Ambition, Fantasy, and Disillusionment: AI's 'Golden Age' and the First AI Winter

The General Problem Solver (GPS): AI’s First Great Leap#

In 1957, within the laboratories of Carnegie Mellon University, three scientists witnessed the birth of history. The General Problem Solver (GPS), developed collaboratively by Allen Newell, Herbert Simon, and J.C. Shaw, successfully ran, marking the first major breakthrough in artificial intelligence.

GPS’s revolutionary nature lay in its adoption of “means-ends analysis” methodology. This program could decompose complex problems into a series of sub-goals, then seek means to achieve these sub-goals until ultimately solving the entire problem. More importantly, GPS achieved separation between problem-solving strategies and specific problem knowledge, meaning the same reasoning mechanism could be applied to different types of problems.

The collaboration between Newell and Simon stands as a classic in scientific history. After earning a physics degree from Stanford University, Newell encountered game theory at Princeton University, laying the mathematical foundation for his later AI Chronicle.

GPS excelled at solving formalized problems like the Tower of Hanoi and logical proofs, giving researchers hope for creating “general intelligence.” However, as later developments would prove, this success in controlled environments remained separated from true human intelligence by an unbridgeable chasm.

SHRDLU: The Language Miracle in a Blocks World#

If GPS demonstrated the possibility of machine reasoning, then Terry Winograd’s SHRDLU program, developed between 1968-1970, gave people their first glimpse of machines “understanding” natural language.

On a computer at MIT’s Artificial Intelligence Laboratory, SHRDLU created a virtual “blocks world.” Users could converse with the program in English, directing it to move blocks of different colors and shapes. When a user input “Put the red block on the blue block,” SHRDLU could not only understand this instruction but also execute the corresponding operation, even answering complex questions about the blocks world’s state.

Winograd himself was a dual expert in linguistics and computer science. At Stanford University, he served as both a computer science professor and held positions in the linguistics department. This interdisciplinary background enabled him to skillfully combine linguistic theory with computer programming.

SHRDLU’s demonstrations caused a sensation in academic circles. Media rushed to report on this computer program that could “understand” English, and the public began imagining a future where machines could freely converse with humans. However, Winograd himself quickly realized the fundamental limitations of this program—it could only operate in an extremely simplified blocks world and couldn’t handle real-world complexity.

ELIZA: The Digital Incarnation of a Psychotherapist#

Perhaps no early AI program better embodied the contradictory psychology of that era than Joseph Weizenbaum’s ELIZA. In 1966, this MIT computer scientist developed a relatively simple program that unexpectedly touched upon deep psychological mechanisms of human-computer interaction.

ELIZA’s most famous version was the DOCTOR script, which simulated a psychotherapist’s conversation with patients. The program used simple pattern matching and text substitution techniques to transform user inputs into seemingly profound questions. When a user said “I feel depressed,” ELIZA might respond “Why do you feel depressed?” or “Tell me more about that feeling.”

What shocked Weizenbaum was that many users developed deep emotional dependencies on ELIZA. His secretary, when using the program, even asked him to leave the room to protect her privacy with the “therapist.” This phenomenon later became known as the “ELIZA effect”—people’s tendency to attribute human qualities to computer programs, even when they rationally know these are merely executing preset rules.

Weizenbaum later wrote: “I had not realized that extremely short exposures to a relatively simple computer program could induce powerful delusional thinking in quite normal people.” This discovery would fundamentally change his view of artificial intelligence, transforming him from an enthusiastic AI supporter into one of its sharpest critics.

Other Important Achievements#

The Golden Age’s achievements extended far beyond these three landmark programs. Arthur Samuel’s checkers program demonstrated early possibilities of machine learning, capable of improving its strategy through self-play and eventually reaching amateur expert level. The rudiments of expert systems also emerged during this period, as researchers began exploring how to encode human expert knowledge into computer programs.

Combinatorial Explosion: The Nightmare of Computational Complexity#

The term “combinatorial explosion” sounds technical, but it describes one of AI’s most fundamental challenges. Simply put, when problem scale increases slightly, the number of possible solutions grows exponentially, quickly exceeding any computer’s processing capacity.

Take chess as an example: while the rules are relatively simple, the number of possible games is estimated to exceed 10^120—a number greater than the atoms in the observable universe. Early AI programs attempted to solve problems through exhaustive search but quickly discovered this approach was completely powerless when facing real-world complexity.

Researchers tried using heuristic search to alleviate this problem, but these methods often only worked effectively in specific domains and couldn’t achieve true generality. While GPS performed excellently on formalized problems, when facing slightly more complex real-world issues, it would become trapped in endless search spaces.

The Microworld Trap: False Prosperity in Simplified Environments#

SHRDLU’s success largely depended on its extremely simplified “blocks world.” In this world, there were only geometric shapes of a few colors, without shadows, textures, or complex physical properties. More importantly, all rules of this world were hardcoded, and the program couldn’t learn new concepts or adapt to environmental changes.

Winograd later deeply reflected on this problem. He realized that SHRDLU’s “understanding” was completely illusory—the program didn’t truly understand language meaning; it was merely manipulating symbols within a predefined rule system. When attempting to apply similar methods to the real world, researchers found that the number of rules needed was astronomical, and interactions between these rules would produce unpredictable consequences.

This “microworld trap” wasn’t limited to SHRDLU. Many early AI systems performed excellently in simplified environments but would immediately fail once removed from these controlled conditions. This fragility became a common ailment of early AI systems and an important reason for disappointment among the public and funding agencies.

The Common Sense Knowledge Problem: The Irreproducibility of Human Intuition#

Perhaps most perplexing was the so-called “common sense knowledge problem.” Humans rely on vast amounts of background knowledge and common sense reasoning in daily life, most of which is implicit—we’re not even aware we’re using it.

For example, when we hear “John opened the door with a key,” we automatically understand that keys are used to unlock doors, that John can pass through once the door is open, and so forth. But for computers, these “obvious” inferences require explicit programming. Worse still, the amount of common sense knowledge is infinite and highly context-dependent.

Marvin Minsky proposed “frame” theory to attempt solving knowledge representation problems, but with limited effect. The common sense problem is considered “AI-complete”—meaning solving this problem requires achieving human-level artificial general intelligence, which was precisely the ultimate goal researchers were trying to achieve.

Report Background: The British Government’s Scientific Assessment#

In 1972, the British Science Research Council commissioned Sir James Lighthill to assess the nation’s artificial intelligence research status. Lighthill’s selection was no coincidence—he was the Lucasian Professor of Mathematics at Cambridge University, a position once held by Newton and later by Hawking.

Lighthill enjoyed worldwide reputation in fluid mechanics, founding the discipline of aeroacoustics and proposing the famous “Lighthill’s eighth power law,” making important contributions to jet engine noise control. As a scientist who achieved practical results in applied mathematics, his assessment of AI Chronicle carried special authority.

Report’s Core Views: Comprehensive Questioning of AI Promises#

The Lighthill Report, published in 1973, delivered merciless criticism of AI Chronicle. The report’s core conclusion was: “In no part of the field have the discoveries made so far produced the major impact that was then promised.”

Lighthill particularly criticized fundamental research areas like robotics and language processing. He pointed out that while these studies were theoretically interesting, they had made virtually no progress in practical applications. More seriously, he analyzed combinatorial explosion and microworld problems in detail, considering these technical obstacles fundamental and impossible to solve through simple technical improvements.

The report also questioned AI Chronicle’s resource allocation. Lighthill believed that relative to the massive funding and human resources invested, AI Chronicle’s output was disappointing. He recommended reallocating resources to more promising research areas.

Report Impact: AI Chronicle’s Funding Crisis#

The Lighthill Report’s impact was immediate. Based on this report’s recommendations, the British government drastically cut AI Chronicle funding, with all university AI Chronicle projects losing government support except those at Edinburgh University and Essex University.

This influence quickly spread to other countries. The U.S. DARPA (Defense Advanced Research Projects Agency) also began reevaluating its AI investment strategy, with many previously well-funded research projects forced to scale down or stop completely. Academia experienced brain drain, with many AI Chronicleers turning to other more promising fields.

After the report’s publication, it also triggered a famous public debate. On May 9, 1973, at the Royal Society in London, Lighthill engaged in heated debate with AI field leaders including John McCarthy and Donald Michie. Although AI Chronicleers mounted strong defenses of their work, public and policymaker confidence had already suffered serious damage.