Glossary
What each term means in this course, and where it is taught
This page is a reference, not a lesson. It answers one question: “I met this term somewhere in the course; what does it mean here?” Each entry gives a plain definition, the distinction most likely to cause confusion, and the page that teaches the concept. Definitions follow the lessons; where a term is used differently elsewhere in machine learning or statistics, the entry says so.
A
Activation function
The fixed nonlinear function a unit applies after its weighted sum, such as the logistic sigmoid, which squashes any real number into the range between 0 and 1. Without it, stacked layers would still compute an affine function of the input, so it is what lets a network represent nonlinear relationships. The word “activation” is also used for the value a unit outputs on a particular input; that value is computed from the input, not learned, and is not a parameter.
Taught in Neurons, Layers, and Learning §4. See also affine function, unit.
Adam
An optimizer that adapts the step size for each parameter using running averages of past gradients; the original Transformer was trained with it. [S-42] Like plain gradient descent, it consumes the gradients that backpropagation computed and performs the parameter update.
Taught in How Pretraining Works §4. See also optimizer.
Affine function
A weighted sum plus a constant. Composing affine functions gives another affine function, however many are stacked, which is why a network needs a nonlinearity between layers. [S-15]
Taught in Neurons, Layers, and Learning §3.
Agent
In this course, a system in which software repeatedly calls a model with an updated context, executes the tool requests the model generates under permission rules, and returns the observations for the next call, so that the model directs its own sequence of steps and tool use within constraints. [S-56] An agent is not a new kind of neural network: the model inside is an ordinary transformer, and the agentic behavior comes from the loop, the tools, the state, and the controls around it. The word has no universally fixed definition; a system whose path is mostly written in advance is better called a workflow. [S-50]
Taught in From Assistant to Agent §4–§6. See also agent loop, workflow.
Agent loop
The cycle that runs an agent: assemble a context from the objective and the current state; call the model; receive either a tool request or a response; check a request against permissions; execute it; place the observation into the context; call the model again; continue, ask the user, or stop. The model does not keep running between calls; continuity is supplied by the context and state the software maintains.
Taught in From Assistant to Agent §5.
Argmax
The single highest-scoring candidate. Selecting the next token by argmax is deterministic and introduces no variation; softmax with the temperature approaching zero converges to it when the top score is unique. [S-10] Deterministic decoding can still produce unsupported text.
Taught in Softmax, Sampling, and Temperature §4 and §6. See also sampling, decoding.
Assistant
The product a person actually types to: a post-trained model plus a system prompt, management of the conversation state, often retrieval and tools, safety checks, and other product logic. It is not identical to the model. Much of its behavior can be changed with no training run at all, by changing what enters the context.
Taught in Post-training §9. See also model, system prompt.
Attention
See self-attention. When this course says “attention” without qualification it means causal self-attention inside a transformer block, one mechanism among several, not the transformer as a whole.
Attention head
One instance of the self-attention computation, with its own learned query, key, and value matrices. A transformer layer runs several heads in parallel so that several attention patterns operate at once over the same positions; this is multi-head attention, and the heads’ outputs are concatenated and passed through one more learned matrix. [S-37] Nothing in the architecture or the training objective assigns a head a fixed, human-readable job. [S-45]
Taught in How Self-Attention Moves Information §6.
Attention weight
The nonnegative number, summing to one across the permitted source positions, that sets how much of each source’s value vector is added into a destination position, in one head at one layer. It is a local mixing weight. It is not a probability that one token caused another, and it is not by itself an explanation of the model’s output, which depends on every head, every layer, and the feed-forward sublayers. [S-45]
Taught in How Self-Attention Moves Information §3 and §8. Not to be confused with a weight in the sense of a parameter.
Autoregressive model
A model that predicts each token from the tokens before it, so that the probability of a whole sequence is the product of one-step conditional probabilities. [S-40] Generation appends each selected token to the context and predicts the next; training evaluates every position of a known sequence in one forward pass under the causal mask.
Taught in How Pretraining Works §1–§2.
B
Backpropagation
The procedure that computes the gradient of the loss with respect to every parameter, by applying the chain rule backward from the loss through the network in one backward sweep, reusing values saved from the forward pass. It computes gradients and does nothing else; a separate optimizer step uses those gradients to change the parameters. [S-14]
Taught in Neurons, Layers, and Learning §9. See also optimizer.
Base model
A model that has been pretrained but not post-trained. Given an instruction it may answer it, continue it, or produce more text like it, because all of those follow such text somewhere in the training data.
Taught in Post-training §1. See also assistant.
Bias
A parameter added to a unit’s weighted sum regardless of the inputs, shifting the result up or down. In this course the word always means this parameter, not a statistical or social bias.
Taught in Neurons, Layers, and Learning §2–§3.
Big-O notation
Shorthand for an asymptotic upper order of growth as the input size \(n\) increases. In the comparisons used here, \(O(n)\), \(O(n^2)\), and \(O(1)\) stand for linear-order, quadratic-order, and constant-order scaling. The course uses it to compare recurrent and self-attention layers.
Taught in Why Transformers §1.
Byte pair encoding (BPE)
A procedure for building a subword vocabulary: start from single characters or bytes and repeatedly merge the most frequent adjacent pair into a new vocabulary entry, so that frequent sequences can become larger vocabulary units while rarer forms can be represented from shared pieces. Adapted from a data-compression algorithm; it sits between character-level and byte-level units and whole words, and many current tokenizers use it or a variant, often with a byte-level base. [S-27] [S-28]
Taught in Tokens, Embeddings, and the Context Window §2–§3.
C
Causal mask
The rule, applied inside self-attention before softmax, that gives zero weight to positions later than the current one, so a position can use itself and earlier positions but never later ones. [S-22] It preserves next-token prediction during training, when the whole sequence is present, and it is what lets every position’s loss be computed in one forward pass.
Taught in How Self-Attention Moves Information §4 and How Pretraining Works §2.
Compaction
Compressing or summarizing accumulated context into a smaller representation so that work can continue with less context. Some systems perform this automatically as the context nears its limit. [S-61] A handoff is a deliberate, human-written state summary that serves a related purpose; the two are not the same operation.
Taught in Using Models Well §3. See also handoff.
Context
Everything the model conditions on for one call: the prompt, instructions, the conversation so far, retrieved text, tool definitions and tool results, supplied files, and the tokens generated so far. All of it is tokens in one finite sequence, and none of it changes a parameter. The course also calls it inference-time context, to keep it apart from training. [S-31]
Taught in Tokens, Embeddings, and the Context Window §8–§9 and How Pretraining Works §8. See also context window, parameter.
Context engineering
Managing what goes into the context at each step: the smallest set of high-signal tokens sufficient for the next step, loaded when a step needs it rather than all at the start, with stale or duplicated material removed. [S-34] Named practices include just-in-time loading, progressive disclosure, compaction, and handoffs. [S-61]
Taught in Using Models Well §3.
Context window
The finite sequence of tokens a model can reference when producing output, including the output produced so far; its size is a fixed maximum number of tokens that differs by model. [S-31] It is distinct from the parameters and from the training data: each model step receives a finite context supplied or managed by the surrounding system, and products may persist information externally and supply it again on later steps or requests.
Taught in Tokens, Embeddings, and the Context Window §8.
Contextual representation
The vector at a position after the network’s layers have updated it using the other positions. Unlike the token’s embedding, which is the same wherever the token occurs, it depends on the surrounding text: the token bank in “river bank” and in “central bank” starts from the same embedding and ends with different representations.
Taught in Tokens, Embeddings, and the Context Window §7. See also embedding.
Cross-entropy
The loss that measures how little probability a predicted distribution gives to the observed outcome. For a softmax distribution over a vocabulary it is the same quantity as the negative log-likelihood of the observed next token.
Taught in Neurons, Layers, and Learning §7 and How Pretraining Works §3. See also negative log-likelihood.
D
Decoder and encoder
Two ways of using self-attention. An encoder lets every position attend to every other, to represent a complete input; a decoder applies the causal mask so that each position attends only to itself and earlier positions, which is what an autoregressive language model needs. [S-22] [S-24] When this course says “self-attention” without qualification it means the causal, decoder form.
Taught in How Self-Attention Moves Information §4 and Why Transformers §3.
Decoding
The rule that selects the next token from the probability distribution the model produces: deterministically, by argmax, or stochastically, by sampling. Decoding settings such as temperature change the output without changing any parameter.
Taught in Softmax, Sampling, and Temperature §6. See also argmax, sampling.
Direct preference optimization (DPO)
A preference-based post-training method that updates the language model directly from comparison pairs, using a loss that resembles ordinary classification, given a fixed preference dataset. It removes the separate reward model and the online reinforcement-learning loop of the classic RLHF pipeline. [S-47] Both routes remain in use.
Taught in Post-training §5. See also reinforcement learning from human feedback.
Dot product
The sum of the products of corresponding entries of two vectors of the same length, giving one number. Self-attention uses it to compare a query with each key.
Taught in How Self-Attention Moves Information §2–§3.
E
Embedding
The learned vector that a token ID selects from the embedding table, a matrix with one row per vocabulary entry. The entries are parameters, adjusted during training so that prediction improves. [S-29] They are not dictionary definitions, and nearness of two rows reflects statistical relationships in the training text rather than a property anyone specified.
Taught in From Text to Numbers §5 and Tokens, Embeddings, and the Context Window §6. See also contextual representation, token ID.
External state
Everything the software or the tools persist outside the model: files an agent writes, a repository, database rows, saved notes, stored summaries. It can be read back into a later context, in this task or a future one; doing so changes no parameter.
Taught in From Assistant to Agent §7. See also state, memory.
F
Feed-forward layer (MLP)
One of the main computations in a transformer block: two linear transformations with a nonlinearity between them, applied to each position separately and identically, with no exchange between positions. [S-38] [S-19] Named architectures arrange and order the block’s sublayers differently.
Taught in How Self-Attention Moves Information §7. See also transformer block.
Fine-tuning
Further training of an already-trained model on new data, changing its parameters. Supervised instruction tuning is one form of it; work on pretraining and transfer showed that one pretrained model could be fine-tuned for many different tasks. [S-24]
Taught in Post-training §3 and Why Transformers §5.
Forward pass
Running the network end to end on a given input with the current parameters to produce an output. During training it produces the predictions the loss is computed from; backpropagation then works backward through the same computation.
Taught in Neurons, Layers, and Learning §6.
G
Gradient
For each parameter, how fast the loss changes for a small change in that parameter with the others held fixed. Computed by backpropagation; used by the optimizer.
Taught in Neurons, Layers, and Learning §8.
Gradient descent
The update rule that moves the parameter vector a small step, scaled by the learning rate, in the negative-gradient direction, the direction of steepest local decrease. With a sufficiently small step this is locally downhill; an optimizer need not update every parameter by a nonzero amount. Pretraining runs this loop an enormous number of times on mini-batches, usually with a more elaborate optimizer.
Taught in Neurons, Layers, and Learning §8 and §10, and How Pretraining Works §4.
H
Handoff
A compact state summary written to carry work into a fresh session or a different model: the objective, the current state, the decisions already settled, the source of truth, the open issues, the next bounded unit, and what must not change. It replaces the transcript rather than preserving it.
Taught in Using Models Well §8. See also compaction.
I
In-context (few-shot) behavior
A model performing a task from demonstrations or instructions placed in its input, without any parameter update. GPT-3 was reported to do this across many tasks, with the ability strengthening with scale. [S-26] The course describes it as observed task performance, not as learning in the training sense.
Taught in Why Transformers §5.
Inference
Using a trained model, with its parameters fixed, to produce output from a context. Request-specific information the model conditions on is supplied through inference-time context; decoding settings and surrounding product logic can also affect the resulting output without changing model parameters. Training and post-training are the only stages that change parameters.
Taught in How Pretraining Works §8–§9.
Inference-time context
See context.
Instruction tuning
K
Key
In self-attention, the vector each source position is transformed into for comparison with the destination’s query; the dot product of query and key gives the compatibility score. [S-36] One of three learned transformations, with the query and the value.
Taught in How Self-Attention Moves Information §2–§3. See also query, value.
Key/value caching
Storing the keys and values computed for earlier positions during token-by-token generation and reusing them at each step instead of recomputing them. [S-43] An efficiency measure that leaves the mathematics of autoregression unchanged; cached tokens still occupy the context window.
Taught in How Pretraining Works §9.
L
Language model, large language model (LLM)
A model that assigns probabilities to sequences of tokens; the ones in this course do so autoregressively, predicting each next token from the ones before it. “Large” adds only scale. In the course’s vocabulary hierarchy, the LLM is the trained model, a transformer is its architecture, and self-attention and the feed-forward layer are two distinct mechanisms inside each transformer block. [S-19]
Taught in The Short Story and Why Transformers.
Layer
A set of units computed in parallel from the same input, whose outputs form a vector that is the input to the next layer. [S-12] In a transformer the word is also used for the repeated blocks and for the attention and feed-forward sublayers inside them.
Taught in Neurons, Layers, and Learning §5.
Learning rate
The small step size that scales each gradient-descent update.
Taught in Neurons, Layers, and Learning §8.
Logit
In this course, and in machine-learning usage generally, the raw pre-softmax score the network produces for each candidate token. This is not the same object as the statistical logit function \(\log\big(p/(1-p)\big)\), which takes a probability and returns a real number: an individual multiclass score \(z_i\) is not \(\operatorname{logit}(p_i)\). The exact relationship holds only for differences, \(\log(p_i/p_j) = (z_i - z_j)/T\). [S-09]
Taught in Softmax, Sampling, and Temperature §3–§4.
Loss
A single number measuring how poor the model’s current output is under the training criterion; training adjusts the parameters to make it smaller. For a language model the loss at a position is the negative log of the probability given to the observed next token.
Taught in Neurons, Layers, and Learning §7 and How Pretraining Works §3.
M
Maximum likelihood
Choosing parameters that maximize the probability the model assigns to the observed training data; minimizing the summed negative log-likelihood is the same thing. [S-05] Pretraining is maximum-likelihood training, but it is not a conventional econometric exercise aimed at identifying interpretable structural parameters or reporting classical standard errors. [S-06]
Taught in How Pretraining Works §3.
Memory
An ambiguous word that should be qualified whenever it is used. Three different things get called memory: the model’s parameters, which only training changes; the current context, which lasts one call; and persistent external state, which software stores and supplies again later. An assistant that “remembers” a project has been given a file or a store that is read back into its context; no parameter was updated.
Taught in From Assistant to Agent §7 and How Pretraining Works §8. See also external state.
Mini-batch
A small set of training sequences drawn from the training text for one optimization step. Its gradient is a noisy estimate of the gradient over all the data, which is why the procedure is called stochastic optimization.
Taught in How Pretraining Works §4.
Model
In this course, a parameterized neural network that maps inference-time context to a probability distribution over the next token. Pretraining and post-training change it. A system prompt, retrieval, tools, and product logic change the assistant built around it without touching it.
Taught in Post-training §9. See also assistant.
Multi-head attention
See attention head.
N
Negative log-likelihood
The loss \(-\log p(x_t \mid x_{<t})\): the negative logarithm of the probability the model gave to the token that actually came next. It is zero when that probability is 1 and rises without bound as the probability approaches zero. Summed over positions, it equals the negative log of the whole sequence’s probability.
Taught in How Pretraining Works §3. See also cross-entropy, maximum likelihood.
Neural network
A function built from many units, each computing a weighted sum plus a bias followed by a nonlinearity, arranged in layers, with trainable parameters adjusted by training. [S-12] A transformer is one neural-network architecture; a language model is a trained instance of one.
Taught in Neurons, Layers, and Learning §5 and §11.
Nonlinearity
See activation function.
Normalization
A step inside each transformer block that rescales the vectors so that their values stay in a workable range through many layers. Where it sits relative to the residual connection differs between named architectures: after the residual sum in the original Transformer, before each sublayer in GPT-2. [S-38] [S-39]
Taught in How Self-Attention Moves Information §7.
O
Optimizer
The rule that uses the gradients backpropagation computed to update the parameters: plain gradient descent, or an adaptive rule such as Adam. The optimizer performs the update; backpropagation does not. [S-14]
Taught in Neurons, Layers, and Learning §9–§10 and How Pretraining Works §4.
P
Parameter
An adjustable number inside the model whose value is set by training: weights, biases, embedding-table entries, attention projection matrices, normalization scales, and the final output projection. Parameters change only during training and post-training; they are fixed at inference time. A value computed for a particular input, such as an activation or an attention weight, is not a parameter.
Taught in Neurons, Layers, and Learning §1 and How Pretraining Works §4.
Permissions
The rules in the software around a model that decide which requested actions run automatically, which wait for approval, and which are refused. They are enforced by the loop regardless of what the model generates, and a well-behaved model is not a substitute for them. [S-54]
Taught in From Assistant to Agent §10.
Policy
In reinforcement learning, the thing being optimized. In the post-training methods this course describes, the policy is the language model itself, whose parameters are updated to raise a reward computed on what it produces.
Taught in Post-training §5 and §7.
Positional information
Whatever mechanism lets a transformer tell the order of positions apart, since the token embeddings alone carry none and self-attention has no built-in notion of order. The original Transformer added a positional encoding to each embedding; other architectures supply position information in other ways. [S-30]
Taught in From Text to Numbers §6.
Post-training
Additional parameter-updating training applied after pretraining to shape a model’s behavior, capabilities, preferences, or safety characteristics. It is a family of methods, including supervised instruction tuning, preference-based methods (through a reward model and reinforcement learning, or directly), model-generated feedback, and reinforcement learning against checkable rewards; RLHF names one route, not the family. [S-55] A system prompt is not post-training.
Taught in Post-training §2.
Preference data
Judgments about which of several candidate responses to a prompt is preferred, collected from people, from models, or from rules, and used by preference-based post-training. Different from a demonstration, which supplies a single target response.
Taught in Post-training §4–§6.
Pretraining
Large-scale training on the next-token objective over a very large body of text, updating the model’s trainable parameters. [S-05] The targets come from the text itself. The objective rewards assigning higher probability to the observed continuation; independent truth verification is not part of the objective. [S-08]
Taught in How Pretraining Works.
Prompt
The text a user supplies for one request; part of the context. This course treats a prompt as a task specification: the objective, the source of truth, the inputs, the constraints, the required output, and the acceptance criteria.
Taught in Using Models Well §2.
Prompt caching
A provider feature that can reuse computation for a repeated input prefix, reducing processing cost or latency on later calls. In the implementation discussed in the course, cached input still occupies logical context capacity: caching changes what the tokens cost, not whether they count. [S-62] Pricing and caching behavior differ between providers.
Taught in Using Models Well §11.
Q
Query
In self-attention, the vector the destination position is transformed into, compared by dot product with each permitted source’s key. [S-36] It is a learned projection, not a question; reading it as “what the destination is looking for” is an intuition only.
Taught in How Self-Attention Moves Information §2–§3. See also key, value.
R
Reasoning tokens (intermediate reasoning)
Tokens a model generates before its final answer. On some difficult tasks, generating intermediate reasoning tokens provides additional inference-time computation and can improve performance; that is the supported statement. [S-49] They are output produced by the same next-token mechanism as any other output, not a transcript of internal computation and not thought in the human sense.
Taught in Post-training §7.
Recurrent network
A network that reads a sequence one position at a time, carrying a running summary, its hidden state, forward from each step to the next. Relating distant positions takes many sequential steps, and the computation cannot be parallelized within a sequence; that is what self-attention changed. [S-16] [S-21]
Taught in Why Transformers §1–§2.
Reinforcement learning
An optimization framework in which behavior is updated using a reward computed on what the model produced, rather than by matching a fixed target string. In post-training it is used with a learned reward model, or with rule-based rewards on tasks whose outcomes can be checked.
Taught in Post-training §5 and §7.
Reinforcement learning from AI feedback (RLAIF)
Preference-based training in which a model, rather than a person, judges which of two responses is better under written principles, and a preference model trained on those judgments supplies the reward. Constitutional AI is the published example. [S-48]
Taught in Post-training §6.
Reinforcement learning from human feedback (RLHF)
One route through preference-based post-training: train a separate reward model on human comparisons, then update the language model by reinforcement learning to score well under it while staying close to its starting point. [S-46] Some writers use “RLHF” for any preference-based post-training; this course uses it only for this route, and never as a name for post-training as a whole. [S-55]
Taught in Post-training §5. See also direct preference optimization, reward model.
Residual connection
Adding each sublayer’s input to its output inside a transformer block, so that the sublayer computes an update to the representation rather than a replacement for it. [S-38]
Taught in How Self-Attention Moves Information §7.
Retrieval
Fetching relevant text from an external collection, a search, a file, or a store and placing it in the context for the current call, so that generation combines the model’s fixed parametric knowledge with information supplied at request time. [S-33] It changes the inputs to this request, not the parameters. The published name for one such design is retrieval-augmented generation.
Taught in Tokens, Embeddings, and the Context Window §10 and From Assistant to Agent §7.
Reward model
A separate network, trained on comparison data, that takes a prompt and a response and outputs a score predicting how a rater would rank it; used as the reward signal in the classic RLHF pipeline. It is an imperfect stand-in for judgment, which is why the model being updated is kept close to its starting point.
Taught in Post-training §5.
S
Sampling
Drawing the next token at random according to the probability distribution the model produced, so that the same prompt can yield different continuations without any change to the model. [S-07] Contrast argmax.
Taught in Softmax, Sampling, and Temperature §6.
Scaling laws
The empirical regularity, reported for language models in 2020, that loss improves as a power law with model size, dataset size, and training compute over a wide range. [S-25] An observed regularity for the studied setup, not a law of nature.
Taught in Why Transformers §5.
Self-attention
The mechanism inside each transformer block that lets a position update its vector using other positions: each destination’s query is compared with each permitted source’s key, softmax turns the scores into weights, and the destination receives the weighted sum of the sources’ values. [S-36] It is one of two main computations in the block, not the whole transformer.
Taught in How Self-Attention Moves Information §2–§3. See also transformer block.
Skill
A reusable set of instructions, procedures, examples, or resources describing how to carry out a class of tasks, loaded into inference-time context when relevant. [S-53] It changes what the model conditions on for a task and leaves the parameters unchanged; it is not retraining.
Taught in From Assistant to Agent §8.
Softmax
The fixed function that turns a list of real-valued scores into a probability distribution: exponentiate each score, divided by the temperature, and divide by the sum. Every output is positive, the order of the scores is preserved, and adding the same constant to all scores changes nothing. [S-01] Used at the output to obtain next-token probabilities and inside attention to obtain attention weights.
Taught in Softmax, Sampling, and Temperature §4.
Source of truth
Where the answer to a question should come from: the model’s general knowledge, a supplied document, current documentation, a database, running code. Deciding it before asking tells you what belongs in the context and whether a tool is needed.
Taught in Using Models Well §4.
State
What carries a task from one model call to the next. Three kinds are easy to run together: the parameters, fixed during a task; the context window, reassembled every call; and external state, persisted by software or tools.
Taught in From Assistant to Agent §7.
Supervised fine-tuning (SFT)
Post-training on pairs of a prompt and a demonstrated desired response, using the same next-token loss as pretraining on the response tokens; called instruction tuning when the pairs are instructions and answers. [S-46] It raises the probability of responses that resemble the demonstrations; it does not guarantee compliance or correctness.
Taught in Post-training §3.
System prompt
Instructions a product places in the context in front of the conversation, often carrying a persona and rules. It is inference-time context: it changes what the model conditions on and no parameter, so it is never a form of post-training.
Taught in Post-training §2 and §9, and Tokens, Embeddings, and the Context Window §9.
T
Temperature
A setting \(T\) that divides every score before softmax. Lower values concentrate probability on the top-scoring tokens and approach argmax as \(T \to 0\); higher values flatten the distribution toward uniform. [S-04] It changes variability, not truth, and leaves the parameters untouched.
Taught in Softmax, Sampling, and Temperature §6.
Token
A unit of text drawn from a fixed vocabulary: often a whole word, but also a word fragment, punctuation, or another piece the tokenizer defines. Not the same as a word; in many commonly used tokenizers a leading space or a capital letter changes the token.
Taught in From Text to Numbers §3 and Tokens, Embeddings, and the Context Window §4.
Token ID
The integer index of a token in the vocabulary. An address, not a measurement: the network never adds or compares IDs, and numerically close IDs are not similar tokens.
Taught in From Text to Numbers §4.
Tokenizer
The program that splits text into tokens according to a vocabulary and rules decided before training. Different model families use different tokenizers, so the same text can be split differently by each.
Taught in From Text to Numbers §3.
Tool
A capability the software around a model exposes to it: a name, a description, and an input schema that the model reads, plus an implementation that the software runs.
Taught in From Assistant to Agent §2.
Tool call
At the model boundary, generated output in a structured format requesting a tool with specified inputs. The surrounding software interprets that output, checks it against permissions, executes or routes the action, and returns the result to a later model call; the model does not itself run anything. [S-51]
Taught in From Assistant to Agent §2.
Transformer
A neural-network architecture built from repeated blocks, each containing self-attention and a feed-forward layer, with residual connections and normalization. Transformer architectures are the dominant foundation for many modern large language models; a transformer is not the beginning of neural networks, and it is not the same thing as attention. [S-19]
Taught in Why Transformers and How Self-Attention Moves Information §7.
Transformer block
The repeating unit of a transformer: an attention sublayer that moves information between positions, a feed-forward sublayer that computes on each position separately, plus residual connections and normalization. [S-38] A model stacks many blocks.
Taught in How Self-Attention Moves Information §7.
U
Unit (neuron)
The basic computation in a neural network: a weighted sum of the inputs plus a bias, followed by a nonlinearity. [S-12] Often called a neuron; the course does not rely on the biological analogy.
Taught in Neurons, Layers, and Learning §5.
V
Value
In self-attention, the vector each source position is transformed into for transfer: the destination receives the weighted sum of the values, weighted by the attention weights. [S-36] It is separate from the key, so that matching and transfer can be learned separately.
Taught in How Self-Attention Moves Information §2–§3. See also query, key.
Vector
An ordered list of real numbers of a fixed length. Embeddings, queries, keys, values, and representations are all vectors.
Taught in From Text to Numbers §5.
Verification
Making success observable rather than asserted: running the tests, recomputing a number, rendering the document, reading the diff. A passing test provides evidence that the specified checks passed. It does not establish that the specification, research design, or resulting number is correct.
Taught in From Assistant to Agent §9 and Using Models Well §6.
Vocabulary
The fixed inventory of tokens a model uses, decided before training and then never changed; typically tens of thousands of entries (GPT-2’s had 50,257). [S-28]
Taught in From Text to Numbers §3.
W
Weight
A parameter that multiplies an input in a unit’s weighted sum; more generally, any entry of a learned matrix. Not to be confused with an attention weight, which is computed for a particular input rather than learned.
Taught in Neurons, Layers, and Learning §2–§3. See also attention weight.
Workflow
A system in which the model calls and tool uses follow code paths written in advance by a programmer; the model fills in each step but does not choose the steps. [S-50] Contrast agent. It is often the simplest system that reliably solves a task.
Taught in From Assistant to Agent §4. See also agent.