Glossary

All technical terms of the chronicle in alphabetical order. The number at the end names the section in which the term is introduced.

A

• Activation function – Rule that turns the weighted sum of a neuron’s inputs into its output value. 3
• Agents – Language models that act with tools and complete multi-step tasks in a loop of planning, acting and checking. 7
• AGI – Artificial general intelligence, an AI as versatile as a human. Whether it is attainable is disputed. 8
• AI effect – As soon as a method works reliably, it no longer counts as AI but as software. Introduction
• AlexNet – Convolutional network with eight layers that won the ImageNet competition in 2012 and marks the breakthrough of deep learning. 5
• Alignment – Aligning a model with what people want from it rather than with what is most probable in the training material. 6
• AlphaFold 2 – DeepMind model that in 2020 predicts the three-dimensional shape of proteins from their chain of building blocks. 6
• AlphaGo – DeepMind program that beat the Go professional Lee Sedol in 2016, through search plus two deep networks. 5
• AND, OR, NOT – The three basic building blocks of logic from which every computer calculation is made. 1
• Architecture – Blueprint of a network: which layers it has and how they are connected. 6
• Artificial neuron – Mathematical imitation of a nerve cell that sums its inputs and fires when a threshold is reached. 1
• Associative memory – Hopfield’s network (1982), which stores patterns as wholes and completes them from fragments. 3
• Attention – Method by which a network, for every word it generates, looks back at all words of the input with weights. 5
• Autonomous weapons systems – Weapons that select and engage their target without human release; subject of negotiations at the United Nations since 2014. 8

B

• Backpropagation – Method that traces a network’s error backwards through all layers and assigns each weight its share. 3
• Bayesian networks – Models that represent dependencies between things as probabilities. 4
• Benchmarks – Standardised tests used to compare models. 7
• Black box – A deep network whose reasons for a decision nobody can name precisely after training. 5

C

• Certainty factor – Value between 0 and 1 with which MYCIN rated the reliability of a rule. 3
• Chain of thought – A model is asked to think step by step before answering. 7
• Chatbot – A program one converses with in writing. 2
• ChatGPT – Chat interface on a GPT model, released in November 2022. 7
• Chinese room – Searle’s thought experiment (1980): someone who manipulates Chinese characters only by a rulebook understands no Chinese; a program, he argued, does nothing else. 3
• Class – Category into which an example is sorted in classification. 2
• Classification – Sorting examples into classes. 2
• CLIP – OpenAI model (2021) that places images and texts in the same vector space. 6
• Combinatorial explosion – Multiplication of possibilities with every step, as with chess moves. 2
• Context window – Number of tokens a model can take into account at once. 7
• Convolutional neural network (CNN) – Network that scans images for patterns with travelling filters. 3

D

• Decision tree – Sequence of yes/no questions leading to a decision. 4
• Decoder – Part of a network that generates text word by word from an internal representation. 5
• Deep learning – Learning in deep networks with many layers that build their features in stages by themselves. Term coined by Hinton, 2006. 4
• Deepfakes – Faked images, voices or videos of real people, produced by generative models. 7
• DENDRAL – First expert system (from 1965), which derived molecular structures from spectrometer data. 3
• Descriptors – Hand-made image features such as SIFT and SURF that describe edges and corners as lists of numbers. 4
• Diffusion models – Image generators that learn to denoise progressively noised images step by step. 7
• Distillation – Training a small model to imitate the answers of a large one. 7
• DQN – DeepMind network (2013) that learned Atari games solely from screen pixels and the score. 5
• Dropout – Randomly switching off neurons during training so that the network cannot rely on individual ones. 5

E

• ELIZA – Early chatbot by Weizenbaum (1966) that answered with keywords and sentence patterns. 2
• Emergent abilities – Abilities that small models lack and that appear by themselves beyond a certain size. 6
• Encoder – Part of a network that reads an input and condenses it into an internal representation. 5
• Epoch – One complete pass through all training examples. 5
• Error – Number measuring how far a network’s output deviates from the correct answer. 3
• Expert system – Program that decides in a narrow field with rules, like a human expert. 3

F

• Factual knowledge – Textbook knowledge of a field, as opposed to heuristic knowledge from experience. 3
• Feature engineering – Human-devised computation of features from raw data. 4
• Features – Properties of the data useful for a task, such as edges in images. 3
• Filters – Small pattern templates that travel across the image in a convolutional network. 3
• Fine-tuning – Adapting a pre-trained model to a specific task with a few examples. 6
• Forward chaining – Reasoning from known facts by applying all matching rules. 3
• Foundation model – A model pre-trained once on which many applications are built. Term from 2021. 6
• Frames – Data structures after Minsky that bundle all properties of a concept and inherit from broader concepts. 3
• Fuzzy logic – Logic with gradations between true and false, by Zadeh, 1965. 2

G

• GANs – Generative adversarial networks: a forger network and an inspector network compete. 5
• General Problem Solver – Program by Newell, Shaw and Simon (1957) meant to solve any problem by breaking it into sub-goals. 2
• Generalisation – A model’s ability to classify examples it has never seen correctly. 2
• Generative AI – Models that create new content instead of merely classifying what exists. 7
• GPU – Graphics processor with thousands of simple arithmetic units that greatly speeds up training neural networks. 4
• Gradient – For each weight, the direction and strength in which the error changes. 3
• Gradient descent – Moving the weights step by step in the direction that reduces the error. 3

H

• Hallucinations – Convincing-sounding but false statements by a language model. 7
• Heuristic knowledge – Rules of experience an expert acquires only through years of practice. 3
• Heuristics – Rules of thumb that weed out hopeless possibilities early. 1
• Hidden layers – Layers between input and output whose outputs nobody can specify. 3
• Hidden Markov models – Statistical models for sequences of hidden states visible only through their effects; the basis of speech recognition until 2012. 4

I

• ImageNet – Database of millions of hand-labelled photos (2009) and the associated competition. 4
• In-context learning – A model continues a pattern shown in the prompt without its weights changing. 6
• Inference – Use of a trained model, as opposed to training. 7
• Inference engine – Part of an expert system that chains rules and draws conclusions. 3
• Instruction tuning – Post-training on examples in which an instruction is followed by the desired answer. 7
• Intelligence explosion – Good’s idea of 1965: a machine cleverer than its builders builds still cleverer machines. 7

J

• Jailbreak – Prompt that circumvents a model’s trained-in limits. 7

K

• Knowledge acquisition bottleneck – Specialists can put only part of their knowledge into rules. 3
• Knowledge base – Collection of the if-then rules of an expert system. 3

L

• Large language model (LLM) – Language model with billions of parameters. 6
• Layers – Groups of neurons arranged one behind the other; the outputs of one layer are the inputs of the next. 2
• Learning rate – Step size in gradient descent, say 0.1: every weight is changed by a tenth of its gradient. 3
• Linearly separable – Task in which a single straight line separates all examples of one class from the other. 2
• LISP – Programming language by McCarthy (1958), for decades the language of AI research. 2
• Logic Theorist – Program by Newell, Shaw and Simon (1956) that proved mathematical theorems. 1
• LSTM – Long short-term memory: recurrent network with a memory cell and gates, by Hochreiter and Schmidhuber, 1997. 4

M

• Machine learning – Umbrella term for programs that improve with experience instead of being programmed. Term coined by Samuel, 1959. 2
• Mixture of experts – Model made of many sub-networks, of which only a few compute per token. 7
• Model – Calculation rule with finished weights, the result of training. 2
• Model collapse – Loss of rare patterns when a model is trained on the output of its predecessors. 7
• Monte Carlo tree search – Search procedure that plays games to the end at random and follows up the moves that pay off most often; AlphaGo’s second building block. 5
• Moravec’s paradox – What people find hard (chess, algebra) is easy for computers; what every child masters (seeing, grasping, walking) is hard for computers. 3
• Multimodal models – Models that process images, sound and video as well as text. 7

N

• Nearest-neighbour algorithm – Method that compares a new example with the most similar known ones. 4
• Neural network – Many connected artificial neurons in which the output of one is the input of the next. 1
• Neurons – Nerve cells that receive signals and fire themselves when stimulated enough. 1
• Noise – Change of every pixel by a random amount; after many steps only a shapeless scatter of light and dark dots remains, which diffusion models learn to remove step by step. 7
• NP-completeness – Theory (around 1971) according to which whole classes of tasks cannot be solved in foreseeable time. 2

O

• Open-weights models – Models whose weights anyone can download and run on their own hardware. 7
• Overfitting – A network memorises its training examples instead of grasping the general pattern. 5

P

• Parameters – The adjustable numbers of a model, above all its weights. 6
• Perceptron – Artificial neuron after Rosenblatt (1958) that sets its own weights. 2
• Pixels – Picture elements of a digital image, each with a brightness value. 3
• Pooling layer – Layer that shrinks an image by keeping only the strongest value from a neighbourhood of pixels. 3
• Pre-training – Learning general language understanding from vast amounts of text, before fine-tuning. 6
• Prompt – The input text given to a language model. 6
• Prompt injection – Hidden instructions in content that an agent reads and wrongly carries out. 7

Q

• Quantisation – Storing weights with fewer decimal places so that models become smaller and faster. 7

R

• RAG – Retrieval-augmented generation: the model answers on the basis of documents retrieved beforehand. 7
• Random forest – Many different decision trees that vote. 4
• Reasoning models – Models that work through intermediate steps before answering. 7
• Recommender systems – Systems that predict from past behaviour which video, product or post will please next. 5
• Recurrent neural network (RNN) – Network that reads a sequence element by element while carrying a memory. 4
• Reinforcement learning – Learning from reward and penalty instead of correct answers; the system tries out actions. 5
• ReLU – Activation function that outputs negative sums as 0 and positive ones unchanged. 5
• Residual connections – Shortcuts through which the signal can skip layers; make very deep networks trainable. 5
• RLHF – Reinforcement learning from human feedback: humans rate answers, the model learns to prefer well-rated ones. 7
• Robot – Artificial machine that does work otherwise done by humans; the word comes from Čapek’s “R.U.R.” (1920). 1

S

• Search procedure – Searching a tree of possibilities for a path to the goal. 2
• Self-attention – Every word of a text weighs how important every other word of the same text is for understanding it. 6
• Self-supervised learning – The text supplies the correct answer itself, for example by predicting masked words. 6
• Sequence-to-sequence – An encoder reads a sequence, a decoder generates another from it, such as a translation. 5
• Sequences – Data in which the order matters, such as text or speech. 4
• Scaling laws – Performance rises predictably with model size, data volume and computing power. 6
• Shakey – First mobile robot that planned its own actions (SRI, 1966–1972); origin of the A* search algorithm. 2
• Singularity – The point, so named by Vinge in 1993, at which machines surpass human intelligence and development escapes prediction. 7
• Stored program – Design principle in which the program resides in memory like data rather than in the wiring. 1
• Strong AI – A machine that genuinely understands and has consciousness. 3
• Supervised learning – Learning from examples for which the correct answer is supplied. 2
• Support vector machines – Method that seeks the dividing line with the largest distance to both classes. 4
• Symbol – A sign that stands for something, such as a word or a chess piece. 2
• Symbolic AI – Approach that reproduces thinking as manipulating symbols according to rules. 2
• Synthetic data – Training data that models generate themselves. 7
• System prompt – Invisible instruction from the provider placed before every conversation. 7

T

• Temperature – Setting that determines whether a model always picks the most probable token or also less probable ones. 7
• Test data – Held-back examples used to check whether a model classifies unfamiliar examples correctly. 5
• Threshold – Limit at which an artificial neuron outputs a 1. 1
• Tokenizer – Method that splits text into tokens, learned from the most frequent character sequences. 6
• Tokens – Words or parts of words that a language model computes with. 6
• Tools – Instruments such as web search or code execution that an agent can use. 7
• Training – Practising on examples while the weights are adjusted step by step. 2
• Training data – The collection of examples from which a model learns. 2
• Transfer learning – Using a pre-trained network as the starting point for a new task. 5
• Transformer – Architecture of 2017 based entirely on self-attention that processes all words at once. 6
• Turing machine – Imaginary device of 1936 that can compute anything computable. 1
• Turing test – A person chats with a stranger; if human and machine cannot be told apart, the machine counts as intelligent. 1

U

• Unsupervised learning – The method receives only data and searches for structure in it by itself. 4

V

• Vanishing gradient problem – The error information shrinks on its way through many layers, so that front layers do not learn. 3
• Vector – List of numbers, such as the representation of a word in the space of word embeddings. 5
• Vector database – Store for documents as vectors in which passages close in content can be found. 7
• Vision-language-action model – Language model that reads camera images and instructions and produces movement commands for a robot from them (RT-2, 2023). 8

W

• Weak AI – Simulation of intelligent behaviour for particular tasks; applies to every system so far. 3
• Weight – Number on a connection by which the incoming signal is multiplied; learning means changing weights. 1
• Word embeddings – Representation of words as vectors in which related words lie close together. 5
• World models – Models that learn space, time and physics instead of only text. 8

X

• XOR – Exclusive or: true when exactly one of two inputs is true; not linearly separable. 2