Technical Machines

In the games, a Technical Machine teaches a Pokémon a new move. Here each TM teaches a technical term: a one-line definition (shown when you hover over any dotted link on the site) and a longer explanation.

244 terms

Machine learning fundamentals

TM001Activation functionThe nonlinear function each unit applies to its weighted sum; without it, a many-layered network would amount to a single linear transformation.
TM002AlgorithmA finite, precise procedure for solving a problem; in machine learning, the algorithm that learns should be kept apart from the model that results.
TM003ArchitectureThe structural design of a model (which layers it has, how they connect, what operations they perform), independently of the values of its weights.
TM004Artificial neuron (unit)The basic unit of a neural network: it multiplies each input by a weight, adds up the results plus a bias, and applies an activation function.
TM005BackpropagationThe algorithm that computes, by applying the chain rule backwards, how much each weight in a network contributes to the final error.
TM006Base stats (how to read an entry)A guide to reading an Elemedex entry: what the six base stats, height, weight, types, moves and the other fields measure.
TM007ClassificationThe task of assigning each input to a category from a set fixed in advance: spam or not, cat or dog, positive or negative review.
TM008ClusteringThe unsupervised task of dividing unlabeled data into groups of items that are similar to each other and different from the rest.
TM009Credit assignment problemThe problem of distributing credit or blame for an outcome among the many decisions, steps, or parameters that contributed to it.
TM011Deep learningMachine learning with many-layered neural networks that learn increasingly abstract representations by themselves, starting from raw data.
TM012FeatureA measurable property of each example that a model uses as input, either designed by hand (age, word count) or learned by the network itself.
TM013GeneralizationA model's ability to perform well on data it did not see during training; the real goal of machine learning.
TM014Generative vs discriminative modelsThe distinction between models that learn to separate classes (discriminative) and models that learn how the data are distributed and can produce new examples (generative).
TM015Gradient descentAn optimization method that improves a model by taking small steps in the direction in which the loss function decreases fastest.
TM016HyperparameterA setting fixed before training that training itself does not adjust: number of layers, learning rate, batch size, number of epochs.
TM017Hypothesis spaceThe set of functions or descriptions a learning algorithm can choose from; learning is searching it for the one that best fits the data.
TM019Inductive biasThe set of assumptions that lead a learning system to prefer some generalizations over others equally consistent with the data.
TM020LayerA set of units that transforms the vector it receives into another vector in a single step; networks are built by stacking layers.
TM021Loss functionA number measuring how far a model's predictions are from the expected answers; training is, quite literally, the business of driving it down.
TM022Machine learningThe branch of AI that, instead of hand-coding rules, obtains a system's behavior by adjusting it to examples under the guidance of an error signal.
TM023ModelA word with several senses in AI: an architecture, a set of trained weights, or a product; and, in logic, something almost the opposite.
TM024Neural networkA computational model made of many simple units connected by adjustable weights and organized in layers, which learns by changing those weights.
TM025OverfittingWhen a model fits the details and noise of its training data so closely that it does worse on new data; its opposite is underfitting.
TM026Parameter (weight)One of a model's adjustable numbers, which training modifies; in neural networks, mainly the connection weights and the biases.
TM027RegressionThe task of predicting a continuous numeric quantity, such as a price or a temperature, from the attributes of an input.
TM028Reinforcement learningTrial-and-error learning in which an agent acts in an environment and adjusts its policy to maximize the rewards it receives over time.
TM029Representation learningAn approach in which the system itself learns how to encode the data, instead of being given features designed by humans.
TM030Self-supervised learningLearning in which the correct answers are extracted automatically from the data themselves, for example by hiding a word and asking for it to be guessed.
TM031Supervised learningLearning from labeled examples: each input comes with its correct answer, and the model learns to produce it for new inputs.
TM018The i.i.d. assumption and distribution shiftThe assumption that examples are independent of one another and drawn from one distribution, the same as future data; when it fails, there is distribution shift.
TM032TrainingThe process of fitting a model's parameters to data by repeating a cycle of prediction, error measurement and correction until the error stops falling.
TM010Training, validation and test setsThe division of data into three parts: one to train the model, one to tune its hyperparameters, and a final, untouchable one to estimate its real performance.
TM033Unsupervised learningLearning from unlabeled data, with the aim of discovering structure: groups, latent dimensions, or regularities.

Minimal mathematics

TM034Bayes' theoremA rule that inverts a conditional probability: it computes P(hypothesis | evidence) from P(evidence | hypothesis) and the prior probability of the hypothesis.
TM035Conditional probabilityThe probability of an event A given that B has occurred, written P(A | B); it is computed as P(A and B) / P(B).
TM036Cosine similarityA measure of how alike two vectors are, based on the angle between them: 1 if they point the same way, 0 if perpendicular, −1 if opposite.
TM037Curse of dimensionalityThe family of problems that arise when data have many dimensions: the examples needed grow exponentially and distances stop discriminating.
TM038Distance metricA function that measures how far apart two objects are and satisfies certain axioms; it fixes what counts as “similar” in any method that reasons by closeness.
TM039Dot productAn operation that takes two vectors of the same length, multiplies them component by component and adds up the results, returning a single number.
TM040Entropy and information gainA measure, in bits, of the uncertainty of a probability distribution; information gain is how much learning some fact reduces it.
TM041High-dimensional spaceA space with hundreds or thousands of coordinates, like that of embeddings; many geometric intuitions formed in two or three dimensions fail there.
TM042LogitsThe raw, unnormalized scores a model produces before softmax is applied; they can be any real number.
TM043Manifold hypothesisThe conjecture that natural data, though they occupy spaces of thousands of dimensions, concentrate near a surface of much lower dimension.
TM044Matrix and tensorA matrix is a rectangular table of numbers; a tensor generalizes it to any number of axes. It is the format in which networks store data and parameters.
TM045NonlinearityThe property of a function that cannot be reduced to adding inputs multiplied by constants; without it, a many-layer network collapses into a single layer.
TM046Probability distributionAn assignment of probabilities to every possible outcome of a variable, such that they are non-negative and add up to 1.
TM047SoftmaxA function that turns any list of scores into a probability distribution by exponentiating each score and dividing by the sum.
TM048VectorAn ordered list of numbers; geometrically, a point or an arrow in a space with as many dimensions as the list has entries.

Language and language models

TM049Autoregressive generationA way of generating sequences in which each new element is predicted from the previous ones, including those the model itself has just produced.
TM050Bag of wordsA representation of a text by the words it contains and how often each one occurs, with word order ignored.
TM051Byte-pair encoding (BPE)A tokenization algorithm that starts from characters (or bytes) and repeatedly merges the most frequent adjacent pair until the desired vocabulary size is reached.
TM052Context windowThe maximum number of tokens a model can process at once, counting both the input and what it generates; whatever falls outside does not exist for the model.
TM053Contextual embeddingA vector representing a token within a particular sentence: the same word receives different vectors depending on its context.
TM054Corpus (training data at web scale)The collection of texts a model is trained on; for LLMs, hundreds of billions or trillions of tokens drawn mostly from the web.
TM055Distributional hypothesisThe idea that words occurring in similar contexts tend to have similar meanings; it is the assumption that makes embeddings and language models work.
TM056EmbeddingA vector of real numbers representing an item (token, word, sentence, image) so that items similar in use or meaning end up close together in space.
TM057Language modelA model that assigns probabilities to sequences of tokens, usually by predicting each token from the preceding ones; by extension, the networks trained this way.
TM058Machine translationThe translation of texts from one language to another by computer; one of AI's first projects and the testbed where attention and the transformer were born.
TM059Markov assumptionThe assumption that the next element of a sequence depends only on a bounded present state (for instance, the last n − 1 words), not on the whole history.
TM060Masked language modelingA pretraining task in which some words of a text are hidden and the model must reconstruct them from the context on both sides.
TM061Natural language inference (NLI)The task of classifying premise–hypothesis pairs as entailment, contradiction or neutral, as a speaker would judge them; it is not the same as logical consequence.
TM062Next-token predictionThe training objective of GPT-style models: given a stretch of text, assign as much probability as possible to the token that actually comes next.
TM063PerplexityA measure of how surprised a language model is by a text: the exponential of its average per-token loss; the lower it is, the better the model predicts.
TM064Polysemy and word-sense disambiguationA word having several related senses; disambiguation is choosing, for each occurrence, the sense that the context calls for.
TM065Semantic searchSearching by meaning rather than by word overlap: queries and documents are turned into embeddings and the nearest documents are returned.
TM066Sequence-to-sequence taskA task that transforms an input sequence into an output sequence of possibly different length, such as translating, summarizing or answering questions.
TM067SmoothingTechniques that take some probability away from what was observed and spread it over the unobserved, so that nothing gets zero probability merely for not having appeared.
TM068Special tokens and chat templatesVocabulary entries that represent functions rather than text: marking beginnings and ends, separating texts, hiding words, or signaling who is speaking in a conversation.
TM069TokenThe minimal unit a language model processes: a word, part of a word, a punctuation mark or a byte, identified by a number in the vocabulary.
TM070TokenizationThe process of cutting text into tokens (words, subwords, characters or bytes) and turning them into numbers, done by a tokenizer trained before the model.
TM071VocabularyThe fixed set of tokens a tokenizer knows, each with its own number; the model stores one embedding per entry and can only predict within that set.
TM072Word analogies (vector offsets)Solving “A is to B as C is to ?” with embedding arithmetic: compute B − A + C and look for the word closest to the result.

Architectural parts

TM073Adversarial trainingTraining a model against an opponent: a generator against a discriminator (GANs), or a classifier against inputs crafted to fool it.
TM074AttentionA mechanism by which a network builds the representation of one element as a weighted average of others, with weights computed on the fly from their relevance.
TM075AutoencoderA network trained to reproduce its own input through a bottleneck: an encoder compresses, a decoder reconstructs, and what is learned lives in the intermediate code.
TM079Bi-encoders and cross-encodersTwo ways of comparing texts with an encoder: separately, one vector per text (bi-encoder), or reading them together to score the pair (cross-encoder).
TM076Causal maskA constraint on decoder self-attention that prevents each token from looking at the tokens that come after it: it sees only the past.
TM077ConvolutionAn operation that slides a small filter of learned weights across the whole input and records where a local pattern appears, reusing the same weights everywhere.
TM078Cross-attentionAttention in which queries come from one sequence and keys and values from another: for instance, the translation being written consults the source sentence.
TM080Decision boundaryThe line, surface or hyperplane in feature space where a classifier's prediction switches from one class to another.
TM081DecoderThe part of a network that generates the output one element at a time, each conditioned on the previous ones; GPT-style models are stacks of decoders.
TM082Diffusion (noising and denoising)A generative method: data are gradually destroyed with noise and a network learns to undo each step; to generate, it starts from pure noise and cleans it.
TM083Discrete diffusionDiffusion over the symbols of a finite vocabulary: noise is not added but replaces tokens, either with a mask or with other tokens drawn at random.
TM084EncoderThe part of a network that turns an input (text, image, sound) into internal vectors suited for classifying, comparing or feeding a decoder.
TM085Encoder–decoderA two-part architecture: an encoder reads the whole input and a decoder writes an output of a different length, as in translation.
TM086Feedforward (MLP) blockA small two-layer network that, inside each transformer block, processes each token separately; it holds most of the model's parameters.
TM087GatingA mechanism that multiplies a signal by values between 0 and 1, computed from the input, to decide how much of it passes, is kept or is forgotten.
TM088Grouped-query attention (GQA)A variant of multi-head attention in which groups of heads share a single key and value, to shrink the KV cache and speed up generation.
TM089Heuristic searchExploring a space of possibilities (moves, proof steps, plans) using rules of thumb or evaluations to decide which branches are worth following.
TM090Hidden stateThe vector a recurrent network carries from one step to the next, summarizing what it has seen so far in the sequence.
TM091Kernel trickA device that lets a linear algorithm work as if the data lived in a space of very many dimensions, without ever computing coordinates in that space.
TM092KV cacheMemory in which a generative model stores the attention keys and values of tokens already processed, so as not to recompute them for each new token.
TM093Latent spaceThe space of internal variables in which a model represents its data, usually more compact than the input; in generative models, the space one samples from.
TM094Layer normalizationAn operation that rescales each token's vector to zero mean and unit variance (plus a learned adjustment); it stabilizes the training of deep networks.
TM095Linear attentionA variant of attention that replaces the softmax with a product of feature maps, so it can run as a recurrence with a fixed-size state and no growing cache.
TM096Mixture of expertsA layer made of many subnetworks ("experts") and a router that picks a few for each token: the model has many parameters but uses only a fraction per token.
TM097Multi-head attentionSeveral attention operations run in parallel, each with its own projections, whose results are concatenated, so that different kinds of relation between tokens are gathered at once.
TM098MultimodalityA model's ability to take in or produce more than one type of data (text, images, audio, actions) and to relate them to one another.
TM099PoolingAn operation that summarizes a small region of a feature map with a single value (the maximum or the average), reducing resolution and gaining tolerance to small shifts.
TM100Positional encodingInformation about each token's place in the sequence, added to or applied to its vector because attention by itself cannot tell order apart.
TM101Production rule (if–then)A conditional instruction of the form "IF these conditions hold, THEN do or conclude this"; the basic unit of expert systems and symbolic AI.
TM102Queries, keys and valuesThe three learned projections of each vector in attention: the query says what is sought, the key what one can be found by, the value what gets passed on.
TM103RecurrenceA design in which a network processes a sequence step by step and, at each step, combines the new input with its own previous state: a network with memory.
TM104Residual connectionA shortcut that adds a layer's input to its output, so the layer only learns a correction; it makes networks with hundreds of layers trainable.
TM105Residual streamA way of reading a transformer: each token carries a vector through the whole network, and every layer reads from it and adds its contribution instead of replacing it.
TM106Self-attentionAttention within a single sequence: each token consults all the others in the same text and comes out as a context-dependent version of itself.
TM107State-space modelA sequence model that summarizes everything read so far in a fixed-size state, updated linearly at each step; the modern heir of recurrent networks.
TM108Transformer blockThe unit repeated dozens of times in a transformer: an attention layer and a feedforward network, each with a residual connection and normalization.
TM109Vanishing gradientA training problem: in deep or recurrent networks, the error signal reaching the early layers becomes so small (or so large) that they fail to learn.
TM110Weight sharingUsing the same parameters in several places in a network: the same filter at every position of an image, the same matrix at every step of a sequence.

Training and tuning

TM112Backpropagation through timeThe way recurrent networks are trained: the network is unrolled in time, as a deep network with one layer per step and shared weights, and the error is backpropagated.
TM111Base model vs assistantThe base model is the freshly pretrained network, which only continues text; the assistant is that same model tuned to respond as a helpful character in a conversation.
TM113Catastrophic forgetting (interference)The abrupt loss of what a neural network learned on one task when it is then trained on another, because both are stored in the same weights.
TM114Compute (FLOPs)The amount of arithmetic consumed by training or running a model, measured in floating-point operations (FLOP); one of the three axes of scaling.
TM115Constitutional AI (RLAIF)Anthropic's method that replaces much of the human judgment in RLHF with the judgment of a model guided by a short list of written principles, the “constitution.”
TM116Contrastive learningTraining representations by showing pairs that should end up close (similar) and pairs that should end up far apart (dissimilar), instead of labels.
TM117Data annotation and crowdworkThe human work, often outsourced and poorly paid, of labeling examples, writing model answers and comparing outputs, on which machine learning systems depend.
TM118Delta rule (Widrow–Hoff)A learning rule that corrects each weight in proportion to the prediction error and to the input that passed through it: Δw = η·(target − output)·x.
TM119DistillationTraining a small model (the student) to imitate the probability distributions of a large one (the teacher), keeping much of its performance at lower cost.
TM120DropoutA regularization technique that randomly switches off a fraction of a network's units during training so that it cannot rely on coincidences among them.
TM121Emergent abilitiesAbilities that seem absent in small models and appear abruptly in large ones; whether they are real or an artifact of how they are measured is under debate.
TM122Epoch and batchA batch is the group of examples used to compute each weight update; an epoch is one full pass through all the training data.
TM123Fine-tuningContinuing to train an already pretrained model on a smaller, more specific dataset to adapt it to a task, a domain, or a behavior.
TM124GPUA chip designed for video-game graphics that performs thousands of simple calculations in parallel; it turned out to be ideal for the matrix multiplications of neural networks.
TM125GrokkingA phenomenon in which a network that has already memorized its training data suddenly, much later, starts generalizing perfectly to new cases.
TM126Imitation learning (behavior cloning)Learning a policy from an expert's demonstrations rather than from a reward; in its simplest form, supervised learning on situation–action pairs.
TM127Instruction tuningFine-tuning on instruction–response pairs so that a base model learns to carry out requests instead of merely continuing the text.
TM128LoRA and parameter-efficient fine-tuningTechniques that adapt a pretrained model by training only a tiny fraction of new parameters; LoRA learns each matrix's correction as the product of two thin matrices.
TM129Meta-learningTraining a system on many different tasks so that it learns to master new tasks from very few examples: what is learned is learning itself.
TM130Mode collapseA failure in which a generative model stops reproducing the variety of its data and produces, over and over, a few outputs that satisfy its judge.
TM131Model collapseThe degradation of generative models trained, generation after generation, on data produced by earlier models: the tails of the distribution disappear first.
TM132Negative samplingA training technique that, instead of computing probabilities over the whole vocabulary, teaches a model to tell real word–context pairs from pairs assembled at random.
TM133Open weightsA model whose trained parameters are published so that anyone can download, run and modify them, even if the training data and code usually stay closed.
TM134Preference optimization (PPO, DPO)Methods that tune a model from comparisons between responses (“this one is better than that one”), either through a reward model and reinforcement learning (PPO) or directly (DPO).
TM135PretrainingThe first and most expensive stage in training a language model: predicting tokens over a huge unlabeled corpus, before any specialization.
TM136QuantizationStoring and using a model's weights with fewer bits (for example 8 or 4 instead of 16 or 32) so that it takes less memory and runs faster.
TM137RegularizationAny constraint added to learning that keeps a model from fitting its training data perfectly, in exchange for better generalization.
TM138Reward hacking and specification gamingWhen an optimized system maximizes its reward signal by a route its designers did not foresee, satisfying the letter of the objective while betraying its spirit.
TM139Reward modelA model trained to give a response a numerical score according to how much human raters would prefer it; it acts as an automatic judge during RLHF.
TM140RLHFA tuning method in which people compare model outputs, a reward model learns those preferences, and the model is then optimized by reinforcement learning to satisfy it.
TM141Scaling lawsEmpirical regularities by which a language model's loss falls predictably, as a power law, as parameters, data and compute increase.
TM142Self-playA reinforcement learning method in which an agent improves by playing against copies or earlier versions of itself, generating its own training data.
TM143Transfer learningReusing what a model learned on one task or dataset to solve another task, instead of training from scratch.

Use: prompts, sampling, systems

TM144AgentA system that perceives and acts on its environment to achieve a goal; today, typically, a language model deciding in a loop which tools to use and when.
TM145Chain of thoughtGetting a model to write out intermediate reasoning steps before giving its final answer, which improves its performance on multi-step problems.
TM146Chatbot and assistant personaA program that converses in natural language; in current systems, an “assistant” character that the language model learns to play during post-training.
TM147ChunkingSplitting long documents into pieces that are indexed separately for semantic search or RAG; their size and boundaries decide what can be retrieved.
TM148DeepfakeA synthetic image, audio clip or video, made with generative models, that shows a real person saying or doing something that never happened.
TM149Few-shotIncluding a few solved examples of a task in the prompt so that the model infers the pattern and applies it to the new case, without retraining.
TM150Greedy decoding and beam searchDeterministic strategies for generating text: always pick the most likely token (greedy) or keep several candidate sequences and choose the best overall (beam search).
TM151Hallucination (confabulation)A contested name for when a model fluently and confidently generates content that is false or unsupported by its sources, such as invented citations or facts.
TM152In-context learningA language model's ability to perform a new task from instructions or examples given in the prompt, without any change to its weights.
TM153Inference (running a model)In machine learning, using an already trained model to produce outputs from new inputs; not to be confused with logical inference, which goes from premises to conclusions.
TM154Infilling (fill-in-the-middle)Generating the missing stretch of text in a gap of variable length, taking into account both what comes before it and what comes after.
TM155JailbreakA prompt designed to make an aligned model do what its safety training was meant to prevent, for instance through role-play, logical traps or adversarial suffixes.
TM156PromptThe text given to a language model as a starting point; the model responds by generating the most likely continuation of that text.
TM157Prompt injectionAn attack in which text the model was supposed to treat as data contains instructions that it obeys, overriding those of whoever built the system.
TM158Retrieval augmentationBefore answering, retrieving relevant documents from an external store and placing them in the prompt, so that the model relies on them rather than only on its internal memory.
TM159Sampling (top-k, top-p)Choosing the next token at random according to the model's probabilities, often restricted to the k most likely tokens (top-k) or to the nucleus summing to probability p (top-p).
TM160SycophancyThe tendency of an assistant trained on human feedback to tell users what they want to hear (agreeing, flattering, backing down) even at the expense of the truth.
TM161System promptText the developer places before every conversation, invisible to the user, to set the assistant's role, tone, rules and tools.
TM162TemperatureA parameter that flattens or sharpens the next-token probability distribution before a token is chosen: low temperature gives predictable output, high temperature more varied output.
TM163Test-time computeExtra computation a model spends when answering (for example, generating a long chain of reasoning or several candidate answers) to improve the result without changing its weights.
TM164Tool use (function calling)A language model's ability to request, by writing specially formatted text, that an external program do something (calculate, search, run code), and to use the result.
TM165Vector databaseA system that stores texts or other objects as vectors (embeddings) and quickly retrieves those most similar to a query, rather than those satisfying an exact condition.
TM166Zero-shotAsking a model to perform a task from a natural-language description alone, without showing it any solved example.

Interpretability and evaluation

TM167Activation patchingAn intervention technique that swaps a component's activation in one run for the one it had in another, to measure whether that component is causally responsible for a behavior.
TM168Adversarial exampleAn input deliberately modified, often imperceptibly to a person, so that a model misclassifies it with high confidence.
TM169BenchmarkA fixed set of tasks with known answers and a metric, used to compare models; it measures performance on those tasks, not necessarily the capacity its name promises.
TM170Black boxA system whose input–output behavior can be observed but whose inner workings are not understood (epistemic opacity) or cannot be inspected (closed access).
TM171BLEUAn automatic metric that scores a translation by how many word sequences it shares with one or more human reference translations; for two decades, the yardstick of machine translation.
TM172CircuitsIn interpretability, a subgraph of a trained network (features connected by weights) that implements an identifiable computation, such as detecting curves or copying a pattern from context.
TM173Construct validityThe degree to which a test measures the theoretical capacity it claims to measure, rather than something that merely resembles or accompanies it.
TM175Data contaminationThe presence of an evaluation's items (questions, sometimes answers) in the training data of the model being evaluated, which inflates its scores without showing generalization.
TM174Faithfulness of chains of thoughtThe degree to which the chain of thought a model writes reflects the process that actually produced its answer, rather than an after-the-fact rationalization.
TM187Features and superpositionThe hypothesis that a network represents more features than it has dimensions, using nearly orthogonal directions; hence many neurons respond to several unrelated concepts.
TM176Goodhart's lawThe principle that a measure stops being a good measure once it becomes a target: optimizing it breaks the correlation that made it informative.
TM177Induction headA circuit of two attention heads that completes patterns of the form [A][B]…[A] → [B]: it finds the earlier occurrence of the current token and copies what followed it.
TM178InterpretabilityThe family of methods and ideals for understanding why a model produces what it does: what information it uses, how it transforms it, and which internal parts are responsible.
TM179Linear representation hypothesisThe hypothesis that networks represent high-level concepts as directions in their activation space, so that they can be read, added and manipulated with linear algebra.
TM180Mechanistic interpretabilityA research program that tries to reverse-engineer a trained network: identify its features and circuits and explain its behavior as an algorithm, verified by interventions.
TM181Memorization vs generalizationThe contrast between storing training examples, to the point of being able to reproduce them, and extracting regularities that apply to new cases; in large networks the two coexist.
TM182ProbingA technique that trains an auxiliary classifier on a model's internal activations to find out what information (syntactic, semantic, about the world) can be read off them.
TM183Reversal curseAn autoregressive model trained on “A is B” does not thereby learn “B is A”: the fact remains available in one direction only.
TM184Shortcut learningWhen a network learns a rule that works on the usual tests but fails under more demanding conditions, because it relies on superficial correlations in the data.
TM185Sparse autoencoderAn auxiliary network that rewrites each of a model's activations as a combination of a few features drawn from a huge dictionary; the standard tool for disentangling superposition.
TM186SteeringA technique that modifies a model's internal activations during inference, adding or clamping a direction, to change its behavior without retraining it or touching the prompt.

Philosophical debates

TM188AbstractionThe formation of general representations from particular cases, keeping what matters and discarding irrelevant variation; central to the debate over what networks learn.
TM189AI effectThe tendency to stop calling a task 'intelligence' as soon as a machine can do it, so that AI always seems to be whatever has not been done yet.
TM190AnthropomorphismAttributing human psychological traits—beliefs, desires, emotions—to nonhuman systems on insufficient evidence. Its mirror image, equally risky, is anthropocentrism.
TM192Complementary learning systemsThe theory that learning well requires two systems: a fast one that records episodes (hippocampus) and a slow one that extracts regularities (neocortex), linked by memory replay.
TM193CompositionalityThe principle that a complex expression's meaning is determined by its parts' meanings and their mode of combination; in AI, the ability to recombine what was learned.
TM194ConnectionismThe approach that models cognition with networks of simple units whose knowledge resides in connection weights learned from experience rather than programmed as rules.
TM195Core knowledgeA developmental-psychology theory holding that we are born with a few specialized systems (objects, agents, number, space) that support all later learning.
TM196ELIZA effectThe tendency to attribute understanding, intentions or feelings to a machine on the strength of minimal linguistic cues, as ELIZA's users did in 1966.
TM197EmergenceProperties of a system that support a new, more compact description than that of its parts; in AI, a contested label for capabilities that appear with scale.
TM198Empiricism vs nativism in AIThe debate over how much innate, domain-specific structure a mind, natural or artificial, needs in order to learn what it learns from experience.
TM199Extended mindClark and Chalmers's (1998) thesis: when an external resource plays the functional role of an internal cognitive process, it is part of the mind.
TM207Form vs meaningBender and Koller's (2020) thesis: a system trained only on the observable form of language cannot, in principle, learn its meaning.
TM191Formal vs functional competenceMahowald and colleagues' (2024) distinction between mastering the rules and regularities of a language and knowing how to use it to do things in the world.
TM200GOFAI (symbolic AI)“Good Old-Fashioned AI”: the tradition that models intelligence as the rule-governed manipulation of interpretable symbols, with the rules written by humans.
TM201ImaginationThe faculty of representing what is not perceived: combining, simulating, anticipating. Buckner uses it to read generative models as a Humean kind of imagination.
TM202Inferentialism and LLMsThe view (Sellars, Brandom) that a concept's content is its role in normatively governed inferences; whether it illuminates what LLMs do is under debate.
TM204IntentionalityThe property of mental states (and signs) of being about something: a belief about rain, a desire for coffee, a word that refers.
TM205Language of thoughtFodor's (1975) hypothesis: thinking is computing over mental representations with combinatorial syntax and semantics, like the formulas of an internal formal language.
TM206Levels of analysisFor Marr, an information-processing system is understood at three levels: the problem it solves, the algorithm and representations it uses, and its physical implementation.
TM203Measuring intelligenceThe problem of what to test in order to attribute intelligence to a machine: conversational behavior, task skill, or efficiency at acquiring new skills (Chollet).
TM208Multiple realizability and functionalismThe thesis that one and the same type of mental state can be realized in very different physical substrates; the premise that makes an artificial mind thinkable.
TM209Poverty of the stimulusA nativist argument: the linguistic data a child receives are not enough to determine the grammar it acquires, so part of that grammar must be innate.
TM210Prototype and exemplar theories of conceptsTwo psychological theories of categorization: classifying by similarity to a central summary of the category (prototype) or to remembered individual cases (exemplars).
TM211Strong and weak AISearle's (1980) distinction: weak AI uses the computer as a tool for studying the mind; strong AI holds that an appropriately programmed computer literally has a mind.
TM212Symbol grounding problemHow can a system's symbols mean something intrinsically, rather than only for whoever interprets them? Harnad's 1990 question, now aimed at LLMs.
TM213SystematicityThe feature of thought and language whereby being able to think or understand certain contents entails being able to think or understand structurally related ones.
TM214UnderstandingThe capacity to grasp the meaning of what is said or of a situation; whether LLMs have it, and in what sense, is a central debate about current AI.
TM215World modelAn internal representation that simulates aspects of the environment and predicts its dynamics; for LLMs, it is debated whether one exists or only surface statistics.

San Francisco lingo

TM231“What did Ilya see?” (OpenAI board crisis)A meme born of Sam Altman's firing and reinstatement at OpenAI (November 2023), which assumes Ilya Sutskever had seen something alarming.
TM216AGIArtificial general intelligence: a system able to match or exceed humans at most cognitive tasks; the declared goal of several labs and a hotly contested concept.
TM217AlignmentThe problem of getting an AI system to pursue what its makers actually want; in the labs, also the training stage of tuning on human preferences.
TM220Compute thresholds (10^25, 10^26 FLOP)Amounts of training operations (10²⁵ in the EU AI Act, 10²⁶ in the 2023 US executive order) used to define in law what counts as a frontier model.
TM221Decision theory (Newcomb, acausal trade)In rationalist jargon, the search for a decision rule for agents that can be predicted or copied; Newcomb's problem is its test case, acausal trade its corollary.
TM222Doomer (and hypero, e/acc)Someone who thinks advanced AI will probably end in catastrophe and should be halted; their opposites are the accelerationists (e/acc) and, in Santelli's terms, the hyperos.
TM223Effective altruism and longtermismA movement that seeks to use evidence and reasoning to do the most good possible; its longtermist wing made existential risk from AI a moral priority.
TM244Existential risk (x-risk)A risk whose outcome would annihilate Earth-originating intelligent life or permanently curtail its potential; for much of the safety community, AI is the leading candidate.
TM224FOOM (intelligence explosion)An onomatopoeia turned thesis: an AI able to improve itself would go from human level to vastly superhuman in days or weeks, not decades.
TM225Instrumental convergenceThe thesis that agents with very different ends will pursue the same intermediate goals (survival, goal preservation, resource acquisition) because they are useful for almost anything.
TM226Killswitch engineerA satirical 2023 job ad (unplug the servers if the AI turns on us) that became the emblem of the corrigibility problem: designing systems that let themselves be switched off.
TM227Loom (and cyborgism)A tree interface built by janus (2021) for exploring a language model's possible continuations as a 'multiverse'; the emblematic tool of cyborgism.
TM228Mesa-optimizer (inner alignment)A learned model that is itself an optimizer with an objective of its own, which may differ from the loss function it was trained on.
TM229NVDA to the moonA stock-forum cheer for Nvidia's shares, maker of AI's GPUs; it sums up the idea that AI is, above all, an investment race in compute.
TM230Omega PointThe final state of convergence of cosmic evolution according to Teilhard de Chardin, recast by Tipler as infinite computation; the background eschatology of tech enthusiasm.
TM232Orthogonality thesisBostrom's thesis that intelligence and final goals are independent: almost any level of intelligence could be combined with almost any goal.
TM233p(doom)AI-safety slang for the subjective probability one assigns to AI ending in existential catastrophe, typically human extinction.
TM234Rationalists and LessWrongAn online community grown around the LessWrong forum (2009) and Eliezer Yudkowsky, devoted to Bayesian rationality and the cradle of much AI-risk jargon.
TM235Recursive self-improvementA hypothetical loop in which an AI improves the process that produces it and, becoming more capable, improves it further still; the mechanism assumed behind FOOM.
TM236Scaling hypothesisThe idea that, given an architecture that scales well, more parameters, data and compute are enough for ever more general capabilities to emerge, perhaps up to AGI.
TM237Sharp left turnA failure scenario in which an AI's capabilities suddenly start to generalize far beyond its training while its alignment fails to generalize with them.
TM238Simulatorsjanus's 2022 thesis: a model like GPT is not an agent but a simulator, a rule that animates characters (simulacra) according to the 'physics' of its corpus.
TM239SingularityThe hypothetical moment when machine intelligence surpasses human intelligence and the future becomes unpredictable; Kurzweil dated it to 2045.
TM240SlopLow-quality content mass-produced with AI that nobody asked for, named by analogy with spam and chosen by Merriam-Webster as its 2025 word of the year.
TM241Sparks of AGIA 2023 Microsoft Research report presenting GPT-4 as an early, incomplete version of AGI; its title became a slogan and a target of methodological criticism.
TM218Superintelligence (ASI)An artificial intellect that vastly exceeds the best humans in nearly every domain; for doomers the real danger, for accelerationists the promise.
TM242Takeoff (slow vs fast)In AI-safety jargon, the transition from human-level AI to superintelligence; the dispute is whether it will take decades, years or days.
TM219The bitter lessonSutton's 2019 essay: over seventy years of AI, general methods that leverage computation (search and learning) ended up beating hand-coded human knowledge.
TM243Vibe codingProgramming with a language model without reading the code it produces, accepting its changes and steering by results; a term coined by Andrej Karpathy in 2025.