Chapter audiobook
हम यहाँ तक कैसे पहुँचे
Listen to the complete teaching text, with spoken descriptions of diagrams. The lesson activities remain available below.
53:01 remaining at 1×
Press play to listen and follow the words here.
With this player focused: Space plays or pauses; arrow keys skip fifteen seconds.
Chapter sections and transcript
Download the transcriptHindi synthetic narration using Resemble AI Chatterbox-Multilingual-hi, MIT. Original Addaly teaching text adapted for listening.
A name from the 1950s
In 1950 Alan Turing published a paper that replaced the question of whether machines think with an imitation-game discussion: could an interrogator distinguish a machine from a person through written exchanges? It helped frame a research question; it was not a complete definition of intelligence. The field label used in the Dartmouth proposal came later.
In 1955 John McCarthy and three colleagues used “artificial intelligence” in a proposal for a summer research project at Dartmouth College in the United States. They proposed ten people working for two months in 1956 and hoped for a significant advance in language, learning and other aspects of intelligence. The meeting took place that summer. It helped establish a name and a research community, rather than finishing the programme the proposal imagined.
The proposal is worth remembering because an ambition and a demonstrated result are different things. In McCarthy’s later account, participants attended for different periods and pursued different agendas. A research meeting can matter even when it does not settle its largest questions.
Two recurring approaches
Two recurring approaches help organise the history, although they are not exhaustive or mutually exclusive. You can write the rules, or you can learn from examples. Real systems can combine learned components with search, rules and databases.
The rule-writers dominated early. The Logic Theorist in 1956 proved theorems from a book of mathematics. ELIZA, in 1966, held a conversation by matching patterns and reflecting them back, and its author was disturbed by how many people confided in it.
Learning research also began early. Frank Rosenblatt described the perceptron in 1958, a model whose connection strengths could be adjusted using examples. Minsky and Papert’s 1969 analysis showed important limitations of particular perceptron architectures. Those results did not prove that every multilayer network was incapable of learning, and neural-network research continued through later changes in funding and fashion.
Both approaches contributed useful tools. Explicit rules can make a narrow procedure inspectable; learning can capture patterns that are difficult to enumerate. Neither guarantees that a demonstration will generalise, and neither disappeared when the other became more prominent.
The pattern
One useful, simplified chronology groups several waves of activity like this:
- 1956 to the 1970s. Symbolic programs and ambitious research projects attracted support. Critical evaluations, including the 1973 British Lighthill report, accompanied later funding retrenchment in parts of the field.
- 1980s to the early 1990s. Expert systems brought commercial deployments. Maintenance costs, unmet expectations and changes in the specialised-computer market contributed to another downturn.
- 2012 onward. The ImageNet result accelerated interest in deep learning. Earlier neural-network work, including speech recognition, helped prepare the way; large language models later brought another wave of public attention.
These waves include both useful systems and expectations that outran evidence. Their causes differed; they are not three identical repetitions. The next lessons examine generalisation and maintenance as practical risks, without treating them as the only explanations for changes in research funding.
Search, expertise and learning in games
One event confuses the pattern and is worth placing correctly. In 1997 IBM’s Deep Blue beat Garry Kasparov, the world chess champion. It was reported as the arrival of machine intelligence.
Deep Blue combined specialised search hardware, a complex evaluation function, databases and chess expertise. Its authors report both manual tuning and automated training tools for some evaluation weights; it was not a self-play-trained deep neural network. Search speed varied with the position, reaching around two hundred million positions a second in quieter positions. AlphaGo’s 2016 victory over Lee Sedol, four games to one, combined learned policy and value networks with Monte Carlo tree search. Go’s enormous search space made exhaustive search impractical, not all search impossible. A later AlphaGo version defeated Ke Jie in 2017.
The lesson to carry from this: “a machine beat a human at X” is not enough to identify its method or its limits. Inspect the task, training, search, data and evaluation before drawing conclusions about what it might do elsewhere.
A timeline to keep
1950 Turing’s paper. 1955 the Dartmouth proposal and 1956 the workshop. 1958 Rosenblatt’s perceptron description. 1966 ELIZA. 1970s funding contractions in parts of AI. 1980s expert systems. 1986 a major popularisation of back-propagation for multilayer networks. 1997 Deep Blue. 2012 AlexNet’s ImageNet result. 2016 AlphaGo and Lee Sedol. 2017 the transformer paper. 2020 GPT-3. 2022 ChatGPT’s public research preview.
What the pattern does and does not tell you
It tells you to distinguish a demonstration from a dependable deployment, and a funding announcement from an earned return. It does not tell you when, or whether, another downturn will come. Adoption, useful performance, operating costs, research progress and financing can move differently. A short historical timeline is a guide to questions, not a reliable forecast of the next cycle.
Check your understanding
What does Deep Blue’s 1997 victory establish about the method it used?
Pick the one you would defend. Your answer stays private.
Your place and first choices stay on this browser. Classroom credit is recorded separately.
Stuck on something?Ask Meenal, your AI tutor