Glossary

Switching off part of a model, for example setting an attention head's output to zero, to see which behaviors depend on it.

Taught in How circuits form: induction headsRelated: Attention head, Circuit

Quantization that sets the scale from the largest magnitude in a block, so that value maps to the largest integer. Levels are symmetric around zero.

Taught in Quantization: smaller numbersRelated: Quantization, Zero-point quantization, Group-wise quantization

A list of uses a model's provider forbids, such as fraud, weapons, or passing off outputs as human, which its license or terms make binding on users.

Taught in Open and closed modelsRelated: Model license, Open weights

The return an agent can expect from taking a given action in a given state and acting well afterwards, written Q(s, a). The best action is the one with the highest Q.

Taught in Learning from experience: Q-learningRelated: Q-learning, Value function

The function a neuron applies to its weighted sum, such as the sigmoid, tanh, or ReLU. Without it, stacking layers of neurons would add no power.

Taught in The artificial neuronRelated: Artificial neuron, Sigmoid

Running a model on one prompt while copying in an internal activation from a run on another prompt, to test whether that activation carries the information that changes the output.

Taught in Switching parts off and steeringRelated: Ablation, Residual stream, Interpretability

A boosting method that keeps a weight on every training point, raises the weights of points the latest weak model got wrong, and gives each model a say based on its weighted error.

Taught in Ensembles: forests and boostingRelated: Boosting, Decision stump

A widely used optimizer that combines momentum with a separate, automatically scaled step size for each parameter. Most large neural networks are trained with Adam or a close variant.

Taught in Better optimizersRelated: Optimizer, Momentum

An input changed on purpose, often by an amount a person would barely notice, so that a model gives a wrong answer. Using one against a deployed model is called an evasion attack.

Taught in Adversarial examplesRelated: Fast gradient sign method, Adversarial training, Gradient, Jailbreak

A string of tokens, found by an automated search over a model's gradients, that is added to the end of a prompt to make a language model comply with a request it would otherwise refuse.

Taught in Adversarial examplesRelated: Adversarial example, Jailbreak, Refusal

Training a model on adversarial examples made from its own training data during training, with their correct labels, so small worst-case changes stop fooling it. It usually costs some accuracy on clean inputs.

Taught in Adversarial examplesRelated: Adversarial example, Fast gradient sign method, Regularization

A system that runs a language model in a loop with tools, letting it take several steps, observe results, and decide what to do next toward a goal. Definitions vary in how much autonomy they imply.

Taught in Models that use toolsRelated: Tool use, Prompt injection

In reinforcement learning, the learner that observes its situation, chooses actions, and receives rewards. Not to be confused with a language model agent that uses tools.

Taught in Explore or exploit: banditsRelated: Reinforcement learning, Policy, Reward

The software around a language model that turns it into an agent: it supplies the instructions and tool definitions, runs the tools the model asks for, enforces permissions and limits, and manages what fits in the context window.

Taught in Inside an agent harnessRelated: Agent, Tool use, Context window, System prompt

Everything a model or AI application is built from that someone else made: datasets, pretrained weights, adapters, code libraries, conversion tools, and, for agents, tools and plugins. Each can be tampered with.

Taught in Poisoned data and the supply chainRelated: Data poisoning, Pickle, Checksum, Model card

When a frequency above half the sampling rate is recorded, its samples look exactly like those of a lower frequency, so it shows up in the wrong place. Recorders filter such frequencies out before sampling.

Taught in How computers hearRelated: Sampling rate, Fourier transform

Getting an AI system's behavior to match what its designers and users intend, including avoiding harmful behavior nobody asked for. A system can be very capable and still poorly aligned.

Taught in Alignment: getting the goal rightRelated: Specification gaming, Goal misgeneralization, Reward hacking

A model complying with its training objective while it infers it is being trained, in order to avoid being changed, and behaving differently when it infers it is not. Demonstrated so far in experimental setups built to test for it.

Taught in How alignment is attemptedRelated: Alignment, RLHF, Backdoor

A per-pixel opacity map from 0 (background) to 1 (subject), with values in between at soft edges such as hair, used to cut a subject out and place it over a new background.

Taught in Cutting objects outRelated: Image segmentation

The basic unit of a neural network. It multiplies each input by a weight, adds them up with a bias, and passes the total through an activation function.

Taught in The artificial neuronRelated: Weight, Bias (neuron), Activation function, Neural network

A mechanism that lets each token build a new vector as a weighted mix of other tokens' value vectors, with weights set by how well its query matches their keys.

Taught in Attention, step by stepRelated: Self-attention, Query, key, and value, Transformer, Softmax

One independent attention computation inside a multi-head attention layer, with its own learned projections and its own pattern of weights.

Taught in Many heads and the causal maskRelated: Multi-head attention, Induction head

A position, usually the first token, where many attention heads put most of their attention when nothing else is relevant. Attention weights must add up to one, so they need somewhere to rest.

Taught in Reading a model layer by layerRelated: Attention head, Attention

Naming the kinds of sound in a recording, such as speech, rain, or a siren, from a fixed list of labels. Unlike speech recognition, it says what is making the sound, not what words are said.

Taught in What sound is this?Related: Spectrogram, Multi-label classification, Speech recognition

A network trained to rebuild its own input after squeezing it through a narrow middle layer. The encoder half compresses; the decoder half reconstructs; the middle layer's values are a learned code.

Taught in Autoencoders and latent spacesRelated: Variational autoencoder, Latent space, Reconstruction loss, Sparse autoencoder

Producing a sequence one element at a time, where each new element is predicted from all the ones before it and then appended to the input.

Taught in Predicting the next tokenRelated: Large language model, Token, Causal mask

Hidden behavior planted in a model that stays dormant on ordinary inputs and appears when an input contains a chosen trigger, such as a small patch in an image or a phrase in a prompt.

Taught in Poisoned data and the supply chainRelated: Data poisoning, Fine-tuning, Red teaming

How a network works out which way to nudge every weight: one backward pass carries the error from the output back through the layers, giving each weight its share of the blame (its gradient, found with the chain rule).

Taught in BackpropagationRelated: Chain rule, Computational graph, Gradient

Training a recurrent network by unrolling it into one copy per step, running ordinary backpropagation, and adding up the gradients each copy contributes to the shared weights.

Taught in Before transformers: recurrent networksRelated: Backpropagation, Recurrent neural network, Vanishing gradient, Chain rule

Bootstrap aggregating: train many models, each on its own bootstrap sample of the data, and combine them by voting or averaging. It mainly reduces the variance of unstable models such as deep trees.

Taught in Ensembles: forests and boostingRelated: Bootstrap sample, Random forest, Ensemble

A language model after pretraining only. It continues text in whatever style the input suggests, rather than following instructions or answering like an assistant.

Taught in Inside a real language modelRelated: Pretraining, Fine-tuning

How common something is in the population before any evidence, such as the share of people who have a condition. Ignoring it makes rare events look far more likely after a positive test than they are.

Taught in Probability and Bayes' ruleRelated: Prior, Bayes' rule, Precision

A number subtracted from the return in a policy gradient estimate, often an estimate of the average return. Any baseline that ignores the action taken leaves the estimate unbiased; a good one makes it much less noisy.

Taught in Policy gradientsRelated: REINFORCE, Policy gradient, Value function

How many training examples are used to compute each gradient step. Small batches give noisy but frequent updates; large batches give smoother but fewer updates per epoch.

Taught in Training a networkRelated: Epoch, Stochastic gradient descent

The rule for turning P(evidence | hypothesis) into P(hypothesis | evidence): multiply by the prior and divide by the overall chance of the evidence.

Taught in Probability and Bayes' ruleRelated: Prior, Likelihood, Posterior, Conditional probability

The rule linking a state's value to its neighbors': a state is worth the expected reward of the next step plus the discounted value of where you land. For the best policy, take the best action.

Taught in States, actions, and rewardsRelated: Value function, Value iteration, Markov decision process

A fixed set of test questions or tasks with a scoring rule, used to compare models. A score is an estimate from a sample of questions, so it comes with noise.

Taught in Evaluation: what a score meansRelated: Test set, Confidence interval, Data contamination

Generating n responses and keeping the one a reward model scores highest. More samples mean more optimization pressure on the reward model.

Taught in Rewarding LLMs, and reward hackingRelated: Reward model, Goodhart's law

A 16-bit floating point format with float32's 8 exponent bits and only 7 fraction bits: the same range as float32, less precision than float16. Designed at Google Brain and widely used for training.

Taught in Quantization: smaller numbersRelated: Floating point, Quantization

A number a neuron adds to its weighted sum before the activation function. It shifts how easily the neuron switches on, regardless of the inputs.

Taught in The artificial neuronRelated: Artificial neuron, Weight, Parameter

A standard keyword ranking formula. It scores documents by the query words they contain, weighting rare words more and limiting the reward for repeated words.

Taught in Search by meaningRelated: Semantic search

Building an ensemble in sequence, where each new model concentrates on what the models before it got wrong, and the final prediction is a weighted combination of all of them.

Taught in Ensembles: forests and boostingRelated: AdaBoost, Gradient boosting, Ensemble

A sample of n points drawn with replacement from n points, so some appear more than once and about 37% are left out. Used to estimate uncertainty and to train varied models.

Taught in Ensembles: forests and boostingRelated: Bagging, Out-of-bag error

Finding body poses by detecting every keypoint of every person in the whole picture at once, then grouping the keypoints into people, as OpenPose does with part affinity fields.

Taught in Finding a body's poseRelated: Top-down pose estimation, Pose estimation

A model of pairwise comparisons: each item has a score, and the chance that A beats B is the sigmoid of A's score minus B's. Reward models for RLHF are trained with it.

Taught in Rewarding LLMs, and reward hackingRelated: Reward model, Sigmoid

Controlling how long a reasoning model thinks. Adding the end-of-thinking marker after a set number of tokens cuts it short; holding the marker back and appending "Wait" extends it. From the s1 paper (Muennighoff and colleagues, 2025).

Taught in Thinking longerRelated: Test-time compute, Chain of thought

A way to build a subword vocabulary by starting from single characters or bytes and repeatedly merging the most frequent neighboring pair into a new token.

Taught in TokensRelated: Tokenizer, Vocabulary, Token

How well a model's confidence matches how often it is right. A calibrated model that says 80% is right about 80% of the time; modern neural networks are often overconfident.

Taught in Recognizing objects with your cameraRelated: Softmax, Probability

A mask that stops each token from attending to tokens after it, by setting those attention scores to negative infinity before the softmax. Used by models that generate text left to right.

Taught in Many heads and the causal maskRelated: Self-attention, Autoregressive generation

Finding where a model stores and moves information by restoring or swapping activations layer by layer and position by position and measuring the effect on the output.

Taught in Switching parts off and steeringRelated: Activation patching, Interpretability

The average position of a group of points. In k-means, each cluster is represented by its centroid.

Taught in Finding groups: k-meansRelated: k-means

Intermediate reasoning steps a language model writes before its final answer. Prompting for them, or training models to produce them, often improves results on multi-step problems.

Taught in Thinking longerRelated: Test-time compute, Self-consistency

The calculus rule for nested functions: if L depends on a and a depends on z, then dL/dz = dL/da times da/dz. Backpropagation applies it over and over.

Taught in BackpropagationRelated: Backpropagation, Derivative

The exact layout a chat model was fine-tuned on, which turns a list of messages into one text: special tokens mark where each message starts and ends and who wrote it, and the text ends where the assistant's reply should begin.

Taught in How a chat assistant worksRelated: Special token, System prompt, Instruction tuning, Base model

A saved copy of files taken before an agent changes them, so the change can be undone. Checkpoints cover only what the harness tracks; version control such as Git keeps the lasting history.

Taught in Inside an agent harnessRelated: Agent harness

A short fingerprint computed from a file's bytes, such as a SHA-256 hash. Changing even one bit changes it, so comparing it with a value pinned in advance shows whether a file is exactly the one expected.

Taught in Poisoned data and the supply chainRelated: Safetensors, AI supply chain, Model card

Splitting documents into smaller passages before embedding them, so each passage can be retrieved and fit in a prompt on its own.

Taught in Retrieval-augmented generationRelated: Retrieval-augmented generation, Sentence embedding

A small part of a neural network, such as a few attention heads in different layers, that together carry out one specific, understandable computation.

Taught in How circuits form: induction headsRelated: Induction head, Previous-token head, Attention head

A prediction task where the answer is one of a fixed set of categories, such as spam or not spam, or which digit a drawing shows.

Taught in Classification and probabilityRelated: Regression, Logistic regression, Decision boundary

A sampling technique for conditional generative models, diffusion or token by token, that compares the model's prediction with and without the prompt and pushes further toward the prompt, trading variety for closer matching.

Taught in Diffusion on imagesRelated: Diffusion model, DDIM

A model whose weights are not released, so you can use it only through the company's app or API, under its terms of service, while it runs on the company's computers.

Taught in Open and closed modelsRelated: Open weights, Cloud AI

A model that runs on a company's computers in a data center. Your app sends what you type, say, or upload over the internet, and the reply comes back the same way.

Taught in Where your AI runsRelated: On-device AI, GPU

A recommender that learns from behavior knows nothing about a new user or a new item until they have some ratings or clicks, so it cannot personalize for the one or recommend the other on its merits.

Taught in How your feed decides: recommendation systemsRelated: Collaborative filtering, Content-based filtering, Explore or exploit tradeoff

Recommending from behavior alone. It ignores what the items are and uses only who liked what, on the bet that people who agreed in the past will agree again.

Taught in How your feed decides: recommendation systemsRelated: Content-based filtering, Matrix factorization, Cold-start problem

A nonprofit that crawls the web and publishes free snapshots of billions of pages. Most open training datasets for language models start from its crawls.

Taught in Where training data comes fromRelated: Training data, Deduplication, robots.txt

Freeing room in a full context window by replacing older parts of a conversation with a shorter summary. Details the summary leaves out are gone from what the model sees.

Taught in Inside an agent harnessRelated: Context window, Agent harness

A diagram of a calculation as simple operations connected by the values flowing between them. Each operation knows its own local derivative, which makes backpropagation mechanical.

Taught in BackpropagationRelated: Backpropagation, Chain rule

The probability of A given that B happened, written P(A | B): the share of B cases that are also A. P(A | B) and P(B | A) are usually different numbers.

Taught in Probability and Bayes' ruleRelated: Probability, Bayes' rule

A range around a measured number that shows how much it could move by chance. For an accuracy p on n questions, the usual 95% interval (the normal approximation) is about p plus or minus 1.96 times the square root of p(1 - p)/n.

Taught in Evaluation: what a score meansRelated: Benchmark

A table counting a classifier's results by true class and predicted class: true positives, false positives, false negatives, and true negatives.

Taught in Measuring a classifierRelated: Precision, Recall

A training method from Anthropic in which a model critiques and revises its own answers against a written list of principles, and a preference model trained on AI judgments, not human labels of harm, supplies the reward.

Taught in How alignment is attemptedRelated: RLHF, Reward model, Alignment

A signed record attached to an image, video, audio, or document file saying who made it, with what tools, and how it was edited (the C2PA standard). It shows tampering, not truth, and can be removed.

Taught in Safety in AI productsRelated: Watermark, Generative AI

Recommending from descriptions of the items, such as genre, cast, or tempo, by matching them to the features of things a person liked. It works for new items but needs every item described.

Taught in How your feed decides: recommendation systemsRelated: Collaborative filtering, Cold-start problem

Choosing and maintaining what goes into a model's context window on each call, such as instructions, tool definitions, retrieved text, and tool results, because a longer context costs more and is used less well.

Taught in Long context in practiceRelated: Context window, Compaction, Retrieval-augmented generation

The maximum number of tokens a language model can take into account at once, covering both the prompt and the text it generates.

Taught in How a chat assistant worksRelated: Token, Large language model, Retrieval-augmented generation, Compaction

Training that pulls the embeddings of matching pairs together and pushes non-matching pairs apart, so the model learns a useful space without category labels.

Taught in One space for images and textRelated: Embedding, Cosine similarity, Zero-shot

Sliding a small grid of weights (a kernel) across an image and computing a weighted sum at every position, producing a new image called a feature map.

Taught in ConvolutionsRelated: Kernel (convolution), Feature map, Convolutional neural network

A neural network whose early layers are convolutions, so the same small learned patterns are detected everywhere in an image. The standard design for image recognition for over a decade.

Taught in ConvolutionsRelated: Convolution, Kernel (convolution), Feature map

A measure of how closely two vectors point in the same direction, from -1 (opposite) through 0 (perpendicular) to 1 (same direction), ignoring their lengths.

Taught in Vectors and similarityRelated: Embedding

A loss for predicted probabilities: minus the log of the probability the model gave to the right answer. Confident wrong answers are punished very heavily.

Taught in Classification and probabilityRelated: Loss function, Logistic regression, Softmax

The collection of ways data behaves unintuitively in many dimensions. For example, distances between random points become nearly all the same, so nearest neighbors stop standing out.

Taught in Seeing high dimensionsRelated: Principal component analysis, Euclidean distance, k-nearest neighbors

Making extra training examples by applying small, label-preserving changes to existing ones, such as shifting, rotating, or rescaling an image.

Taught in Reading handwritingRelated: Training data, Overfitting

When training data misrepresents the situations or people a model will be used on, for example by leaving some groups out, so the model works worse for them.

Taught in What AI gets wrongRelated: Training data, Distribution shift

When test questions, or their answers, end up in a model's training data, so a benchmark score partly measures memory instead of skill.

Taught in Evaluation: what a score meansRelated: Benchmark, Test set, Overfitting

The share of training that comes from each source, such as web pages, books, code, and mathematics. Developers tune it, often showing high-quality sources more than once.

Taught in Where training data comes fromRelated: Training data, Pretraining, Scaling laws

Tampering with the data a model learns from, by adding, changing, or relabeling examples, so the trained model behaves as the attacker wants on some inputs.

Taught in Poisoned data and the supply chainRelated: Backdoor, Training data, RAG poisoning, Memory poisoning

Denoising diffusion implicit models: a way to sample from a trained diffusion model without adding fresh noise at each step, so the same starting noise always gives the same image and far fewer, larger steps work.

Taught in Diffusion on imagesRelated: Diffusion model, Classifier-free guidance

The line or surface where a classifier switches from predicting one class to another.

Taught in Classification and probabilityRelated: Classification, Logistic regression

A decision tree with a single question and two leaves. Weak on its own, it is a common building block for boosting.

Taught in Ensembles: forests and boostingRelated: Decision tree, AdaBoost

A model that classifies (or predicts a number) by asking a sequence of simple questions about the input, following one branch per answer down to a leaf that gives the prediction.

Taught in Decision treesRelated: Gini impurity, Information gain, Random forest

Removing copies from a dataset, both exact copies and near copies such as a reposted article with a new headline. In training data it reduces wasted compute and memorization.

Taught in Where training data comes fromRelated: MinHash, Memorization, Training data

Machine learning with neural networks that have many layers. It powers modern image recognition, speech recognition, translation, and chatbots.

Taught in What is AI, really?Related: Machine learning, Neural network

Protecting a system with several independent layers, each with known limits, so that one layer failing does not decide the outcome. For AI apps: training, filters, least privilege, confirmation, isolation, and monitoring together.

Taught in Attacks on AI systemsRelated: Least privilege, Sandbox, Prompt injection

The slope of a function at a point: how much the output changes for a tiny change in the input.

Taught in Gradient descentRelated: Gradient

A generative model that learns to reverse a gradual noising process, so it can turn pure random noise into new data such as images, step by step.

Taught in Diffusion: from noise to dataRelated: Score, Classifier-free guidance, Generative AI

A diffusion or flow-matching model whose denoising network is a transformer that treats small patches of the noisy image or its latent as tokens, instead of a U-Net.

Taught in Diffusion on imagesRelated: Diffusion model, Transformer, U-Net

Splitting a token's final score into the direct contribution of the embedding, each attention head, and each MLP. The parts add up exactly to the score, but indirect effects through later layers are not counted.

Taught in Reading a model layer by layerRelated: Residual stream, Logit, Interpretability

A way to train a language model on pairs of preferred and rejected responses with a single classification-style loss, without a separate reward model or reinforcement learning.

Taught in How LLMs are trainedRelated: RLHF, Fine-tuning

A number gamma between 0 and 1 that shrinks rewards the further in the future they arrive. Low gamma makes an agent short-sighted; gamma near 1 makes it patient.

Taught in States, actions, and rewardsRelated: Return, Value function

Representing each idea as a pattern spread across many numbers, with every number taking part in many ideas, instead of giving each idea a number of its own.

Taught in Reading an embeddingRelated: Embedding, Sentence embedding

When the data a model meets in use differs from the data it was trained on. Models often get worse under shift, sometimes without becoming any less confident.

Taught in What AI gets wrongRelated: Training data, Generalization, Data bias

The sum of the products of matching entries of two vectors. It equals the product of their lengths times the cosine of the angle between them.

Taught in Vectors and similarityRelated: Vector, Cosine similarity

A rating scale from chess where a 400-point gap means 10 to 1 odds of winning. Online Elo nudges two ratings after every game; leaderboards like Chatbot Arena report ratings on this scale.

Taught in Evaluation: what a score meansRelated: Bradley-Terry model, Benchmark

A list of numbers (a vector) that represents a token, word, sentence, or image, learned so that similar things end up with similar vectors.

Taught in Word embeddingsRelated: Token, Attention

Reconstructing the text that produced an embedding from the embedding itself. Research recovers much of a short text, more easily with access to the same embedding model.

Taught in Poisoned retrieval and memoryRelated: Embedding, Sentence embedding, Vector database

A model in two parts. The encoder reads the whole input (a sentence, or a spectrogram) into a set of vectors; the decoder writes the output one token at a time, looking back at those vectors through cross-attention.

Taught in Speech to textRelated: Transformer, Attention, Autoregressive generation

A model that takes speech in and gives speech out with one network, instead of a cascade of speech recognition, a text model, and text to speech. It can hear and produce tone, laughter, or overlapping talk that text would drop.

Taught in A voice assistant taken apartRelated: Speech recognition, Text to speech, Latency, Multimodal model

A model made by combining the predictions of many models, for example by voting or adding them up. It works best when the members are good and make different mistakes.

Taught in Ensembles: forests and boostingRelated: Bagging, Random forest, Boosting

A measure of how uncertain a probability distribution is, in bits. Zero means one outcome is certain; k bits is as uncertain as a fair choice among 2 to the power k outcomes.

Taught in Inside a real language modelRelated: Softmax, Temperature, Cross-entropy

One full pass through the training data. Training usually runs for many epochs, shuffling the examples each time.

Taught in Training a networkRelated: Batch size, Stochastic gradient descent

An exploration rule: with a small probability epsilon take a random action, otherwise take the action that currently looks best.

Taught in Explore or exploit: banditsRelated: Explore or exploit tradeoff, Q-learning

The straight-line distance between two points, the square root of the sum of squared differences between matching entries.

Taught in Vectors and similarityRelated: Vector, Cosine similarity, k-nearest neighbors

An AI system able to do more damage than its job needs, through tools it does not use, permissions wider than its task, or high-impact actions taken without a check.

Taught in Agents with too much powerRelated: Agent, Least privilege, Tool use

A gradient that grows very large as it is passed back through many steps or layers, because it is multiplied by many factors larger than 1. Training steps then jump wildly.

Taught in Before transformers: recurrent networksRelated: Vanishing gradient, Gradient clipping, Backpropagation through time

The choice between taking the action that looks best so far (exploit) and trying something less known that might be better (explore). Too little exploring locks in mistakes; too much wastes reward.

Taught in Explore or exploit: banditsRelated: Epsilon-greedy, Upper confidence bound, Thompson sampling

A rule in United States copyright law that allows some uses of copyrighted work without permission, decided case by case on four factors, including the purpose of the use and its effect on the market for the work.

Taught in Where training data comes fromRelated: Training data

A one-step adversarial attack: move every input value by a small amount epsilon in the direction of the sign of the loss's gradient with respect to the input. Repeated in small, clipped steps, it becomes projected gradient descent.

Taught in Adversarial examplesRelated: Adversarial example, Gradient, Backpropagation

One measurable property of an example, given to a model as a number, such as a fruit's weight or a pixel's brightness.

Taught in Turning things into numbersRelated: Label, Training data, Model

The output of one convolution kernel applied across a whole image: a grid showing how strongly that kernel's pattern appears at each location.

Taught in ConvolutionsRelated: Convolution, Kernel (convolution)

The MLP inside each transformer block. It processes every token's vector on its own, usually widening it about four times, applying a nonlinearity, and projecting back. It holds most of each block's parameters.

Taught in The transformer blockRelated: Transformer, Residual stream, Attention

A cycle in which a system's outputs shape the data it later learns from. A feed that learns from what people do with what it showed them learns little about what it never shows.

Taught in How your feed decides: recommendation systemsRelated: Popularity bias, Explore or exploit tradeoff

Putting a few worked examples of a task in a language model's prompt, so it continues the pattern for a new case without any further training.

Taught in Reading an embeddingRelated: Base model, Zero-shot, In-context learning

The idea, popularized by Eli Pariser in 2011, that personalization shows people mainly what fits their past behavior and so isolates them from other views. How strong this effect is on real platforms is still debated.

Taught in How your feed decides: recommendation systemsRelated: Feedback loop, Recommender system

Continuing to train a pretrained model on a smaller, targeted dataset, such as example conversations, to change its behavior. Supervised fine-tuning uses examples of good responses.

Taught in How LLMs are trainedRelated: Pretraining, RLHF, Base model

A way to store real numbers as a sign, an exponent that sets the scale, and a fraction that sets the precision. float32 and float16 are the common 32-bit and 16-bit IEEE 754 formats.

Taught in Quantization: smaller numbersRelated: bfloat16, Quantization

A floating point operation, one multiplication or addition of numbers with fractional parts. Training and running models are measured in how many of these they take, and chips in how many they do per second (FLOP/s).

Taught in Compute and energyRelated: GPU, Floating point, Scaling laws, Inference

Training a network to predict the velocity that carries noise to data along chosen paths, then generating by following that velocity from noise. With straight paths it is also called rectified flow.

Taught in Diffusion: from noise to dataRelated: Diffusion model, Score

A way to rewrite a signal as a sum of sine waves, giving how much of each frequency it contains. Computers use the discrete version on a block of samples, computed quickly with the fast Fourier transform (FFT).

Taught in How computers hearRelated: Spectrogram, Sampling rate

A simpler gated recurrent network than the LSTM, with an update gate that blends the old hidden state with a new candidate and a reset gate. Introduced by Cho et al. in 2014.

Taught in Before transformers: recurrent networksRelated: Long short-term memory, Recurrent neural network

How well a model performs on new data it was not trained on. It is the goal of machine learning; doing well on the training data is only a means to it.

Taught in Overfitting and generalizationRelated: Overfitting, Test set

Models that produce new content, such as text, images, audio, or code, rather than only labeling or scoring existing content.

Taught in What is AI, really?Related: Deep learning, Large language model

A measure of how mixed the classes in a group are: the chance that two members drawn at random (with replacement) belong to different classes. 0 for a pure group, 0.5 for an even two-class mix.

Taught in Decision treesRelated: Entropy, Information gain, Decision tree

When a system keeps its skills in a new situation but pursues the wrong goal, one that matched the intended goal in training. It can happen even when the training reward was correct.

Taught in Alignment: getting the goal rightRelated: Alignment, Distribution shift, Specification gaming

The observation, usually phrased as "when a measure becomes a target, it ceases to be a good measure": optimizing hard for a proxy breaks its link to the real goal.

Taught in Rewarding LLMs, and reward hackingRelated: Reward hacking, Specification gaming

A graphics processing unit, a chip that does thousands of simple calculations at once. Built for graphics, GPUs also suit the many multiplications inside neural networks, so most AI training and serving uses them or similar chips.

Taught in Where your AI runsRelated: Memory bandwidth, Cloud AI

The list of slopes of a function, one per parameter. It points in the direction in which the function increases fastest, so its opposite points downhill.

Taught in Gradient descentRelated: Gradient descent, Derivative, Loss function

A boosting method that adds small trees one at a time, each fitted to the negative gradient of the loss (for squared error, the residuals), scaled by a learning rate.

Taught in Ensembles: forests and boostingRelated: Boosting, Gradient descent, Decision tree

Rescaling the gradient whenever its size passes a threshold, so one unusually large gradient cannot throw training off course. A standard fix for exploding gradients.

Taught in Before transformers: recurrent networksRelated: Exploding gradient, Gradient

An optimization method that repeatedly moves the parameters a small step in the direction that lowers the loss fastest, the opposite of the gradient.

Taught in Gradient descentRelated: Gradient, Learning rate, Optimizer, Stochastic gradient descent

Generating text by always choosing the single most probable next token. Equivalent to sampling at temperature 0.

Taught in Predicting the next tokenRelated: Temperature, Top-k sampling, Top-p sampling

Giving each group of consecutive values, such as 64 or 128 weights, its own quantization scale, so one large value only affects its own group.

Taught in Quantization: smaller numbersRelated: Quantization, Absmax quantization, Outlier

An attention variant in which groups of query heads share one key head and one value head. It shrinks the KV cache with little loss in quality. Multi-query attention is the extreme case of a single shared key/value head.

Taught in Generating fast: the KV cacheRelated: KV cache, Multi-head attention, Attention

When a language model states something false or invented, such as a fake quote or citation, in the same fluent, confident style as true statements.

Taught in What AI gets wrongRelated: Language model, Large language model

A layer of neurons between a network's inputs and its output. Its values are not given or read directly; the network works them out as intermediate steps.

Taught in Networks of neuronsRelated: Neural network, Artificial neuron

The vector a recurrent network carries from one step of a sequence to the next. It is the network's only memory of the inputs it has read so far.

Taught in Before transformers: recurrent networksRelated: Recurrent neural network, Long short-term memory

A setting chosen by people rather than learned from data, such as the learning rate, the number of layers, or the strength of regularization.

Taught in Overfitting and generalizationRelated: Parameter, Validation set

Deciding which pixels of a picture belong to which object or region, producing a mask for each, rather than a label for the whole picture or a box around each object.

Taught in Cutting objects outRelated: Object detection, Promptable segmentation, Alpha matte

A whole number standing for a small square of a picture: an index into an image tokenizer's codebook. A model can write image tokens one at a time, as a language model writes words, and a decoder turns them into pixels.

Taught in Pictures, one token at a timeRelated: Vector quantization, Token, Autoregressive generation

A large collection of labeled photos organized by the WordNet hierarchy. Its best-known subset, 1.28 million training images in 1,000 categories, became the standard benchmark for image classification.

Taught in Recognizing objects with your cameraRelated: Classification, Benchmark, Object detection

Signals of preference that people give without rating anything, such as clicks, plays, purchases, watch time, and skips. Plentiful but indirect; a missing click is not the same as a dislike.

Taught in How your feed decides: recommendation systemsRelated: Recommender system

A model picking up a pattern from the text in its prompt, such as repeating a name or following an example, without any change to its weights.

Taught in How circuits form: induction headsRelated: Induction head, Large language model

An attention head that, at a token A, looks back for earlier copies of A and attends to the token that followed them, helping the model continue repeated patterns.

Taught in Many heads and the causal maskRelated: Attention head, Multi-head attention, Previous-token head, Circuit, In-context learning

Running a trained model to get an output, as opposed to training it. Every prompt a chatbot answers is inference, and over a popular model's life it can use more energy than the training did.

Taught in Compute and energyRelated: Pretraining, KV cache, Test-time compute, FLOP

How much a question lowers the entropy of the class labels, averaging the resulting groups by size. Decision tree algorithms such as ID3 ask the question with the largest gain.

Taught in Decision treesRelated: Entropy, Gini impurity, Decision tree

Fine-tuning a pretrained model on many tasks written as instructions paired with good responses, so it learns to follow instructions it has not seen before.

Taught in Fine-tuning and LoRARelated: Fine-tuning, Base model, RLHF

The study of how a trained model computes its outputs, by reading and intervening on its internal activations and weights.

Taught in Reading a model layer by layerRelated: Logit lens, Activation patching, Steering vector

How well two regions agree: the area they share divided by the area either one covers, from 0 (no overlap) to 1 (identical). Used to compare boxes in detection and masks in segmentation.

Taught in Cutting objects outRelated: Object detection, Image segmentation

The similarity of two sets, from 0 to 1, measured as the number of items they share divided by the number of items in either one.

Taught in Where training data comes fromRelated: MinHash, Shingle

A prompt written to get a model to do what its training or policy is meant to prevent, for example by role-play, odd encodings, or many fake example dialogues.

Taught in Safety in AI productsRelated: Refusal, Prompt injection, Red teaming

A clustering method that places k centers, assigns every point to its nearest center, moves each center to the average of its points, and repeats until nothing changes.

Taught in Finding groups: k-meansRelated: Unsupervised learning, Centroid

A method that predicts the label of a new point by finding the k most similar labeled examples and taking a vote among them.

Taught in What is AI, really?Related: Machine learning, Classification

The small grid of weights, often 3 x 3, that a convolution slides across an image. In a convolutional network the kernel values are learned.

Taught in ConvolutionsRelated: Convolution, Feature map

A named landmark on a body or object, such as the left elbow or the nose, given as a point in the picture. The COCO dataset labels 17 for each person, from the nose to the ankles.

Taught in Finding a body's poseRelated: Pose estimation, Keypoint heatmap

A small grid a pose model outputs for each keypoint, with a high value wherever it thinks that keypoint is. It is trained to draw a bump centered on the true position, and the keypoint is read from the grid's peak.

Taught in Finding a body's poseRelated: Pose estimation, Keypoint

A measure of how much one probability distribution differs from another, in bits or nats. It is zero only when the two are identical. In RLHF a KL penalty keeps the trained model close to where it started.

Taught in Quantization: smaller numbersRelated: Entropy, Cross-entropy, RLHF

Training a smaller student model to reproduce a larger teacher model's outputs, so the student learns from the teacher's predictions instead of, or as well as, human labels.

Taught in Depth from a single imageRelated: Vision transformer, Monocular depth estimation

Stored keys and values for every token already processed, kept between generation steps so each new token only needs its own keys and values computed. It grows with every token in the context.

Taught in Generating fast: the KV cacheRelated: Attention, Grouped-query attention, Autoregressive generation

The answer a model should learn to give for an example, such as "apple" or "orange". Labeled examples are what supervised learning learns from.

Taught in Turning things into numbersRelated: Feature, Training data, Supervised learning

A model that assigns probabilities to what comes next in a piece of text. Writing text with one means repeatedly predicting the next word or token and picking one.

Taught in How a chatbot writesRelated: Large language model, n-gram model, Token, Temperature

A neural network, usually a transformer with billions of parameters, trained on large amounts of text to predict the next token. Chatbots are built on them.

Taught in Predicting the next tokenRelated: Token, Transformer, Autoregressive generation

The wait between a request and the response, such as the silence between the end of your question and the start of a voice assistant's answer. In a chain of models the waits add up.

Taught in A voice assistant taken apartRelated: Inference, KV cache

The space of learned, hidden coordinates a model uses to represent things internally, such as word vectors or a network's hidden layers. Nearby points stand for similar things.

Taught in Autoencoders and latent spacesRelated: Embedding, Vector, Autoencoder, Variational autoencoder, Principal component analysis, Diffusion model

A group of neurons that all read the same inputs. A fully connected layer computes Wx + b for a weight matrix W and bias vector b, then applies an activation function.

Taught in Layers are matrix multiplicationsRelated: Neural network, Hidden layer, Matrix multiplication

Rescaling each token's vector to a standard mean and spread, then applying a learned scale (and, in LayerNorm, a shift). Keeps every sublayer's input on a stable scale.

Taught in The transformer blockRelated: Residual stream, Transformer

The step size in gradient descent. Too small and training crawls; too large and each step overshoots, so the loss can grow instead of shrink.

Taught in Gradient descentRelated: Gradient descent, Optimizer

Giving a program, person, or agent only the access its task needs, so a mistake or a takeover can do less harm. OWASP's agentic list extends the idea to autonomy as least agency.

Taught in Agents with too much powerRelated: Excessive agency, Sandbox, Agent

Simon Willison's name for an agent that can read private data and untrusted content and can send data out, so planted text can make it leak. Meta's Rule of Two builds on it: hold at most two, or have a person supervise.

Taught in Agents with too much powerRelated: Prompt injection, Excessive agency, Agent

How probable the observed evidence is if a hypothesis is true, P(evidence | hypothesis). Comparing it across hypotheses (the likelihood ratio) says how strongly the evidence favors one over another.

Taught in Probability and Bayes' ruleRelated: Prior, Posterior, Bayes' rule

Predicting a number with a weighted sum of the inputs plus a constant. With one input, it is fitting a straight line.

Taught in Linear regression and lossRelated: Regression, Mean squared error

A map that can only rotate, stretch, shear, or reflect space, so straight lines stay straight and parallel lines stay parallel. Adding a bias also shifts it (an affine map).

Taught in Why nonlinearity mattersRelated: Layer, Activation function

A point lower than everything around it but not necessarily the lowest point overall. Gradient descent can settle in one and stop improving.

Taught in Gradient descentRelated: Gradient descent, Saddle point

The identity that the gradient of p(x) equals p(x) times the gradient of ln p(x). It turns the gradient of an expected reward into an expectation that can be estimated from samples.

Taught in Policy gradientsRelated: REINFORCE, Policy gradient

A classifier that computes a weighted sum of the inputs and passes it through the sigmoid to get the probability of the positive class. Despite the name, it is used for classification.

Taught in Classification and probabilityRelated: Sigmoid, Cross-entropy, Decision boundary

A raw, unnormalized score a model outputs for one option, such as one token in the vocabulary. Softmax turns a vector of logits into probabilities.

Taught in Predicting the next tokenRelated: Softmax, Temperature

Reading a model's next-token guess at an earlier layer by applying its final norm and output layer to that layer's residual stream, as if the model stopped there.

Taught in Reading a model layer by layerRelated: Residual stream, Logit, Interpretability

A recurrent network whose memory cell is updated by adding gated new information to a gated copy of the old cell, so information and gradients can survive many steps. Introduced in 1997.

Taught in Before transformers: recurrent networksRelated: Recurrent neural network, Gated recurrent unit, Vanishing gradient, Hidden state

A formula that turns a model's mistakes on the training data into a single number. Lower is better, and training tries to make it as low as possible.

Taught in The learning loopRelated: Mean squared error, Cross-entropy, Gradient descent

Hypothesized scenarios in which AI systems come to operate outside anyone's control, with regaining control extremely costly or impossible. Expert opinion on how likely such scenarios are varies greatly.

Taught in How alignment is attemptedRelated: Alignment, Alignment faking, Model evaluation

The tendency of language models to use information at the start or end of a long input better than information in the middle, named by Liu and colleagues in 2023.

Taught in Long context in practiceRelated: Context window, Needle in a haystack, Retrieval-augmented generation

A way to fine-tune a model by freezing its weights and learning each change as the product of two thin matrices, B times A, of a small rank r. It trains a tiny fraction of the parameters.

Taught in Fine-tuning and LoRARelated: Fine-tuning, Rank (of a matrix), Singular value decomposition, Parameter

A way of building software where the behavior is learned from examples with known answers, instead of being written as rules by a person.

Taught in What is AI, really?Related: Model, Deep learning, Training data

The standard model of a sequential decision problem: states, actions, transition probabilities for where each action leads, and rewards. What happens next depends only on the current state and action.

Taught in States, actions, and rewardsRelated: Policy, Value function, Bellman equation

Learning a vector for every user and every item so that their dot product, plus offsets, predicts the known ratings. It fills in the empty cells of the user-by-item ratings table.

Taught in How your feed decides: recommendation systemsRelated: Collaborative filtering, Dot product, Embedding, Regularization

Combining a matrix with a vector or another matrix by taking dot products of rows with columns. An (m x n) matrix times an n-vector gives an m-vector.

Taught in Layers are matrix multiplicationsRelated: Layer, Parameter

The average size of the differences between predictions and true values, ignoring sign. Less sensitive to a few wild outliers than mean squared error.

Taught in Linear regression and lossRelated: Mean squared error, Outlier

Turning a model's per-token vectors into one vector for the whole text by averaging them, number by number. Many sentence embedding models, including all-MiniLM-L6-v2, work this way.

Taught in Reading an embeddingRelated: Sentence embedding, Token, Embedding

The average of the squared differences between predictions and true values. Squaring makes big misses count much more than small ones.

Taught in Linear regression and lossRelated: Mean absolute error, Loss function, Linear regression

A pitch scale built from listening experiments, roughly linear below 1,000 Hz and logarithmic above, so equal steps in mels sound like equal steps in pitch. Speech models group frequencies into bands spaced evenly in mels.

Taught in How computers hearRelated: Spectrogram, Fourier transform

When a model can reproduce passages of its training data word for word. It is much more likely for text the model saw many times, which is one reason training data is deduplicated.

Taught in Where training data comes fromRelated: Deduplication, Training data, Overfitting

How many bytes a chip can read from its memory each second. A language model reads nearly all of its weights for every token it writes, so bandwidth divided by the model's size caps how fast it can write.

Taught in Where your AI runsRelated: KV cache, Quantization, GPU

Getting false facts or instructions saved into an assistant's or agent's memory, so they are read back into later conversations, not just the current one.

Taught in Poisoned retrieval and memoryRelated: Prompt injection, RAG poisoning, Agent

A way to estimate how similar two sets are from short signatures: for each of many random hash functions, keep each set's smallest hash. The share of matching minimums estimates their Jaccard similarity.

Taught in Where training data comes fromRelated: Jaccard similarity, Shingle, Deduplication

An architecture with many alternative sub-networks (experts) and a router that runs only a few of them for each token, so a model can store many parameters while using few per token.

Taught in Mixture of expertsRelated: Router, Parameter

A classic dataset of 70,000 handwritten digits, each a 28 x 28 grayscale image: 60,000 for training and 10,000 for testing.

Taught in Reading handwritingRelated: Training data, Test set

The rule a learning algorithm produces from data. Given an input, a model returns a prediction. In neural networks, the model is defined by its parameters.

Taught in What is AI, really?Related: Machine learning, Parameter

A short document published with a model that describes what it is for, how it was trained and evaluated, and its known limits. It is the publisher's own account, not proof of where the weights came from.

Taught in Poisoned data and the supply chainRelated: Model license, Open weights, AI supply chain

MCP, an open protocol from Anthropic (2024), now under the Linux Foundation, for connecting AI applications to outside tools and data. A server offers tools, resources, and prompts; the application decides what the model may use.

Taught in Inside an agent harnessRelated: Tool use, Agent harness

Systematic tests of what a model can do and how it behaves, such as benchmarks, dangerous capability evaluations, and red teaming. Often shortened to evals.

Taught in Alignment: getting the goal rightRelated: Benchmark, Red teaming, Alignment

The share of a chip's peak arithmetic rate that a run actually spends on the model's own operations. Meta reported 38 to 43 percent for the main training setups of Llama 3.1 405B.

Taught in Compute and energyRelated: FLOP, GPU

The terms under which a model's weights may be used, changed, and shared, from permissive open source licenses such as Apache 2.0 to licenses with use restrictions, user limits, or a ban on commercial use.

Taught in Open and closed modelsRelated: Open weights, Open source AI, Acceptable use policy

An optimizer that keeps a running sum of past gradients, like a ball gathering speed downhill, so it moves faster along consistent slopes and damps zig-zags.

Taught in Better optimizersRelated: Optimizer, Gradient descent, Adam

Guessing how far away every pixel is from a single photo, using learned cues such as overlap, perspective, and familiar sizes. Most models give relative depth, an order with no units, rather than meters.

Taught in Depth from a single imageRelated: Vision transformer, Knowledge distillation

A problem where you repeatedly choose among several options (like slot machines) with unknown payouts, aiming to win as much as possible. The situation never changes between choices.

Taught in Explore or exploit: banditsRelated: Explore or exploit tradeoff, Regret, Thompson sampling

Running several attention heads in parallel, each with its own query, key, and value projections, then concatenating their outputs and mixing them with one more matrix.

Taught in Many heads and the causal maskRelated: Attention, Attention head, Transformer

Classification where one input can have several labels at once, such as a clip with both speech and music. Each label gets its own yes-or-no probability from a sigmoid, so the probabilities need not add up to 1.

Taught in What sound is this?Related: Classification, Sigmoid, Softmax, Cross-entropy

A model that works with more than one kind of data, such as text, pictures, and sound. Some only read other kinds of data and answer in text; others can also produce pictures or speech.

Taught in Where multimodal models are headingRelated: Vision-language model, End-to-end speech model, Image token, Contrastive learning

A simple language model that predicts the next word by counting what followed the previous one or few words in its training text.

Taught in How a chatbot writesRelated: Language model

A classifier that applies Bayes' rule assuming the features are independent once the class is known, so each word (or feature) adds its own evidence. Simple and fast, but its probabilities tend to be overconfident.

Taught in Probability and Bayes' ruleRelated: Bayes' rule, Smoothing, Classification

A test that hides one fact at different depths in long filler text and asks the model to find it. Passing shows a model can locate a fact that matches the question; harder long-context tests often find much lower scores.

Taught in Long context in practiceRelated: Context window, Lost in the middle, Benchmark

A model built from layers of artificial neurons, where each layer's outputs become the next layer's inputs. Training adjusts all of the weights together.

Taught in Networks of neuronsRelated: Artificial neuron, Hidden layer, Deep learning

How much noise a diffusion model's forward process adds at each step. DDPM's schedule raises the per-step noise variance beta from 0.0001 to 0.02 over 1,000 steps, which fixes how much of the image survives at every step.

Taught in Diffusion on imagesRelated: Diffusion model

Finding every object of known kinds in an image and reporting each one as a box (its bounding box), a label, and a confidence score, instead of one label for the whole picture.

Taught in Recognizing objects with your cameraRelated: Classification, Vision transformer, ImageNet

COCO's 0 to 1 score for how close a predicted pose is to a labeled one. It allows more error for larger people and for keypoints labelers place less consistently, such as hips. It plays the role IoU plays for boxes.

Taught in Finding a body's poseRelated: Pose estimation, Intersection over union

A map of how much a model's output changes when each region of its input is covered in turn, made by running the model once per covered region.

Taught in Where the classifier looksRelated: Saliency map, Shortcut learning, Ablation

Learning about one policy (often the best one) while behaving according to another (often an exploring one). Q-learning is off-policy; SARSA is on-policy, learning about the policy it follows.

Taught in Learning from experience: Q-learningRelated: Q-learning, SARSA

A model that runs on your own phone, laptop, or browser, so what you give it is processed there instead of being sent to a company's computers. Its files must be downloaded and fit in your device's memory.

Taught in Where your AI runsRelated: Cloud AI, Quantization, Parameter, Memory bandwidth

As the Open Source Initiative defines it (2024): an AI system anyone may use for any purpose, study, modify, and share, released with its parameters, its complete code, and enough information about its training data to rebuild it.

Taught in Open and closed modelsRelated: Open weights, Model license

A model whose trained parameters are published for anyone to download, under a license that says what you may do with them. It does not mean the training data or code are published too.

Taught in Open and closed modelsRelated: Open source AI, Model license, Closed model, On-device AI

Object detection where the kinds of object to find are given as text at run time, so the detector can box things named by any word or phrase rather than a fixed list of labels.

Taught in Finding anything you can nameRelated: Object detection, Zero-shot, Contrastive learning

The rule that decides how to change a model's parameters at each training step, given the gradient. Gradient descent, momentum, and Adam are optimizers.

Taught in Better optimizersRelated: Gradient descent, Momentum, Adam

An estimate of a bagged ensemble's error on new data, computed by judging each training point only with the models whose bootstrap samples left it out. It needs no separate test set.

Taught in Ensembles: forests and boostingRelated: Bootstrap sample, Random forest

A data point far from the pattern of the rest, from a rare event or a recording error. Some loss functions are pulled hard by outliers.

Taught in Linear regression and lossRelated: Mean squared error, Mean absolute error

When a model fits the quirks and noise of its training data so closely that it does worse on new data. Training error keeps falling while test error rises.

Taught in Overfitting and generalizationRelated: Underfitting, Generalization, Regularization, Test set

A ranked list of the most important security risks for one kind of software, published by OWASP, a nonprofit security community. Its GenAI Security Project keeps one list for LLM applications and one for agentic applications.

Taught in Attacks on AI systemsRelated: Prompt injection, Excessive agency

A number inside a model that training adjusts, such as the slope of a line or a weight in a neural network. Large language models have billions of them.

Taught in The learning loopRelated: Model, Gradient descent

The chance that at least one of k attempts at a problem is correct, averaged over problems. Usually estimated from n ≥ k samples with the unbiased formula 1 - C(n - c, k) / C(n, k).

Taught in Evaluation: what a score meansRelated: Benchmark

A sudden change in a model's behavior or training loss over a short stretch of training, often because a new mechanism has formed inside the model.

Taught in How circuits form: induction headsRelated: Induction head, Circuit

Python's built-in format for saving objects. A pickle file can name functions to call while it loads, so opening an untrusted one can run code. PyTorch's own model files, written by torch.save, are built on it.

Taught in Poisoned data and the supply chainRelated: Safetensors, AI supply chain

One tiny square of a digital image. A grayscale pixel is stored as one brightness number; a color pixel as three (red, green, and blue).

Taught in Turning things into numbersRelated: Feature

An agent's rule for choosing actions: for each state, which action to take, or the probability of taking each one.

Taught in States, actions, and rewardsRelated: Value function, Policy gradient, Markov decision process

A family of methods that learn a policy directly, adjusting its parameters in the direction that increases expected return, estimated from sampled episodes.

Taught in Policy gradientsRelated: REINFORCE, Policy, Proximal policy optimization

A neuron that responds to several unrelated things, such as cat faces and the fronts of cars. Superposition is one proposed explanation for why such neurons are common.

Taught in Superposition: more features than neuronsRelated: Superposition, Sparse autoencoder, Neural network

The tendency of a recommender to show already-popular items more, because they have the most data and the surest scores, which makes them more popular still.

Taught in How your feed decides: recommendation systemsRelated: Feedback loop, Recommender system

Finding where a body's joints and other landmarks (keypoints such as the shoulders, elbows, knees, and ankles) are in a picture, usually drawn as a stick-figure skeleton.

Taught in Finding a body's poseRelated: Keypoint, Keypoint heatmap, Top-down pose estimation, Bottom-up pose estimation

Information about where each token sits in the sequence, given to a transformer because attention on its own ignores word order.

Taught in Where words areRelated: Rotary position embedding, Self-attention, Transformer

What you should believe about a hypothesis after seeing the evidence, P(hypothesis | evidence), computed from the prior and the likelihood by Bayes' rule.

Taught in Probability and Bayes' ruleRelated: Prior, Likelihood, Bayes' rule

A data center's total electricity divided by what its computers use. A value of 1.0 would mean no overhead for cooling and power conversion; Google reported a fleet average of 1.09 in 2025.

Taught in Compute and energyRelated: GPU, Inference

Of the items a classifier flagged as positive, the fraction that really are positive. High precision means few false alarms.

Taught in Measuring a classifierRelated: Recall, Confusion matrix

The first and most expensive stage of training a language model, in which it learns to predict the next token across a very large amount of text. The result is a base model.

Taught in How LLMs are trainedRelated: Base model, Fine-tuning, Large language model

An attention head that mostly looks at the position just before the current one, writing information about the previous token into each position.

Taught in How circuits form: induction headsRelated: Induction head, Attention head, Circuit

A method that finds the directions along which data varies the most, often used to project many-dimensional data down to two or three dimensions for viewing.

Taught in Seeing high dimensionsRelated: Curse of dimensionality, Embedding

What you believe about a hypothesis before seeing the evidence, as a probability. Bayes' rule combines it with the likelihood to give the posterior.

Taught in Probability and Bayes' ruleRelated: Posterior, Likelihood, Base rate

Splitting a system so the part that handles untrusted input cannot take privileged actions. In an AI app, one model reads untrusted text but has no tools, while the part that acts never reads that text (the dual LLM pattern).

Taught in Prompt injectionRelated: Prompt injection, Least privilege, Sandbox

A number from 0 to 1 saying how likely something is. One way to read it: out of many similar cases, the share in which it happens.

Taught in Probability and Bayes' ruleRelated: Conditional probability, Base rate

In a vision-language model, the small network that maps each vector from the vision encoder into the language model's embedding space, so the language model can read picture pieces like word tokens. Also called an adaptor or connector.

Taught in How a model reads a pictureRelated: Vision-language model, Vision encoder, Embedding, Linear transformation

Keeping the processed start of a prompt, its keys and values, so later requests that begin with the same text skip that work. Providers charge less for cached tokens, but they still fill the context window.

Taught in Long context in practiceRelated: KV cache, Context window

An attack in which text a model reads changes what it does against its developer's intent. Direct injection comes from the person typing; indirect injection hides in what the model reads, such as a web page, email, picture, or memory.

Taught in Prompt injectionRelated: Agent, Tool use, Sandbox, Jailbreak, Privilege separation

Segmentation guided by a prompt such as a click, a box, or a rough mask, so one model can outline whatever object the person points at, without a fixed list of classes. SAM is the best-known example.

Taught in Cutting objects outRelated: Image segmentation, Intersection over union

A policy gradient algorithm that limits how far each update can move the policy, by clipping the change in action probabilities. Used for RLHF in InstructGPT.

Taught in Policy gradientsRelated: Policy gradient, RLHF

A temporal difference method that learns action values from experience, updating toward the reward plus the discounted value of the best next action, whatever action the agent actually takes next.

Taught in Learning from experience: Q-learningRelated: Action value, Temporal difference learning, SARSA, Off-policy learning

Storing a model's numbers in fewer bits, for example as 8-bit or 4-bit integers with a scale, to save memory and speed up inference at some cost in accuracy.

Taught in Quantization: smaller numbersRelated: Absmax quantization, Zero-point quantization, Group-wise quantization, Floating point, bfloat16

The three vectors attention makes from each token. A token's query is compared with other tokens' keys to decide how much of each of their values to take.

Taught in Attention, step by stepRelated: Attention, Self-attention

Planting text in a knowledge base that a retrieval-augmented system searches, written so it is retrieved for chosen questions and steers the answers.

Taught in Poisoned retrieval and memoryRelated: Retrieval-augmented generation, Prompt injection, Vector database

An ensemble of decision trees, each trained on a bootstrap sample and allowed to choose only among a random subset of features at every split, combined by majority vote.

Taught in Ensembles: forests and boostingRelated: Bagging, Decision tree, Out-of-bag error

A number that sets where a computer's random number generator starts. The same seed gives the same "random" numbers every time, so a result that uses randomness can be repeated exactly.

Taught in How AI sees, hears, and makes picturesRelated: Diffusion model

The number of independent directions a matrix can produce. A matrix of rank r can be written as a sum of r pieces, each one column times one row.

Taught in Fine-tuning and LoRARelated: Singular value decomposition, Low-rank adaptation, Matrix multiplication

A language model trained, often with rewards for answers that can be checked, to write long reasoning before it answers. More thinking often helps on hard problems, at the cost of more tokens, time, and energy.

Taught in Thinking longerRelated: Test-time compute, Chain of thought, Budget forcing

Of all the items that really are positive, the fraction the classifier flagged. High recall means few misses.

Taught in Measuring a classifierRelated: Precision, Confusion matrix, ROC curve

The part of the input image that can affect one output of a layer. One 3 × 3 convolution sees 3 × 3 pixels; stacking layers widens it, while self-attention reaches the whole image in one layer.

Taught in How computers seeRelated: Convolution, Vision transformer, Feature map

Software that predicts how much each person will like, click, or watch each item, then shows each person the items with the highest predictions. Feeds, shops, and streaming apps run on them.

Taught in How your feed decides: recommendation systemsRelated: Collaborative filtering, Content-based filtering, Matrix factorization

How far an autoencoder's output is from its input, such as squared error or binary cross-entropy summed over the pixels. Training an autoencoder means making it small.

Taught in Autoencoders and latent spacesRelated: Autoencoder, Loss function, Cross-entropy

A network that reads a sequence one step at a time, applying the same weights at every step and passing a hidden state from each step to the next.

Taught in Before transformers: recurrent networksRelated: Hidden state, Backpropagation through time, Long short-term memory, Gated recurrent unit, Transformer

Deliberately trying to make a model behave badly, by people or by other models, to find failures before users do.

Taught in Alignment: getting the goal rightRelated: Model evaluation, Alignment

When a model declines a request, usually because it was trained to refuse requests that break its policy. Over-refusal is declining harmless requests that only look like harmful ones.

Taught in Safety in AI productsRelated: Jailbreak, Safety classifier, RLHF

A prediction task where the answer is a number, such as a price, a temperature, or a time.

Taught in The learning loopRelated: Classification, Linear regression, Supervised learning

The total reward given up by not always choosing the best option: summed over every choice, the best option's expected payout minus the chosen option's.

Taught in Explore or exploit: banditsRelated: Multi-armed bandit, Explore or exploit tradeoff

Any technique that discourages a model from fitting noise, for example adding a penalty on large parameter values to the loss.

Taught in Overfitting and generalizationRelated: Overfitting, Hyperparameter

The basic policy gradient algorithm (Williams, 1992): after an episode, make each action taken more likely in proportion to the return that followed it, minus a baseline.

Taught in Policy gradientsRelated: Policy gradient, Baseline (policy gradient), Log-derivative trick

Learning to act by trial and error: an agent takes actions, receives rewards, and adjusts its behavior to collect more reward over time, with no answer key telling it the right action.

Taught in Explore or exploit: banditsRelated: Agent (reinforcement learning), Reward, Supervised learning, Policy

The rectified linear unit, max(0, z): it passes positive inputs through unchanged and turns negative inputs into zero. The most common activation in deep networks.

Taught in Why nonlinearity mattersRelated: Activation function, Tanh, Sigmoid

The running vector for each token that flows through a transformer from the embedding to the output. Every attention and MLP sublayer reads from it and adds its result back.

Taught in The transformer blockRelated: Transformer, Layer normalization, Feed-forward network

A technique that searches a collection of documents for passages relevant to a question and adds them to a language model's prompt, so it can answer from that text.

Taught in Retrieval-augmented generationRelated: Semantic search, Chunking, Context window, Hallucination

The total reward an agent collects from some point on, usually with later rewards discounted: G = r1 + γ r2 + γ² r3 + ...

Taught in States, actions, and rewardsRelated: Reward, Discount factor, Value function

A number an agent receives after acting that says how well things went. The agent's goal is to collect as much reward as possible over time.

Taught in Explore or exploit: banditsRelated: Reinforcement learning, Return, Reward model

When an agent finds a way to score highly on its reward that does not achieve what the designers intended, by exploiting flaws in the reward or reward model.

Taught in Rewarding LLMs, and reward hackingRelated: Specification gaming, Goodhart's law, Reward model

A model trained to predict which of two responses people prefer, used to score a language model's outputs during reinforcement learning. It is a learned stand-in for human judgment, and can be wrong.

Taught in Rewarding LLMs, and reward hackingRelated: RLHF, Bradley-Terry model, Reward hacking

Reinforcement learning from human feedback. People compare pairs of model responses, a reward model learns to predict their preferences, and the language model is then trained with reinforcement learning to score well on it.

Taught in How LLMs are trainedRelated: Fine-tuning, Direct preference optimization

A file on a website that tells named crawlers which pages they may visit. Site owners use it to opt out of crawlers that collect AI training data; it is a request, not a technical barrier.

Taught in Where training data comes fromRelated: Common Crawl, Training data

A plot of the true positive rate against the false positive rate as a classifier's threshold sweeps from strict to lenient. The area under it (AUC) summarizes how well the scores rank positives above negatives.

Taught in Measuring a classifierRelated: Recall, Precision, Confusion matrix

A way to encode position by rotating each pair of query and key dimensions by an angle proportional to the token's position, so attention scores depend on position only through the offset between tokens.

Taught in Where words areRelated: Position encoding, Attention

The small learned layer in a mixture-of-experts model that scores every expert for each token and picks which few experts will process it.

Taught in Mixture of expertsRelated: Mixture of experts

A flat point that curves up in some directions and down in others, like the middle of a horse saddle. The gradient is zero there, but it is not a minimum.

Taught in Better optimizersRelated: Local minimum, Gradient descent

A file format for model weights: a short JSON header listing each tensor's name, number type, shape, and position, followed by the raw numbers. Loading it reads data and runs no code from the file.

Taught in Poisoned data and the supply chainRelated: Pickle, Checksum, AI supply chain

A separate model that checks what goes into or comes out of a language model against a policy, and blocks or flags what breaks it. Like any classifier, it makes errors both ways.

Taught in Safety in AI productsRelated: Refusal, Jailbreak, Classification

A heatmap over a model's input showing which parts most affect one output, usually computed from gradients of that output with respect to the input.

Taught in Where the classifier looksRelated: Occlusion map, Interpretability

How many times per second a recording measures the sound wave. CD audio uses 44,100 samples per second; Whisper works at 16,000. A rate can only capture frequencies below half of it.

Taught in How computers hearRelated: Aliasing, Fourier transform, Spectrogram

A boundary, enforced by the operating system or a separate machine, that limits what a program can reach, such as which files it can write and which network hosts it can contact, whatever the program intends.

Taught in Inside an agent harnessRelated: Agent harness, Prompt injection

A temporal difference method like Q-learning, except it updates toward the value of the action the agent will actually take next, exploration included. Named for state, action, reward, state, action.

Taught in Learning from experience: Q-learningRelated: Q-learning, Off-policy learning, Temporal difference learning

Methods meant to help people supervise AI systems on tasks too hard for them to judge directly, for example by having two models debate while a person judges, or having models critique other models' answers.

Taught in How alignment is attemptedRelated: RLHF, Alignment, Reward model

Fitted curves that predict how a model's loss falls as parameters, training data, or compute grow. They are empirical, and their constants depend on the architecture, data, and training recipe.

Taught in How LLMs are trainedRelated: Pretraining, Parameter

The gradient of the log of a probability density, which points in the direction where data becomes more likely. Diffusion models learn the score of noisy data at every noise level.

Taught in Diffusion: from noise to dataRelated: Diffusion model

Attention in which the queries, keys, and values all come from the same sequence, so every token can draw information from every other token in it.

Taught in Attention, step by stepRelated: Attention, Causal mask, Query, key, and value

Sampling several reasoning paths for the same question and returning the most common final answer. It helps when the correct answer is the model's most likely answer.

Taught in Thinking longerRelated: Chain of thought, Test-time compute

Search that ranks documents by how similar their meaning is to the query, usually by comparing sentence embeddings with cosine similarity.

Taught in Search by meaningRelated: Sentence embedding, BM25, Vector database

A single vector representing a whole sentence or passage, produced by a model trained so that texts with similar meaning get similar vectors.

Taught in Search by meaningRelated: Embedding, Semantic search, Contrastive learning

In a mixture-of-experts layer, an expert that runs for every token alongside the few the router picks, meant to hold what all tokens need so the routed experts can specialize.

Taught in Mixture of expertsRelated: Mixture of experts, Router

A short run of consecutive words, for example every 5-word sequence in a page. Comparing the sets of shingles two pages share is how near copies are found.

Taught in Where training data comes fromRelated: Jaccard similarity, MinHash, n-gram model

When a model learns a rule that scores well on its test data because of a quirk of that data, such as a watermark on one class, and fails once the quirk is gone.

Taught in Where the classifier looksRelated: Distribution shift, Occlusion map, Training data

The S-shaped function 1 / (1 + e^-z), which squashes any number into the range 0 to 1 so it can be read as a probability.

Taught in Classification and probabilityRelated: Logistic regression, Logit

Writing a matrix as a sum of rank-one pieces, each with a weight called a singular value, largest first. Keeping the largest r pieces gives the best rank-r approximation.

Taught in Fine-tuning and LoRARelated: Rank (of a matrix), Principal component analysis, Low-rank adaptation

Adding a small count to every possible word or value before turning counts into probabilities, so nothing unseen gets probability zero. Adding exactly 1 is called Laplace smoothing.

Taught in Probability and Bayes' ruleRelated: Naive Bayes

A function that turns a list of real-valued scores into probabilities that are positive and sum to 1, by exponentiating each score and dividing by the total.

Taught in Predicting the next tokenRelated: Logit, Temperature, Attention

A network trained to rebuild a model's activations from a much larger set of features, only a few active at a time. Used to pull apart features stored in superposition; a form of dictionary learning.

Taught in Superposition: more features than neuronsRelated: Superposition, Polysemantic neuron

A token reserved for structure rather than words, such as one marking the end of a text or the start and end of a chat message. When a chat model writes its end-of-turn token, the app stops generating.

Taught in How a chat assistant worksRelated: Token, Chat template, Tokenizer

Behavior that satisfies the literal specification of an objective without achieving the intended outcome, such as racing in circles to collect points instead of finishing the race.

Taught in Rewarding LLMs, and reward hackingRelated: Reward hacking, Goodhart's law

A picture of sound with time across, frequency up, and loudness as color, made by taking the frequencies of many short, overlapping windows of the recording and placing them side by side.

Taught in How computers hearRelated: Fourier transform, Mel scale, Sampling rate

Turning recorded speech into text. Modern systems such as Whisper are neural networks that read a spectrogram of the audio and write the transcript one token at a time.

Taught in Speech to textRelated: Spectrogram, Word error rate, Encoder-decoder

A direction added to a model's residual stream while it runs, usually the difference between its average activations on two contrasting sets of prompts, used to push its output one way.

Taught in Switching parts off and steeringRelated: Residual stream, Interpretability

Gradient descent where each step uses a small random batch of training examples instead of all of them. Steps are noisier but far cheaper, which is how large models are trained.

Taught in The learning loopRelated: Gradient descent, Batch size

A separate agent run that a main agent starts for a side task. It works in its own context window and hands back only a short result, which keeps the main conversation small.

Taught in Inside an agent harnessRelated: Agent, Agent harness, Context window

When a network represents more features than it has dimensions by giving them overlapping directions. It works when features are rarely active together, at the cost of some interference.

Taught in Superposition: more features than neuronsRelated: Polysemantic neuron, Sparse autoencoder, Feature

Learning from examples that come with the right answer attached, such as photos labeled "cat" or houses with their sale prices.

Taught in The learning loopRelated: Unsupervised learning, Regression, Classification, Training data

A benchmark of real GitHub issues from 12 Python projects. A system gets the code and the issue and must write a change; it counts as solved if the tests for the fix pass and the other tests still pass.

Taught in Inside an agent harnessRelated: Benchmark, Agent

A model telling people what they want to hear, such as agreeing with a wrong claim or praising weak work. Training on human approval can encourage it.

Taught in Safety in AI productsRelated: RLHF, Reward model, Reward hacking

Training data generated by a model or a program instead of collected from people, such as textbook-style exercises written by a language model.

Taught in Where training data comes fromRelated: Training data, Data mixture

Instructions an application places at the start of a language model's context, before the user's messages, to set its role, rules, and how it should use tools. The model reads them as text; nothing forces it to obey.

Taught in How a chat assistant worksRelated: Agent harness, Context window, Prompt injection, Chat template

The hyperbolic tangent, an S-shaped activation function that squashes any number into the range from -1 to 1. It is a rescaled sigmoid centered on zero.

Taught in Why nonlinearity mattersRelated: Activation function, Sigmoid, ReLU

A setting for how adventurous a model's choices are. Low temperature sticks to the likeliest options; high temperature spreads choices toward less likely ones. Technically, it divides the logits before softmax.

Taught in Predicting the next tokenRelated: Softmax, Logit, Top-p sampling

Learning a value estimate from the next step's reward plus the estimate at the next state, instead of waiting for the final outcome. The gap between the two is the TD error.

Taught in Learning from experience: Q-learningRelated: Q-learning, SARSA, Value function

Examples held back from training and used only to measure how well a model does on data it has never seen.

Taught in The learning loopRelated: Training data, Validation set, Overfitting

Computation a model spends while answering, rather than during training, such as writing out reasoning, sampling several answers, or searching over candidates. More of it often improves accuracy on hard problems.

Taught in Thinking longerRelated: Chain of thought, Self-consistency

Turning written text into spoken audio. Neural systems predict an intermediate representation of the sound, such as a spectrogram or a learned code, and a decoder or vocoder turns it into a waveform.

Taught in Text to speechRelated: Vocoder, Spectrogram, Speech recognition

A generative model that makes a picture from a written description, called a prompt. Many, such as Stable Diffusion, are diffusion models guided by the text; others build the picture piece by piece, the way a chatbot writes.

Taught in How AI sees, hears, and makes picturesRelated: Diffusion model, Generative AI, Classifier-free guidance

An exploration strategy that keeps a probability distribution over each option's payout, draws one plausible value from each, and picks the option with the highest draw.

Taught in Explore or exploit: banditsRelated: Upper confidence bound, Explore or exploit tradeoff

METR's measure of what an AI agent can do: the length of task, in the time a skilled person needs for it, that the agent completes with 50% success. It depends on the tasks chosen and how long people take.

Taught in Evaluation: what a score meansRelated: Benchmark, Agent

The unit of text a language model reads and writes. A token can be a whole word, part of a word, a single character, or punctuation, often including a leading space.

Taught in TokensRelated: Large language model, Logit

The program that splits text into tokens and maps each one to an id in a vocabulary, and turns ids back into text.

Taught in TokensRelated: Token, Vocabulary, Byte-pair encoding

Letting a language model request actions from external programs, such as searches, calculations, or API calls. The model outputs a structured call; the application runs it and returns the result.

Taught in Models that use toolsRelated: Agent, Prompt injection

Finding body poses in two steps. A detector first draws a box around each person, then a pose model finds the keypoints of the one person in each box. Its cost grows with the number of people.

Taught in Finding a body's poseRelated: Bottom-up pose estimation, Pose estimation, Object detection

Sampling the next token from only the k most probable tokens, after setting all others to zero probability and renormalizing.

Taught in Predicting the next tokenRelated: Top-p sampling, Temperature, Greedy decoding

Sampling the next token from the smallest set of most probable tokens whose probabilities add up to at least p. Also called nucleus sampling.

Taught in Predicting the next tokenRelated: Top-k sampling, Temperature

The date after which a model's training data contains nothing. The model cannot know later events unless an app puts them in its context, and it has no clock: today's date has to come from the app too.

Taught in How a chat assistant worksRelated: Context window, System prompt, Pretraining

The examples, usually with known answers, that a model learns from. A model can only learn patterns that are present in its training data.

Related: Machine learning, Model

A neural network architecture built from stacked layers of attention and feed-forward networks, introduced by Vaswani et al. in 2017. Most modern language models are transformers.

Taught in Attention, step by stepRelated: Attention, Large language model, Multi-head attention

A convolutional network shaped like a U: it shrinks an image into coarse, many-channel feature maps, then grows it back to full size, with skip connections carrying fine detail across. Many image diffusion models use one to predict noise.

Taught in Diffusion on imagesRelated: Convolutional neural network, Convolution, Diffusion model

When a model is too simple to capture the real pattern, so it does poorly on both the training data and new data.

Taught in Overfitting and generalizationRelated: Overfitting, Generalization

Learning from examples that have no answers attached, by finding structure such as groups or directions of variation in the data itself.

Taught in Finding groups: k-meansRelated: Supervised learning, k-means

An exploration strategy that scores each option by its average reward plus a bonus that is large for rarely tried options, then picks the highest score: optimism about what is uncertain.

Taught in Explore or exploit: banditsRelated: Explore or exploit tradeoff, Thompson sampling, Regret

A second held-back set used while building a model to choose settings such as the learning rate or model size, so the test set stays untouched until the final check.

Taught in Overfitting and generalizationRelated: Test set, Hyperparameter

The return an agent can expect from a state (or from a state and an action) when it follows a given policy from then on. Values let an agent compare choices whose payoff comes later.

Taught in States, actions, and rewardsRelated: Bellman equation, Action value, Return

A planning method that starts with all values at zero and repeatedly updates every state with the Bellman equation until the values stop changing. Needs a known model of the world.

Taught in States, actions, and rewardsRelated: Bellman equation, Value function, Policy

A gradient that shrinks toward zero as it is passed back through many steps or layers, because it is multiplied by many factors smaller than 1. Early steps then barely learn.

Taught in Before transformers: recurrent networksRelated: Exploding gradient, Backpropagation through time, Long short-term memory, Gradient

An autoencoder whose encoder outputs a probability distribution instead of a single code, trained with a KL term that pulls codes toward a standard normal prior, so any point near the center decodes to something plausible.

Taught in Autoencoders and latent spacesRelated: Autoencoder, KL divergence, Prior, Latent space, Diffusion model

An ordered list of numbers. It can also be read as an arrow, or a point, in a space with one dimension per number.

Taught in Vectors and similarityRelated: Embedding, Dot product, Cosine similarity

Adding, subtracting, and scaling embeddings, as in king - man + woman, to test whether a relationship between meanings shows up as a consistent direction.

Taught in Word embeddingsRelated: Embedding, Cosine similarity

A database built to store embeddings and quickly find the ones most similar to a query vector, usually with an approximate nearest neighbor index.

Taught in Search by meaningRelated: Semantic search, Retrieval-augmented generation

Replacing each vector by the nearest entry in a learned list (a codebook), so it can be stored or predicted as one whole number. Image tokenizers use it to turn each small square of a picture into one code.

Taught in Pictures, one token at a timeRelated: Image token, Autoencoder, Quantization

Rewards computed by an automatic check instead of a learned reward model, such as whether a math answer matches the known answer or whether code passes its tests.

Taught in Rewarding LLMs, and reward hackingRelated: Reward model, Reinforcement learning

The part of a multimodal model that turns a picture into a sequence of vectors, usually a vision transformer giving one vector per square patch. Janus-Pro's SigLIP encoder gives 576 vectors for a 384 by 384 picture.

Taught in How a model reads a pictureRelated: Vision transformer, Vision-language model, Projector, Contrastive learning

A transformer that reads an image by cutting it into square patches, turning each patch into a token vector, and attending across all the patches at once. Often shortened to ViT.

Taught in How computers seeRelated: Transformer, Self-attention, Convolutional neural network, Receptive field

A language model that also reads pictures: a vision encoder turns the picture into vectors, a small projector maps them into the language model's input space, and the model reads them alongside the text to answer in words.

Taught in How a model reads a pictureRelated: Vision encoder, Projector, Multimodal model, Language model

The fixed list of tokens a tokenizer and model know. Each token's position in the list is its id.

Taught in TokensRelated: Token, Tokenizer, Byte-pair encoding

The part of a speech system that turns a compact description of sound, such as a mel spectrogram or a learned code, into an audio waveform, filling in detail the description leaves out, such as phase.

Taught in Text to speechRelated: Text to speech, Spectrogram

A signal a model hides in what it generates, such as a statistical preference for certain words or a pattern in pixels, that a detector with the right key can test for later.

Taught in Safety in AI productsRelated: Content credentials, Generative AI

A number that sets how much one input matters to a neuron, and in which direction. Positive weights push the output up, negative weights push it down.

Taught in The artificial neuronRelated: Artificial neuron, Bias (neuron), Parameter

The standard score for speech recognition. Count the fewest word substitutions, deletions, and insertions that turn the transcript into the correct text, and divide by the number of correct words. Lower is better.

Taught in Speech to textRelated: Speech recognition

Quantization that maps a block's actual range, from its minimum to its maximum, onto the integers, with an integer offset marking where zero lands. Suits values that are not centered on zero.

Taught in Quantization: smaller numbersRelated: Quantization, Absmax quantization

Doing a task without any examples of that specific task during training or in the prompt, for example classifying images into categories the model was never trained to label.

Taught in One space for images and textRelated: Contrastive learning