2023 Thinking

John's term. The tendency to hold strong, confident opinions about what AI can and cannot do that were formed during an early encounter with the technology — typically shortly after ChatGPT's public release in November 2022 — and never revisited.

It sounds like this:

• AI is useless because it hallucinates while sounding authoritative.
• AI makes up citations that don't actually exist.
• As a developer I spend more time debugging AI-generated code than writing it myself.
• AI doesn't know anything after the last date of its training data.

Every one of these was a fair description of something real. Several remain partly true — citation fabrication is still a genuine hazard, which is why verification is graded so heavily in this course.

What makes it 2023 thinking is not that the claims were wrong. It is that they were formed once, about a fast-moving technology, and then treated as settled. Validation and testing still matter enormously. So does noticing when the thing you tested has changed underneath your conclusion.

Worth turning on yourself: whatever you believe about AI right now, when did you last check?

Agentic Coding

Directing an AI system that doesn't just suggest code but writes it, runs it, reads the errors, fixes them, and reports back — working through a task over many steps rather than answering one question. Claude Code is an example.

AI Effect

The tendency for a problem to stop counting as artificial intelligence the moment it's solved. Chess, optical character recognition, speech recognition, and spam filtering were each landmark AI achievements; each is now "just software." Often summarized as "AI is whatever hasn't been done yet."

AI Slop

Low-quality AI-generated content produced at volume and published without human review — filler articles, fake product reviews, plausible-sounding nonsense. The term is pejorative and deliberately so. The concern isn't that any single piece is bad, but that the sheer quantity degrades the usefulness of the places we look for information.

AI Winter

A period in which artificial intelligence fails to deliver on its promises, funding collapses, and the field goes quiet — typically following a period of intense hype. There have been two, roughly 1974–1980 and 1987–1993.

Artificial Intelligence

There is no single accepted definition. One possible definition: a subfield of Computer Science focusing on the development of mechanical and software systems capable of performing tasks that normally require human thinking, reasoning, or intelligence. What counts as AI also shifts over time — see AI Effect. The term was coined by John McCarthy in the 1955 proposal for the 1956 Dartmouth workshop.

Another definition, based on the perspective of Stuart Russell and Peter Norvig, authors of the field's standard textbook: the study and design of agents — systems that perceive their environment and take actions to achieve the best possible outcome, or, when the outcome is uncertain, the best expected one. This framing says nothing about humans at all: an agent is intelligent if it acts well, whether or not a person could do the same task.

While modern society often equates AI with recent developments in Large Language Models, the field is much broader and includes:

• Mathematical reasoning and logical deduction
• Perception (facial recognition, image interpretation)
• Learning
• Intelligent decision-support
• Problem-solving
• Strategic game play
• Understanding language
• Pattern recognition
• Planning and scheduling
• Computer-aided programming
• Auto-generation of images, videos, music, and other media
• Robotics, including autonomous vehicles

Big Data

Not a technique — a change in circumstances. From the mid-2000s on, ordinary activity started leaving machine-readable traces at a scale nobody had before: e-commerce transactions, search queries, social networks, card payments, image and video libraries, sequenced genomes, online games. Learning from examples only works if you have the examples, and this is where they came from. Most of the methods behind the current boom are decades old. The data is not.

Holding it takes a database; finding patterns in it is data mining; learning from it is machine learning.

Boolean Logic

Reasoning with values that can only be true or false, combined with and, or, and not. Named for George Boole, whose An Investigation of the Laws of Thought (1854) argued that human reasoning could be written down as algebra. He thought he was doing philosophy. He had in fact built the arithmetic that every digital computer would later run on.

Chatbot

A system you converse with in ordinary language, which replies with generated text. Modern chatbots are built on large language models.

The thing to hold onto: a chatbot writes, where a search engine finds. It is producing text likely to follow your question, not retrieving a document someone wrote. That is the source of both its usefulness and its characteristic failures — see citation fabrication.

Citation Fabrication

When a language model produces a reference that looks entirely correct — plausible authors, a real journal, a well-formed DOI — for a paper that was never written.

It happens for a mundane reason. The model generates text that is likely to follow, and a citation is a highly patterned genre. A fabricated one is every bit as likely-looking as a real one, which is precisely what makes it convincing. The model is not lying in any intentional sense; it has no concept of whether the paper exists.

Lawyers have been sanctioned for filing briefs containing fabricated case law. Papers have been retracted. It is the most common way a confident AI user gets publicly embarrassed, and the reason Homework 1 weights verification above everything else.

The fix is not cleverer prompting. It is opening the source.

Claude Code

A version of Claude that works on files and folders instead of in a chat window: it writes programs, runs them, sees what breaks, fixes it, and reports what it did. It runs in a terminal, inside code editors, and inside Claude Desktop under a separate Code tab — the same engine on different surfaces. It is turned on separately and you do not need it for Homework 1; it is how the larger builds later in the term get made. These course materials were themselves built with it.

Claude Desktop

The Claude application installed on your own Mac or PC, as opposed to using claude.ai in a browser. Same models, same account, same Projects — they sync. What the app adds is reach: its own window and keyboard shortcut, and the ability to work with things on your own computer, such as your files and other applications you connect to it. If a managed or work-issued laptop will not let you install it, the browser version does everything this course asks of you.

Connectionism

The approach to AI that says: don't write down rules — build a network of simple units and let it learn the rules from examples. The rival tradition to Symbolic AI. Neural networks, deep learning, and every large language model descend from it.

Cyc

The longest-running project in the history of AI: an attempt, begun by Doug Lenat in 1984 and still going, to write common sense down — millions of hand-authored facts and rules of the water is wet, things fall when you drop them, people have emotions variety — so that software could reason about the everyday world. Lenat's estimate of the job: "intelligence is ten million rules."

It is the purest expression of symbolic AI and the clearest illustration of its limit. Forty years of human effort went into Cyc, and a language model trained on text arrived at common-sense fluency by a completely different road.

Hold both thoughts, though: Cyc can tell you exactly why it believes something and show you the chain of reasoning. A language model cannot.

Dartmouth Workshop

The summer meeting at Dartmouth College in 1956 where the field got its name. John McCarthy coined the term artificial intelligence in the proposal, and convened it with Marvin Minsky, Nathaniel Rochester and Claude Shannon on a bracing 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."

Seventy years on, that sentence is still the field's open question rather than its settled premise.

Data Mining

Searching a large body of data for patterns nobody asked about in advance. The distinction worth holding on to: most analysis starts with a question and looks for the answer, while data mining starts with the data and looks for anything surprising in it — clusters of customers who behave alike, a combination of symptoms that keeps recurring, a fraud pattern no one had named.

It overlaps with machine learning and often uses the same algorithms, but the goal differs: machine learning builds a model to reuse on new cases, whereas data mining is trying to discover something about the data you already have. It needs big data to be worth doing and a database to be practical, and it sits inside data science.

The catch is that a large enough dataset contains striking patterns by pure chance. Finding one is easy; showing it means anything is the work.

Data Science

The discipline of getting useful, defensible answers out of data — finding it, cleaning it, analysing it, modelling it, and communicating what it does and does not show. It overlaps heavily with machine learning but is broader: a data scientist spends far more time deciding what question is worth asking, and whether the data can honestly answer it, than building models.

Worth knowing because most of the vocabulary we use to talk about what goes wrong with AI — selection bias, base rates, correlation versus causation — comes from data science and statistics rather than from AI research. Those fields have been thinking about how data misleads people for a century longer.

It is the umbrella over most of the other terms here: big data is the raw material, a database is where it lives, data mining is one thing you do with it, and machine learning is the part that builds a model.

Database

Organized storage that lets you ask questions of data instead of reading through it. A spreadsheet holds numbers; a database holds numbers and knows how they relate — which order belongs to which customer, which reading came from which sensor — so a question like "what did shoppers in this age group buy after their first return?" can be answered in one query rather than by hand.

Unglamorous and almost invisible, and everything runs on it: banks, hospitals, airlines, universities, and every website you have ever logged in to. When people say big data changed what was possible, part of what changed was databases learning to hold that much. Nothing in data science or data mining happens without one somewhere underneath.

Deep Learning

Machine learning using neural networks with many layers stacked between input and output, each layer transforming what the layer below it produced. "Deep" refers to the number of layers and nothing more. The idea is old; what made it work was big data, GPUs, and the software libraries that made both usable by ordinary researchers. Everything in the current wave — image recognition, speech, translation, large language models — is deep learning.

ENIAC

The first programmable, general-purpose electronic computer: built by John Mauchly and J. Presper Eckert at the University of Pennsylvania and completed in 1945. Thirty tons, 17,000 vacuum tubes, half a million 1945 dollars — about $8.7 million today. "Programming" it meant physically rewiring it, a job that took days.

It is the bottom rung of the ladder that ends at vibe coding. The distance travelled in eighty years is from rewiring a room to describing what you want in a sentence.

Expert System

A program that captures a human specialist's knowledge as a large set of explicit IF–THEN rules and applies them to give advice. The dominant commercial form of AI in the 1980s. MYCIN diagnosed blood infections from about 600 rules; XCON configured computer orders and grew past 10,000. Their limitation is that rules don't cover enough and don't compose — see Polanyi's Paradox.

Generative AI

AI that produces new artifacts — text, images, audio, video, code — rather than sorting, scoring, or predicting from things that already exist. The distinction that matters here: a system that classifies hands you a label you can check against the world, while a system that generates hands you something plausible, and plausibility is not the same as correctness. See hallucination and citation fabrication.

GOFAI

"Good Old-Fashioned AI." A slightly affectionate, slightly dismissive name for Symbolic AI.

GPU

Graphics processing unit. A chip designed to draw video-game graphics, which turns out to be built from exactly the arithmetic neural networks need — thousands of small multiplications at once, rather than one complicated instruction at a time. A modern consumer card does over a thousand times the work of the fastest supercomputer of 1990, for a couple of thousand dollars. AI's decisive hardware arrived as a side effect of the games industry.

Hallucination

The general term for AI output that is fluent, confident, and not grounded in fact. Citation fabrication is one species of it.

The word is a slightly misleading metaphor — it suggests a malfunction, as though the system normally perceives reality and has briefly stopped. It hasn't. Generating plausible text is the only thing it ever does; hallucination is that same process producing something false. There is no internal alarm that distinguishes the two cases, which is why there is no warning light.

Large Language Model

The kind of system behind Claude, ChatGPT, and Gemini. It is trained on an enormous quantity of text to do one deceptively simple thing: predict what word is likely to come next.

Everything else — answering questions, writing essays, holding a conversation, apparently reasoning — emerges from that single operation repeated at extraordinary scale. Nothing in it stores facts the way a database does, which is why it can be fluent and wrong at the same time. Usually abbreviated LLM.

LGTM

Looks good to me — what a reviewer writes when approving someone else's code. Sometimes it follows a careful read. Sometimes it is a reflex.

The abbreviation has become shorthand for that gap. In Everybody's Lost Their Minds, Jan Schaumann calls the human in the loop "nothing more than a rubber stamp" and links those words straight to his own entry for lgtm: if the machine finds the bug, writes the patch, and reviews the pull request, the person signing off may be approving work nobody in the room understands. It is also the title of a Claude Opus 5.5 music video.

Machine Learning

Building software by showing it examples instead of writing down rules. You supply the data and a measure of what counts as doing well; the system adjusts itself until it does well. Nobody writes the rules it ends up using — which is why it can solve problems no one knows how to explain, and why it is hard to say what it has actually learned. The rival tradition is symbolic AI; see also connectionism and training data.

It is a subset of artificial intelligence and a tool of data science; it became practical when big data arrived. Distinguish it from data mining, which looks for patterns rather than building a model to reuse.

Mis-Citation

A source that exists, opens, and does not say what it was cited for.

It is the third rung of a ladder of citation failure, and the dangerous one:

• No source given. A claim with nothing attached. Easy to catch if you are watching for it.
• Fabricated source. The reference does not exist — the link is dead, the title returns nothing. Annoying, but it announces itself. See citation fabrication.
• Mis-citation. A real paper, by real authors, on roughly the right subject, attached to a number, date, or finding that is not in it.

The check most people actually perform is is there a link, and does it open? Mis-citation passes that test. The only thing that catches it is reading the source for the specific claim it was cited for.

Moravec's Paradox

The observation that tasks humans find hard — chess, calculus, formal logic — turn out to be comparatively easy for machines, while tasks any toddler can do — walking over uneven ground, picking up an egg, telling whether a joke landed — are extraordinarily difficult. Named for Hans Moravec (Mind Children, 1988). A useful test: when you hear a prediction about what AI will do next, ask whether a one-year-old could do it.

Neural Network

Software loosely modelled on the idea of connected brain cells: many simple units, each adding up the numbers coming in and passing a number on, arranged in layers. Nothing about an individual unit is intelligent. The behaviour lives in the strengths of the connections between them, which are adjusted automatically during training. The ancestor is the Perceptron (1957); stack enough layers and you have deep learning. See also Society of Mind — Minsky's argument that intelligence emerges from unintelligent parts is the same shape of idea.

Perceptron

The first machine that learned from examples rather than being programmed. Built by Frank Rosenblatt in 1957 — physical hardware, with motors and dials. It took several inputs, weighted them, added them up, and produced a single output, adjusting its own weights when it got an answer wrong. Every neural network in use today is a descendant.

Polanyi's Paradox

"We can know more than we can tell." — Michael Polanyi, 1966. Most of what we know how to do isn't available to us as rules: you can ride a bicycle, recognize a friend's face, and tell when a sentence sounds wrong, without being able to explain how. This is the deep reason Symbolic AI stalled, and the reason learning-from-examples had to be invented.

Project

A persistent workspace in Claude that holds project knowledge (documents uploaded once and visible in every conversation inside it) and custom instructions (standing directions about how it should work with you).

The practical difference: an ordinary conversation starts from nothing and is gone when you close it. A Project is the difference between asking a stranger the same question forty times and working with someone who has already read the brief.

Prompt

What you type to a chatbot: the question, instruction, or material that sets up the reply. Worth taking seriously as the unit of work in this course. The same question, asked with its context and constraints made explicit, comes back a different quality of answer — and unlike a hunch, a prompt is a record you can save, share, critique, and improve. That is why your assignments ask you to hand them in.

Qualitative Reasoning

The branch of AI concerned with reasoning about the physical world the way people do — in relationships rather than numbers. Drop something and it falls. Heavier things are harder to lift. A pebble in a pond makes ripples. Also called qualitative physics, and for decades a serious research programme, with applications in robotics and in simulating situations where precise numbers are not available.

The measure of how far the ground has moved: handing a current model a cartoon of a man dangling from a rope above sharks and asking what happens if he lets go — and getting a competent, sensible answer, sharks included. What was once a research agenda is now a party trick. Is it still AI? See AI effect.

Sampling Bias

A near-synonym for selection bias, used more narrowly for bias introduced by how the sample was drawn — as opposed to bias that creeps in afterwards, when people drop out or fail to respond. In practice the two terms are used loosely and often interchangeably; if someone draws the distinction sharply, they mean sampling bias is about the draw and selection bias is about the whole path from population to dataset.

Search Engine

A system that indexes documents that already exist and returns pointers to them. Every result is somewhere a human actually wrote something, which is why you can always click through and check.

Its failures are visible — burying the good answer on page four, ranking for advertisers — and it cannot summarise, compare, or reason across sources. That work is left to you. Contrast with chatbot.

Selection Bias

A term from data science and statistics. The umbrella term for what happens when the data you collected is not representative of the population you actually want to draw conclusions about — because of how it got selected.

The cleanest example is the one sitting in this room. If you poll this class about how it feels about AI, the answer will come back curious and broadly positive. That is not a finding about students, or about people. It is a finding about people who signed up for an AI course. The students most dismayed by AI, the ones who want nothing to do with it, are not in the dataset — they chose a different elective. The sample was filtered before a single word was collected.

Two named subtypes worth knowing: survivorship bias, where the filter is survival, and self-selection bias, where the filter is people choosing for themselves whether to take part.

It matters enormously for AI because no amount of clever modelling fixes it. If the selection was skewed, the model faithfully learns the skew, reports high accuracy on data drawn the same skewed way, and then fails the moment it meets the real population. See also training data.

The habit to build: before believing any number, ask who is missing from this dataset, and why?

Self-Selection Bias

Selection bias in which the units decide for themselves whether to be in the sample. Nobody filtered them; they filtered themselves.

Online reviews are the everyday case — they over-represent people who were delighted or furious and under-represent the large, quiet middle who thought it was fine. So are voluntary surveys, call-in polls, and the comments section. So is a class poll about AI in a course people elected to take.

It is especially slippery because the data can look abundant and healthy. Ten thousand reviews feels authoritative. It is still ten thousand people who chose to write one.

Sentiment Analysis

Automatically classifying a piece of text by the attitude it expresses — typically positive, negative, or neutral, sometimes on a finer scale or across specific emotions. It is one of the oldest practical applications of natural language processing, and one of the most widely deployed: companies run it over product reviews, social media, support tickets, and open-ended survey responses, because reading all of that by hand is impossible.

The methods have changed completely while the task stayed the same. Early systems counted words against a hand-built list of "positive" and "negative" terms. Then came classifiers trained on labelled examples — thousands of reviews with their star ratings attached. Today a large language model will do it without being trained on your data at all, just by being asked.

What makes it a good teaching example is how hard it turns out to be, in ways that are invisible from the outside. Sarcasm inverts meaning without changing a single positive word. Negation is easy to miss across a long sentence. "Unpredictable" is praise for a thriller and a complaint about a car. A review can be genuinely positive and negative at once. And the whole exercise compresses something rich and human into one number — which is exactly where the meaning tends to leak out.

Society of Mind

Marvin Minsky's 1986 argument that intelligence isn't one thing: the mind is built from many small processes, none of which is intelligent on its own, and intelligence is what the society of them produces together. The source of this course's name — and a preview of how modern AI works, since a neural network is also a very large number of individually mindless operations.

Survivorship Bias

Selection bias in which the filter is survival: you only get to examine the things that made it, so the things that didn't are invisible and silently excluded from your reasoning.

The classic case is from the Second World War. Analysts examining returning bombers wanted to add armour where the planes showed the most bullet holes. The statistician Abraham Wald pointed out the error: those were the planes that came back. The armour belonged where the returning planes had no holes at all, because hits there were the ones that brought a plane down. The sample had already been filtered by survival before anyone started counting.

Modern versions: studying successful companies to learn what makes companies succeed, without looking at the identically-run ones that failed. Or judging a technology by the projects that shipped.

Symbolic AI

The idea that intelligence is the manipulation of symbols according to explicit, human-written rules — so to build a mind, you write the rules down. Dominant from the 1950s through the 1980s. Also called GOFAI. Its rival is Connectionism.

Training Data

The examples a machine learning system learns from — and the single biggest determinant of what it will and won't be able to do.

A model cannot learn what isn't in its data, and it will learn everything that is, including the accidents. In 2016 a classifier that told huskies from wolves turned out to be detecting snow in the background. In 2018 a model reading chest X-rays was partly identifying which hospital took the scan. Both scored well. Both were useless.

"Garbage in, garbage out" is the old version. The modern version is subtler and worse: bias in, bias out, at enormous scale — and the model will report excellent accuracy while doing it. See selection bias.

Vibe Coding

Building software by describing what you want in ordinary language and letting an AI write the code — iterating on the result rather than on the source. The current top rung of an eighty-year ladder of abstraction that runs from rewiring ENIAC by hand, through assembly language, compilers, and libraries, to here. Coined by Andrej Karpathy in early 2025.

---