AI Fundamentals: A Landscape of Approaches
Artificial intelligence" has meant several genuinely different things at different points in its own history, and a good deal of confusion in how people talk about it today comes from treating those as one continuous technology rather than as a sequence of distinct paradigms, each with its own assumptions about what "intelligence" even means to build. The field was named, deliberately, at a 1955 proposal for a summer workshop at Dartmouth College, which set out to explore the conjecture "that every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it" [1]. Seventy years on, the machines being built to test that conjecture look almost nothing like what its authors imagined — and the differences between the approaches tried along the way are still the most useful map available for understanding what a given AI system can and cannot do.
Four Paradigms, Not One Technique
The earliest paradigm, dominant from the 1950s into the 1980s, is usually called symbolic AI or "good old-fashioned AI": intelligence modelled as explicit manipulation of symbols and rules — logic programs, expert systems built from hand-written if-then rules, search over formally defined problem spaces [2]. It excels where a domain really can be captured as a clean rule set (a chess engine's legal-move generator, a tax calculator) and fails, often badly, wherever the rules themselves are the hard part — which turns out to be most of the interesting cases.
The second paradigm, statistical machine learning, replaces hand-written rules with rules inferred from data: instead of a programmer specifying "if income exceeds X and history includes Y, decline the loan," the system is shown thousands of past decisions and learns the pattern for itself. This shift — from specifying the rule to specifying the objective and letting the system find the rule — is the single most consequential idea separating pre-1980s AI from everything that followed, and it is why a modern AI system's behaviour is usually described in terms of what it was trained on rather than what it was told to do [2].
The third paradigm, deep learning, is a specific and now-dominant way of doing statistical learning: instead of hand-engineering which features of the data matter (this pixel pattern means "edge," that word co-occurrence means "sentiment"), a many-layered neural network learns its own internal representations directly from raw data, layer by layer. This removed the single biggest bottleneck in the previous approach — a human deciding, by hand, what the machine should even look at.
The fourth and most recent paradigm, sometimes called foundation models, is less a new algorithm than a new way of building and deploying deep learning systems: train one very large model on a very broad swath of data once, then adapt that same base model — through further training or simply through the instructions given to it — to a wide range of specific tasks it was never explicitly trained for [4]. Large Language Models, the subject of the rest of this module, are the most visible example of this paradigm, but it also covers image, audio, and multimodal systems built the same way.
(hand-written rules)"] --> B["Statistical ML
(rules learned from data)"] B --> C["Deep Learning
(representations learned too)"] C --> D["Foundation Models
(one base model, many tasks)"] style D fill:#FFC857
None of these paradigms fully replaced the last. Symbolic techniques still run the route-planner in a satnav and the rule-checker in a compiler; classical statistical models still power most credit-scoring and fraud-detection systems, precisely because their behaviour is easier to audit and explain than a deep network's. Choosing "which AI" for a task is still, in practice, choosing which paradigm fits the problem's actual shape.
Narrow and General Intelligence
A second, orthogonal distinction matters as much as the paradigm one: almost every AI system actually deployed today, including the most capable LLMs, is a form of narrow AI — built for, and only reliably good at, a bounded range of tasks, however broad that range might feel from the outside. "General AI" — a system with human-like, transferable intelligence across essentially any domain — remains a research aspiration, not a category any current system belongs to [2]. This distinction is not academic pedantry; it is the single most important thing to keep in mind when evaluating a confident-sounding AI output, because a system that is narrowly excellent at fluent, plausible-sounding text generation is not thereby narrowly excellent at truth — those are different tasks that happen, most of the time, to produce similar-looking output.
A Brief, Honest History
AI's history is not a smooth climb; it is a sequence of booms followed by winters, and understanding why the winters happened is a useful inoculation against overclaiming now. The first, deepest AI winter followed a 1973 report commissioned by the UK's Science Research Council from applied mathematician Sir James Lighthill, which reviewed the field's progress against its own founding promises and concluded, bluntly, that "in no part of the field have the discoveries made so far produced the major impact that was then promised" [3]. UK funding for AI research collapsed in its wake, and a parallel loss of confidence in the US followed soon after. The pattern repeated in the late 1980s, when a second wave of AI — expert systems built on symbolic rules — proved expensive to maintain and brittle outside the narrow domains they were built for. Each winter followed the same shape: a genuine capability jump generated promises the underlying technique could not keep, funding and attention collapsed, and the field's actual, more modest progress continued underneath the hype cycle regardless. That pattern is worth having in mind for every claim made about AI's capabilities in the rest of this module — including the ones that turn out to be true.
A Taxonomy of AI Applications You Will Meet This Module
The rest of this module works through four broad application types, and it is worth distinguishing them now because they fail in different ways and need different kinds of scrutiny:
- Generative systems produce new content — text, in this module's case — from a prompt. Their central risk is confident-sounding output that is wrong, covered in Understanding Large Language Models and Text Generation, Summarisation & Paraphrasing.
- Retrieval systems find and surface existing information rather than generate new content, and are combined with generative systems in Retrieval-Augmented Generation specifically to reduce the generative risk above.
- Conversational / question-answering systems maintain state across a dialogue, covered in Question-Answering Systems.
- Agentic systems plan and execute multi-step tasks with some autonomy, orchestrating the other three — the subject of LLM Orchestration, Context Engineering & Agentic AI.
Almost every AI product you will meet, in this module and afterwards, is some combination of these four, wrapped around a foundation-model core.
References
- McCarthy, J., Minsky, M. L., Rochester, N., & Shannon, C. E. (1955/2006). A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence, August 31, 1955. AI Magazine, 27(4), 12. https://doi.org/10.1609/aimag.v27i4.1904
- Russell, S., & Norvig, P. (2020). Artificial Intelligence: A Modern Approach (4th ed.). Pearson.
- Lighthill, J. (1973). Artificial Intelligence: A General Survey. In Artificial Intelligence: A Paper Symposium. Science Research Council, London.
- Bommasani, R., Hudson, D. A., Adeli, E., et al. (2021). On the Opportunities and Risks of Foundation Models. Stanford Center for Research on Foundation Models. arXiv:2108.07258