who argue that passing behavioral tests doesn't indicate genuine understanding. ELIZA demonstrates syntax without semantics: it manipulates words according to rules without grasping what they mean.
Watson Wins Jeopardy! (2011)
- Combined natural language processing with massive knowledge retrieval. IBM's Watson parsed complex, ambiguous clues and retrieved answers from millions of documents.
- Used statistical confidence scoring rather than "understanding" to select responses. Watson ranked candidate answers by probability, not by grasping their meaning. This raises the question: does performance equal comprehension?
- Demonstrated practical AI capability while leaving the understanding question unresolved. Watson couldn't explain why an answer was correct, only that its statistical model assigned it high confidence.
Compare: ELIZA vs. Watson. Both processed natural language, but ELIZA used simple pattern-matching rules while Watson used statistical learning over massive datasets. Neither demonstrates understanding in the philosophical sense, but Watson's performance is far more sophisticated. Both are relevant examples if you're asked about the Chinese Room argument.
Knowledge-Based and Expert Systems
Rather than pursuing general intelligence, this approach encoded human expertise into specialized systems. It reflected a shift toward narrow AI and tested whether intelligence could be captured as domain-specific rules.
Expert Systems (1970s)
- Encoded expert knowledge as if-then rules. Systems like MYCIN (which diagnosed bacterial infections and recommended antibiotics) and DENDRAL (which identified chemical structures) captured specialist reasoning in structured knowledge bases.
- Used inference engines to chain rules together, mimicking how human experts reason through cases step by step.
- Succeeded in narrow domains but failed to generalize. This revealed the knowledge acquisition bottleneck: much of human expertise involves tacit knowledge (intuitions, judgment calls, contextual awareness) that's extremely difficult to articulate as explicit rules. You can ask a doctor what they'd diagnose, but capturing everything they know about medicine in if-then statements proved impractical.
Connectionism and the Neural Network Revolution
The connectionist paradigm proposed that intelligence emerges from distributed processing across networks of simple units, inspired by the brain's architecture. This approach challenged symbolic AI's assumptions about discrete, explicit representations.
Backpropagation for Neural Networks (1986)
- Enabled multi-layer networks to learn. The backpropagation algorithm adjusts connection weights by calculating the error at the output, then propagating that error signal backward through the network's layers so each connection can be tuned.
- Revived connectionism after earlier neural network approaches (single-layer perceptrons) were shown to have severe limitations. Minsky and Papert had demonstrated in 1969 that perceptrons couldn't solve certain basic problems, which stalled neural network research for over a decade.
- Provided a learning mechanism that didn't require explicit programming of rules. The network discovers patterns from examples, which more closely models how humans learn from experience rather than from being told rules.
ImageNet and Deep Learning Breakthrough (2012)
- A convolutional neural network (CNN) slashed error rates. AlexNet reduced image classification errors dramatically on the ImageNet challenge, outperforming all previous approaches by a wide margin.
- Demonstrated the power of deep architectures with many layers, enabling hierarchical feature learning. Early layers detect simple features like edges, middle layers combine those into shapes, and later layers recognize whole objects. This hierarchy loosely parallels how the visual cortex processes information.
- Sparked the current AI revolution. Deep learning now dominates computer vision, speech recognition, and natural language processing, shifting the field away from hand-crafted features toward learned representations.
Compare: Expert Systems vs. Deep Learning. Expert systems required humans to explicitly encode knowledge as rules, while deep learning extracts patterns automatically from data. This reflects a fundamental debate in cognitive science: is knowledge represented explicitly (as symbols and rules) or distributed implicitly across neural connections?
Game-Playing AI and Strategic Intelligence
Games provide controlled environments to test AI capabilities in strategic reasoning, planning, and learning. Each milestone here represents a different approach to achieving superhuman performance.
Deep Blue Defeats Kasparov (1997)
- Used brute-force search plus evaluation functions. Deep Blue examined roughly 200 million positions per second, relying on raw computational power combined with hand-coded evaluation criteria rather than human-like intuition.
- Demonstrated narrow AI supremacy in a well-defined domain with clear rules and perfect information (both players can see the entire board).
- Did not learn or adapt. Its chess knowledge was hand-coded by human experts, making it a clear example of the symbolic AI approach applied to a specific task.
AlphaGo Defeats World Champion (2016)
- Combined deep learning with reinforcement learning. AlphaGo first trained on a large database of human games, then improved dramatically by playing millions of games against itself.
- Mastered a game considered too complex for brute force. Go has more possible board positions than there are atoms in the observable universe, so exhaustive search is impossible. Success required something closer to intuitive pattern recognition.
- Made moves human experts couldn't explain. This suggests the network discovered strategies beyond established human understanding, raising important questions about interpretability: if an AI system can't explain its reasoning, does it truly "know" something, or is it just pattern matching at a scale we can't follow?
Compare: Deep Blue vs. AlphaGo. Deep Blue used symbolic search and hand-coded evaluation; AlphaGo used neural networks and self-play learning. This contrast illustrates the shift from GOFAI (Good Old-Fashioned AI) to modern machine learning. Think about what each approach reveals about the nature of expertise and intuition: Deep Blue suggests expertise is search through a problem space, while AlphaGo suggests expertise involves learned pattern recognition that can't easily be reduced to explicit rules.
Quick Reference Table
|
| Symbolic AI / Classical Paradigm | Logic Theorist, General Problem Solver, Expert Systems |
| Connectionism / Neural Networks | Backpropagation, ImageNet/Deep Learning, AlphaGo |
| Natural Language Processing | ELIZA, Watson |
| Behavioral Tests of Intelligence | Turing Test, ELIZA |
| Domain-Specific vs. General AI | Expert Systems (narrow), GPS (attempted general) |
| Learning vs. Programmed Knowledge | AlphaGo, Deep Learning (learning) vs. Deep Blue, Expert Systems (programmed) |
| Human-Computer Interaction | ELIZA, Watson, Turing Test |
| Strategic Reasoning | Deep Blue, AlphaGo |
Self-Check Questions
-
Compare and contrast the Logic Theorist and AlphaGo. What do their different approaches reveal about competing theories of cognition in cognitive science?
-
Which two milestones best illustrate the ELIZA effect, and why does this phenomenon matter for debates about machine understanding?
-
If a question asks you to evaluate the Chinese Room argument, which milestones would you use as examples of systems that process symbols without understanding? Explain your choices.
-
What distinguishes the symbolic AI paradigm from the connectionist paradigm? Identify one milestone from each approach and explain how they reflect different hypotheses about the architecture of mind.
-
Both Deep Blue and AlphaGo achieved superhuman performance in games. Why is AlphaGo considered a more significant milestone for understanding human-like intelligence? What does this suggest about the role of learning in cognition?