A small, evolving reference for terms that show up repeatedly across the blog. Inline glossary terms open a compact definition card, while each term here has its own entry page for longer notes and future expansion.
Glossary
A/A Test
An experiment whose two arms serve the identical experience, used to audit the experimentation platform itself: a healthy system should reject the null about 5% of the time.
Glossary
A/B Testing
Comparing two versions of a system by randomly assigning users to each and measuring which performs better on a chosen metric, so changes are judged on evidence rather than opinion.
Glossary
Action Chunking
Having a robot policy predict a short sequence of future actions at once instead of one action per timestep, which reduces the number of closed-loop decisions and tames compounding error.
Glossary
Activation Patching
A causal interpretability technique that transplants internal activations from one forward pass into another to identify which components actually drive a behaviour.
Glossary
Activation Steering
Modifying a model's behaviour at inference time by adding, amplifying, or suppressing concept directions in its internal activations rather than changing its weights or prompt.
Glossary
Active Learning
A training approach where the system selectively chooses which examples should be labeled next.
Glossary
Active Parameters
The parameters a sparse model actually computes with for one token, as opposed to the total it stores; the compute-cost side of the "X total, Y active" description of mixture-of-experts models.
Glossary
Actor-Critic
A reinforcement learning architecture that pairs a policy (the actor) that chooses actions with a learned value function (the critic) that evaluates them, using the critic's feedback to reduce the variance of policy updates.
Glossary
Advantage Function
The gap A(s, a) = Q(s, a) - V(s) between the value of a specific action and the average value of the state, measuring how much better an action is than the policy's usual behaviour.
Glossary
Agentic Loop
The cycle at the heart of every agent turn, where the agent gathers context, takes an action, checks the result, and repeats until a stop condition is met.
Glossary
Arithmetic Intensity
The number of floating-point operations a workload performs per byte of data it moves, which determines whether it is compute-bound or memory-bound.
Glossary
Artificial General Intelligence
A hypothetical AI system with the broad, flexible cognitive ability of a human across essentially any intellectual task, rather than competence confined to a single narrow domain.
Glossary
Attention
A neural mechanism that lets a model weigh which parts of the input matter most when processing or predicting.
Glossary
Attribution Graph
A graph tracing which internal features causally feed which others on a specific prompt, reconstructing the step-by-step circuit a model used to produce an output.
Glossary
Autoencoder
A model that learns a compressed representation of its input and then tries to reconstruct the original input from it.
Glossary
Average Treatment Effect
The mean difference between potential outcomes with and without treatment across a population, E[Y(1) - Y(0)]; the headline quantity most causal analyses target.
Glossary
Bayesian Optimization
A search strategy that uses a probabilistic model of past evaluations to choose where to sample an expensive function next.
Glossary
Behavior Cloning
Training a policy by supervised learning on expert demonstrations, fitting the mapping from observed states to the expert's actions by maximum likelihood.
Glossary
Best Linear Unbiased Prediction
The optimal estimate of a random effect given the data; a shrinkage estimate that pulls each group's raw mean toward the population mean by a precision-determined factor.
Glossary
CUPED
A variance-reduction technique that subtracts the part of an experiment metric predictable from pre-experiment data, shrinking variance by the squared correlation without biasing the effect estimate.
Glossary
Catastrophic Forgetting
The abrupt loss of previously learned knowledge when a neural network is trained sequentially on new data, because gradient updates for the new task overwrite shared weights the old tasks relied on.
Glossary
Causal Inference
The discipline of estimating what would happen under an intervention, rather than what is merely associated in observed data, using assumptions about how the data came to be.
Glossary
Censoring
A partially observed outcome, most commonly right-censoring, where an observation ends before the event of interest occurs, so the true lifetime is known only to exceed the observed duration.
Glossary
Chain of Thought
An intermediate sequence of reasoning steps a model writes out before its final answer, which improves accuracy on multi-step tasks.
Glossary
Churn
The event of a customer ending their relationship with a business, and by extension the rate at which a customer base loses members over a period.
Glossary
Classifier-Free Guidance
A sampling technique that sharpens a diffusion model's adherence to its conditioning by extrapolating between its conditional and unconditional predictions.
Glossary
Coding Agent
An LLM-powered tool that can read a codebase, run commands, edit files, and iterate against feedback in a loop until a task is complete, rather than only suggesting snippets for a human to apply.
Glossary
Collaborative Filtering
A recommendation approach that predicts preferences from patterns in many users' interactions rather than item content alone.
Glossary
Collider
A variable caused by both treatment and outcome; conditioning on it creates spurious association where none existed.
Glossary
Compute-Optimal Scaling
The Chinchilla result that for a fixed training compute budget, loss is minimized by scaling parameters and data together, at roughly 20 training tokens per parameter.
Glossary
Conditional Average Treatment Effect
The average treatment effect among units sharing covariates x, written tau(x); the quantity uplift models estimate to decide whom to target.
Glossary
Confounder
A variable that causes both the treatment and the outcome, opening a backdoor path that makes naive comparisons biased.
Glossary
Container
A lightweight, isolated process that packages an application with its dependencies and shares the host operating system's kernel.
Glossary
Container Image
A read-only, layered package of application code, dependencies, and configuration that serves as the blueprint for running containers.
Glossary
Context Rot
The empirical degradation of an LLM's recall and reasoning as its context window fills, long before the advertised token limit is reached.
Glossary
Context Window
The bounded working memory of an LLM, holding every instruction, file, and tool output the model can currently see. Model performance degrades as it fills, so managing it deliberately is a core skill in agentic work.
Glossary
Continual Learning
The ability of a model to keep learning from a stream of new tasks and experiences after deployment, accumulating skills without forgetting old ones or requiring retraining from scratch.
Glossary
Continuous Batching
A serving scheduler that admits and retires requests at every decode step, keeping the running batch full instead of waiting for a whole batch to finish.
Glossary
Corpus
A structured collection of text or other examples that you analyze, search, or use to train and evaluate a model.
Glossary
Cosine Similarity
A measure of how aligned two vectors are, commonly used to compare embeddings by direction rather than raw magnitude.
Glossary
Counterfactual Evaluation
Estimating how a system would have performed under decisions that were not actually taken.
Glossary
Cross-Validation
An evaluation technique that repeatedly splits the data into training and validation folds so every observation is used for both fitting and scoring.
Glossary
Customer Lifetime Value
The expected total (usually discounted) margin a customer generates over their relationship with a business, combining a survival model of how long they stay with a revenue model of what they pay.
Glossary
Data Parallelism
Training parallelism where every chip holds a copy of the model and processes a different slice of the batch, synchronizing gradients once per step.
Glossary
Delta Method
A first-order Taylor approximation that carries a variance through a smooth function of noisy quantities, giving standard errors for transformed estimates and ratios of means.
Glossary
Difference-in-Differences
A design comparing treated and untreated groups' changes over a treatment date, cancelling stable group differences and shared shocks under the parallel-trends assumption.
Glossary
Diffusion Model
A generative model that learns to reverse a gradual noising process, creating new samples by iteratively denoising pure noise.
Glossary
Directed Acyclic Graph
A graph of variables and one-way causal arrows with no cycles, used to encode causal assumptions and read off what to adjust for.
Glossary
Do-Operator
Pearl's notation do(T = t) for setting a variable by intervention rather than observing it, separating P(Y given do(T)) from the ordinary conditional P(Y given T).
Glossary
Domain Randomization
Training one policy across many randomly perturbed versions of a simulator (varied physics, visuals, latencies) so that the real world falls inside the training distribution.
Glossary
Double Machine Learning
A recipe for using flexible ML models in causal estimation, predicting both outcome and treatment from confounders, and regressing residual on residual with cross-fitting.
Glossary
Draft Model
A small, fast model that proposes candidate tokens for a larger target model to verify in bulk during speculative decoding.
Glossary
ELBO
The evidence lower bound, a tractable quantity that can be maximized in place of an intractable likelihood when training latent-variable models.
Glossary
Elastic Weight Consolidation
A continual-learning method that fights catastrophic forgetting by adding a quadratic penalty anchoring each weight near its old-task value, scaled by the Fisher information measuring how much the old task depends on that weight.
Glossary
Embedding
A dense numeric representation that places similar items close together in vector space.
Glossary
Expected Value
The average outcome you would expect from a decision after weighting each possible result by its probability.
Glossary
Expert Parallelism
Training and serving parallelism for mixture-of-experts models where different experts live on different chips and tokens travel to their assigned experts via AllToAll communication.
Glossary
Exponential Family
A class of probability distributions sharing one algebraic form, whose natural parameter determines the canonical link function of the corresponding GLM.
Glossary
Feature Engineering
The process of turning raw data into model-ready signals that better represent the behaviour or structure a model needs to learn.
Glossary
Fine-Tuning
The process of continuing training on a pre-trained model so it becomes better suited to a narrower task or domain.
Glossary
Fisher Information
A measure of how much a model's likelihood changes as a parameter moves, computed as the expected squared score; near a maximum-likelihood solution it equals the expected curvature of the negative log-likelihood.
Glossary
Fixed Effect
An ordinary regression coefficient treated as an unknown constant shared across all observations, in contrast to a random effect drawn from a distribution per group.
Glossary
FlashAttention
An exact, IO-aware implementation of attention that avoids materializing the full score matrix, cutting memory use and wall-clock time without changing outputs.
Glossary
Generalized Estimating Equations
A method for clustered data that fits the population-averaged (marginal) model directly, using a working correlation structure and robust standard errors instead of random effects.
Glossary
Generalized Linear Mixed Model
A GLM whose linear predictor contains random effects, giving hierarchical versions of logistic and Poisson regression; its likelihood involves an integral with no closed form, so fitting relies on approximations.
Glossary
Generalized Linear Model
A model family where a weighted sum of features predicts the mean of an exponential-family distribution through a link function; linear and logistic regression are its best-known members.
Glossary
Global Workspace Theory
A neuroscience theory holding that conscious access arises when information wins entry to a small-capacity broadcast system that shares it across the brain's specialized unconscious processors.
Glossary
Gradient Boosting
An ensemble technique that builds models sequentially, with each new model trained to correct the errors of the ensemble so far.
Glossary
Group Relative Policy Optimization
A reinforcement learning algorithm that drops PPO's value network and instead estimates advantages by comparing a group of sampled answers to the same prompt against their own average reward.
Glossary
Grouped-Query Attention
An attention variant where groups of query heads share a single key/value head, shrinking the KV cache with little quality loss.
Glossary
Guardrail Metric
A metric an experiment must not harm, tested for non-inferiority alongside the success metric and empowered to veto a launch regardless of wins elsewhere.
Glossary
Hazard Function
The instantaneous or per-period risk of an event occurring at time t, conditional on having survived up to t. High early hazard means early failures; spikes mark risky moments like contract expiries.
Glossary
High-Bandwidth Memory
Stacked DRAM packaged next to an accelerator's compute die, holding model weights and activations and feeding them to the chip at terabytes per second.
Glossary
Hyperparameter
A configuration setting of a learning algorithm, such as tree depth or learning rate, chosen before training rather than learned from the data.
Glossary
Imitation Learning
The family of methods that learn a policy from demonstrations of the desired behavior rather than from a reward signal, spanning behavior cloning, interactive methods like DAgger, and inverse RL.
Glossary
Inference
The stage where a trained model is used to make predictions or generate outputs on new inputs.
Glossary
Instrumental Variable
A variable that shifts treatment uptake but touches the outcome only through treatment, letting a ratio of two clean comparisons recover a causal effect despite unmeasured confounding.
Glossary
Inter-Token Latency
The time between successive output tokens once a response has started streaming, set by the memory-bound decode phase.
Glossary
Interpretable Model
A model whose behaviour can be understood directly enough for humans to inspect, validate, and act on its reasoning.
Glossary
Intraclass Correlation
The correlation between two observations from the same group, equal to the share of outcome variance that lives at the group level; it drives how much information clustered data really contains.
Glossary
Isotonic Regression
A regression method that fits the closest sequence or function that obeys a monotonic ordering constraint.
Glossary
J-Space
The sparse set of verbalizable concept patterns a language model holds in its middle-to-late layers, identified via the Jacobian lens and argued by Anthropic to function as a global workspace.
Glossary
Jacobian Lens
An interpretability technique that decodes intermediate activations by first transporting them to the output basis through the network's averaged input-output Jacobian, revealing what the model is disposed to say later.
Glossary
KL Divergence
A measure of how one probability distribution differs from another; asymmetric, so the direction you compute it in changes what an optimizer learns.
Glossary
KV Cache
A store of every past token's attention key and value vectors, reused at each generation step so the model never recomputes them.
Glossary
Kaplan-Meier Estimator
The classical nonparametric estimator of a survival function from censored data, built as a product of conditional survival fractions, one factor of (1 - d/n) per observed event time.
Glossary
Knowledge Distillation
Training a smaller model to imitate the outputs of a larger, stronger model so it inherits much of the capability at lower cost.
Glossary
Large Language Model
A very large language model, usually transformer-based, trained on broad text data and adapted for many downstream tasks.
Glossary
Latent Space
A compressed internal representation space where a model does its work, keeping the meaningful structure of the data while discarding redundant detail.
Glossary
Learning Rate
A training hyperparameter that controls how large each parameter update is during optimization.
Glossary
Linear Attention
An attention family that replaces the softmax with a factorizable similarity, collapsing the key/value history into a fixed-size state and making cost linear in sequence length.
Glossary
Linear Probe
A small supervised classifier trained on a network's internal activations to test whether a given property is linearly represented in them.
Glossary
Linear Regression
A model predicting the mean of a continuous outcome as a weighted sum of features, classically fitted by least squares.
Glossary
Link Function
The transformation in a GLM that connects the unbounded weighted sum of features to the legal range of the outcome's mean, such as the logit for probabilities.
Glossary
Log-Rank Test
The standard hypothesis test for whether two or more groups share the same survival curve, comparing each group's observed events against its expected share at every event time.
Glossary
Logistic Regression
A GLM for binary outcomes that models the log-odds as a weighted sum of features, yielding calibrated probabilities through the sigmoid.
Glossary
Logit Lens
An interpretability technique that applies a model's unembedding matrix to intermediate-layer activations, revealing which output tokens the model is leaning toward at each layer.
Glossary
Loop Engineering
Designing systems that prompt, verify, and re-dispatch coding agents automatically, so the human builds the loop that runs the agent instead of typing each prompt by hand.
Glossary
Markov Decision Process
The standard formalism for sequential decision problems, defined by states, actions, transition dynamics, a reward function, and a discount factor, where the next state depends only on the current state and action.
Glossary
Maximum Likelihood Estimation
Fitting a model by choosing the parameters under which the observed data would have been most probable.
Glossary
Mechanistic Interpretability
The research program of reverse-engineering the internal computations of neural networks into human-understandable components such as features, directions, and circuits.
Glossary
Minimum Detectable Effect
The smallest true effect an experiment is designed to detect reliably at its chosen power and significance level; the honesty clause of a power calculation.
Glossary
Mixed Effects Model
A regression containing both fixed effects shared by all observations and random effects drawn from a distribution per group, so grouped data is modeled with partial pooling instead of ignored or overfit.
Glossary
Mixture of Experts
A neural architecture that routes each input to a small subset of specialist subnetworks, so a model can hold huge total capacity while spending only a fraction of it per token.
Glossary
Model FLOPs Utilization
The fraction of a hardware fleet's peak compute that ends up in a model's actual forward and backward passes; large training runs typically achieve around 40%.
Glossary
Model Predictive Control
A classical control strategy that re-solves a short-horizon trajectory optimization at every control tick, executes the first action, and re-plans, gaining robustness to model error through constant feedback.
Glossary
Monotonic Constraint
A modelling constraint that forces predictions to only move in one direction as a chosen feature increases.
Glossary
Natural Language Processing
A field focused on helping computers interpret, transform, and generate human language.
Glossary
Novelty Effect
A short-lived lift from users exploring anything new, which fades as curiosity does; its mirror image, the primacy effect, is a temporary penalty while users relearn a changed interface.
Glossary
Odds Ratio
The multiplicative change in the odds of an outcome per unit change in a feature, obtained by exponentiating a logistic regression coefficient.
Glossary
Overall Evaluation Criterion
The single pre-declared metric (or fixed weighting of a few) that decides an experiment's outcome, chosen so that short-term movement predicts long-term business value.
Glossary
Overfitting
A failure mode where a model matches the training data too closely and performs worse on new, unseen examples.
Glossary
Partial Pooling
The middle path between fitting one model to all groups (complete pooling) and fitting each group separately (no pooling); group estimates are pulled toward the overall mean with strength inversely related to group size.
Glossary
Pipeline Parallelism
Training parallelism where consecutive layers live on different chips and activations flow between stages, cheap in bandwidth but prone to idle "bubbles".
Glossary
Policy Gradient
A family of RL methods that parameterise the policy directly and improve it by gradient ascent on expected return, nudging up the probability of actions in proportion to the reward that followed them.
Glossary
Potential Outcomes
The pair of outcomes a unit would experience with and without treatment, of which reality reveals exactly one; the framework that makes causal effects precise.
Glossary
Pre-Training
The initial large-scale training phase that teaches a model general patterns before narrower task-specific adaptation.
Glossary
Prompt
The input text or instructions given to a generative model to shape the response it produces.
Glossary
Prompt Engineering
The practice of designing prompts so a model is more likely to produce the kind of output you want.
Glossary
Propensity Score
The probability of receiving treatment given covariates, e(x) = P(T = 1 | X = x); a one-number summary that makes high-dimensional confounder adjustment tractable.
Glossary
Random Effect
A group-specific coefficient treated as a draw from a shared distribution rather than a free parameter, so the model estimates the distribution's variance and shrinks each group's estimate toward the population center.
Glossary
Randomized Controlled Trial
An experiment that assigns treatment by chance, making treatment independent of potential outcomes so a simple difference in means is the causal effect.
Glossary
Reasoning Model
A language model trained to produce an explicit intermediate chain of thought before its final answer, trading extra inference compute for accuracy on multi-step problems.
Glossary
Regression Discontinuity Design
A design that reads the causal effect off the outcome's jump at a threshold rule, since units just above and below the cutoff are effectively randomized.
Glossary
Regularization
Techniques that limit model complexity or penalize certain behaviors so the model generalizes better to new data.
Glossary
Reinforcement Learning
A family of methods where an agent learns by taking actions, receiving rewards, and improving behavior over time.
Glossary
Reinforcement Learning from Human Feedback
A post-training method that aligns a language model to human preferences by training a reward model on human comparisons, then optimising the policy against that reward with RL.
Glossary
Rejection Sampling
Sampling from a target distribution by drawing from an easier proposal distribution and probabilistically keeping or discarding each draw.
Glossary
Residual Stream
The running per-token vector in a transformer that every attention and MLP block reads from and additively writes back to, acting as the model's shared communication channel.
Glossary
Restricted Maximum Likelihood
A variant of maximum likelihood that integrates the fixed effects out before estimating variance components, correcting the downward bias that plain ML inherits from estimating the mean structure.
Glossary
Retrieval-Augmented Generation
Answering with an LLM by first retrieving relevant documents or memories from an external store (usually via embedding similarity) and placing them in the context window alongside the query.
Glossary
Roofline Model
A performance model saying a workload's speed is capped by either a chip's peak compute or its memory bandwidth times the workload's arithmetic intensity, whichever is lower.
Glossary
Sample Ratio Mismatch
A statistically implausible gap between an experiment's configured traffic split and the observed user counts, signalling broken randomization that invalidates the results.
Glossary
Score Function
The gradient of the log-density of a distribution, pointing at every location toward regions of higher probability.
Glossary
Selection Bias
Bias from the treated and untreated groups differing even without treatment, formally E[Y(0) given treated] minus E[Y(0) given untreated]; the main reason naive comparisons mislead.
Glossary
Self-Attention
Attention applied within a single sequence, so every token builds its new representation as a weighted mix of all the other tokens.
Glossary
Semantic Search
Search that tries to match by meaning and intent instead of only exact keyword overlap.
Glossary
Sequential Testing
Statistical designs that permit analyzing an experiment repeatedly while it runs, by spending the error budget deliberately across looks instead of pretending each look is the only one.
Glossary
Sim-to-Real Transfer
Training a policy in a physics simulator and deploying it on real hardware, closing the gap between simulated and real dynamics with techniques like domain randomization and learned actuator models.
Glossary
Sliding-Window Attention
A sparse attention pattern where each token attends only to a fixed-size window of recent tokens, bounding compute and cache growth on long contexts.
Glossary
Sparse Autoencoder
An autoencoder trained on a model's internal activations with a sparsity penalty, used to decompose activations into a large dictionary of interpretable feature directions.
Glossary
Speculative Decoding
An inference trick where a small draft model proposes several tokens and the large model verifies them in one batched pass, spending spare compute to cut latency.
Glossary
Statistical Power
The probability a test detects an effect of a given size when it truly exists; the complement of the false-negative rate, conventionally targeted at 80%.
Glossary
Superintelligence
A hypothetical intelligence that greatly exceeds the best human performance across virtually every domain, including scientific creativity, strategic reasoning, and social skill.
Glossary
Superposition
A representation strategy where a network stores more features than it has dimensions by assigning each feature an almost-orthogonal direction and tolerating small interference between them.
Glossary
Survival Function
The probability that a lifetime exceeds t, written S(t). It starts at 1, never increases, and its value at each time is the fraction of a population expected to still be "alive" then.
Glossary
Teleoperation
A human directly controlling a robot's movements in real time, via leader-follower arms, VR controllers, or hand-held devices; the dominant way robot manipulation demonstrations are collected.
Glossary
Temporal-Difference Learning
Value-learning methods that update predictions from the gap between successive estimates, r + gamma V(s') - V(s), rather than waiting for the final outcome, trading a little bias for much lower variance.
Glossary
Tensor Parallelism
Training and serving parallelism where each weight matrix itself is split across chips, requiring activation exchanges inside every layer.
Glossary
Tensor Processing Unit
Google's custom AI accelerator, built around large compiler-scheduled systolic arrays and connected into pods of thousands of chips via a switchless torus network.
Glossary
Test-Time Compute
Computation spent during inference rather than training (for example, longer reasoning traces or sampling many candidate answers) to raise quality without changing the weights.
Glossary
Test-Time Training
Architectures whose designated weights are updated by gradient steps during inference itself, so part of the model literally learns from the sequence it is processing.
Glossary
Time to First Token
The delay between sending a request and receiving the first token of the response, dominated by the compute-bound prefill of the prompt.
Glossary
Token
The basic unit a language model reads and predicts, which may be a word, character, or subword fragment.
Glossary
Transfer Learning
Reusing knowledge learned on one task or dataset to help solve a different but related task more efficiently.
Glossary
Transformer
A neural network architecture built around attention mechanisms that became the foundation for many modern language models.
Glossary
Vibe Coding
Building software by conversationally prompting an AI and accepting its output largely unread, embracing the vibes rather than the code. Fast for prototypes, risky for anything that has to last.
Glossary
Virtual Machine
An emulated computer that runs a complete guest operating system on virtualized hardware, providing strong isolation at the cost of size and startup time.
Glossary
Vision-Language-Action Model
A robot policy built on a web-pretrained vision-language model, fine-tuned so that alongside understanding images and text it also outputs robot actions, either as discrete tokens or via a fast generative action head.
Glossary
World Model
A learned model of environment dynamics that predicts future observations or states given actions, letting an agent plan or train inside its own "imagination" instead of the real environment.
Glossary
p-value
The probability, computed assuming no true effect, of observing a result at least as extreme as the one seen; small values are evidence against the no-effect hypothesis.