The AI Revolution: From Expert Systems to Large Language Models

Three Winters and Three Springs of AI

The history of AI is a history of alternating periods of optimism and disappointment. The first wave of optimism (1950s–60s): McCarthy, Minsky, Simon promised to create thinking machines within 20 years. The “AI winter” of the 1970s: the complexity of the tasks was underestimated. The second wave of optimism (1980s): expert systems. The second winter (late 1980s): expert systems proved too expensive and brittle.

The third optimism — the current one — is based on three factors: big data, computational power (GPU), deep learning. Neural networks, inspired by the brain and abandoned in the 1960s, were revived in the 2010s with astonishing results.

AlphaGo (2016) defeated the world champion in Go — a game considered unattainable for machines. GPT-3 (2020) generated text indistinguishable from human writing. GPT-4 (2023) passed medical, legal, MBA exams. This is a qualitative leap.

Large Language Models: What Are They?

LLMs (Large Language Models) are neural networks trained on trillions of words of text to predict the next token. This is not “understanding” in a philosophical sense — it is statistical compression of an enormous corpus of human knowledge.

What they can do: translate, summarize, generate code, answer questions, conduct dialogue, write essays. What they cannot do: reliably distinguish fact from fiction (“hallucinations”), long-term planning, physical interaction with the world, genuine understanding.

The “transformer” architecture (Attention is All You Need, Google, 2017) is the technological basis of GPT. The “attention mechanism” allows the model to “pay attention” to different parts of the input text — this is the key innovative solution.

Question for reflection: LLMs “hallucinate” — confidently report falsehoods. How does your organization integrate AI tools while preserving critical information verification?