Glossary

Every term below is explained in plain words first. Each entry ends with the lesson that teaches it properly, so you can jump there for the worked example. Product and company facts are as of October 2026.

A · B · C · D · E · F · G · H · I · J · K · L · M · N · O · P · Q · R · S · T · U · V · W · Z

Abstention. The model declines to answer a case, and the case goes to a person or a bigger model instead. Abstaining on purpose is how you automate only the cases a model is sure about. Taught in: From probability to automation.

Active learning. Choosing the next items to label by where the model is most uncertain, instead of at random, so each label teaches it more. Taught in: Labels from AI and from the world.

Additive rubric. A grading scheme that gives one point per criterion met and adds them up, such as FineWeb-Edu’s 0-to-5 “educational value” score. Taught in: Designing the questions.

Aleatoric and epistemic uncertainty. Aleatoric uncertainty comes from the case itself being ambiguous, and more data will not remove it. Epistemic uncertainty comes from the model’s own ignorance of inputs like this one, and more data can reduce it. Taught in: How decision models work.

Amdahl’s law. The speedup of a whole job is limited by the part you did not speed up: for a 5x faster job, at least 80% of its time must sit in steps you made instant. Gene Amdahl presented it in 1967. Taught in: Is 5x real?.

Amortized inference. Paying a large training bill once so that every later answer costs only one forward pass. PFNs amortize Bayesian inference this way. Taught in: In-context learning and PFNs.

Attention. The step inside a transformer where each token compares itself with every other token it may see and mixes in information from the ones that match best. Taught in: Transformers in plain words.

Attention mask. A rule saying which tokens may look at which. Decision models use one to let many answer options share a single pass without seeing each other. Taught in: Generating vs deciding.

AUROC. The chance that a model scores a random positive case above a random negative one. 1.0 is perfect and 0.5 is a coin toss. Taught in: Calibration.

Autoregressive loop. How a chat model writes: predict one token, append it to the input, and repeat, so each new token waits for the last. Taught in: Generating vs deciding.

Base rate. How often something happens among all cases like this one, before you know anything about the particular case. Taught in: Probability is the product.

Baseline. The simple method a new model must beat, such as seasonal naive, gradient-boosted trees or a linear model. Many “foundation model wins” shrink against a well-tuned baseline. Taught in: The big debates.

Batch API. A provider endpoint that takes thousands of requests at once, returns them within hours, and charges about half price. It is the cheap way to buy labels. Taught in: Distill a decision model.

Bayes’ rule. The rule for updating a probability when new evidence arrives: start from the prior, weigh how much likelier the evidence is under each answer, and get the posterior. Taught in: Probability is the product.

Bitter lesson. Rich Sutton’s 2019 essay arguing that general methods that use more computation win in the long run over methods built on human knowledge. It is the main risk to any specialized model. Taught in: Wedges, moats and proof.

Bodhi. Fedora’s update system, where testers give each update karma (+1 or −1 votes) before it is pushed to users: a naturally labelled signal of whether an update is good. Taught in: The decision map.

bootc. A way to ship a whole operating system, kernel included, as an OCI image (the container format Docker and Podman use) that machines boot from and update as one unit. Taught in: Where the time goes.

Bot wall. A gate in front of a website that blocks automated clients; Anubis, for example, makes each browser solve a small puzzle first. Treat one as a clear “no”. Taught in: Sourcing from the internet.

Brier score. A score for probability forecasts: the squared gap between the probability and what happened (1 or 0), averaged. Lower is better, and it is proper, so honesty scores best. Taught in: Calibration.

Build farm. A set of servers that runs packages’ build recipes for a distro. Fedora’s official one is Koji; COPR is its self-service build service, where anyone can build into a personal repository; Nix’s is Hydra. Taught in: Where the time goes.

Build loop. The time from “I changed something” to “I know whether it worked”. Shortening it is the biggest lever in distro work. Taught in: Where the time goes.

C4 model. A way to draw software architecture at four zoom levels (system context, container, component, code), each answering one question. The course draws the first three in ASCII. Taught in: How to read the diagrams.

Calibration. Whether a model’s probabilities match reality: a calibrated model is right about 80% of the time on the answers it gives 0.8. Taught in: Calibration.

Cascade. Try a cheap model first and escalate to a more expensive one only when the cheap one is not confident. Taught in: From probability to automation.

Censored outcome. A case whose answer has not arrived yet when you stop watching, such as a bug still open. It is not a “no”. Taught in: Designing the questions.

Chinchilla. A 2022 result that, for a fixed training budget, you should double the training tokens whenever you double the parameters; the 70B Chinchilla beat the 280B Gopher on the same compute. Taught in: The big debates.

Choice. Jev’s question type for “which one of these named options?”, with up to 255 options. It returns the pick, a probability per option and a confidence. Taught in: Anatomy of Jev.

Cohen’s kappa. Agreement between two sets of labels, corrected for the agreement you would get by chance. 1 is perfect agreement and 0 is chance. Krippendorff’s alpha extends the idea to more than two annotators and missing labels. Taught in: Designing the questions.

Common Crawl. A free, monthly crawl of the public web; its August 2026 crawl held 2.14 billion pages. Taught in: Sourcing from the internet.

Conditional probability. The probability of A counted only among cases where B is true, written P(A | B). Taught in: Probability is the product.

Confidence (Jev). A fixed formula Jev applies to the distribution it returned: 0 for an even spread, 1 for all weight on one answer. It is not a second, independent estimate. Taught in: Anatomy of Jev.

Confident learning. Finding likely wrong labels in a dataset from a model’s own out-of-fold predictions; the cleanlab library does this. Taught in: Labels from AI and from the world.

Conformal prediction. A method that turns any model’s scores into a set of answers guaranteed to contain the right one at a rate you choose, using a small labelled calibration set. Taught in: From probability to automation.

Constrained decoding. Generating text while forbidding any token that would break a required format, such as a JSON schema. It fixes the format, not the cost of generating. Taught in: Generating vs deciding.

Context window. The most tokens a model can read in one go. Taught in: Transformers in plain words.

Covariates. Extra series that help a forecast, such as temperature or a holiday calendar known ahead of time. Taught in: TabPFN, TimesFM and friends.

Coverage. The share of cases you let the model decide on its own. Coverage at fixed precision, for example “the share handled alone while 98% of automatic actions are right”, is the automation number that matters. Taught in: From probability to automation.

Criteria drift. Discovering, while grading real outputs, that your idea of a right answer was different from what you wrote down. It is why you label some items yourself before freezing a label guide. Taught in: Designing the questions.

Cross-encoder and bi-encoder. A cross-encoder reads the input and one option together, so each can attend to the other: most accurate, but one pass per pair. A bi-encoder turns each into a vector separately and compares them, so option vectors can be cached. Taught in: How decision models work.

CRPS. The continuous ranked probability score: how far a whole probabilistic forecast is from what happened, averaged over its quantiles. Lower is better. Taught in: TabPFN, TimesFM and friends.

decider-2b. One person’s open Jev replica: Qwen3.5-2B-Base plus a LoRA, trained on one B300 GPU in 56 minutes, with about 95k downloads in a month. Taught in: Distill a decision model.

Decision model. A model that returns one answer from a closed set plus a probability, instead of free text. Jev is the best-known example. Taught in: How decision models work.

Decode. The phase where a model writes its output one token at a time, each step waiting for the last. Decision models skip it. Taught in: Generating vs deciding.

Decoder. A transformer in which each token can see only the tokens before it. That is what lets it generate text; every chat model is a decoder. Taught in: Transformers in plain words.

Decontamination. Removing training items that overlap your test set or public benchmarks, so scores measure skill rather than memory. Taught in: Sourcing from the internet.

Densing law. The observation that capability per parameter doubles about every three months, so a model of fixed quality keeps getting smaller and cheaper. Taught in: The big debates.

Distillation. Labelling data with a big “teacher” model and training a small “student” on those labels, so you get similar answers far more cheaply. The student is capped near the teacher’s own agreement with the truth. Taught in: Distill a decision model.

Distro. A Linux distribution: a curated, tested set of software built from upstream sources, plus the tools to install and update it. Taught in: Where the time goes.

ECE (expected calibration error). Group predictions by stated confidence, measure the gap between confidence and actual accuracy in each group, and average the gaps weighted by group size. 0 is perfect. Taught in: Calibration.

Elo. A chess-style rating from head-to-head wins; on TabArena a 400-point gap means about 10-to-1 odds that the higher-rated method wins on a dataset. Taught in: TabPFN, TimesFM and friends.

Embedding. A list of numbers (a vector) that stands for a token or a piece of text; similar meanings get similar vectors. Taught in: Transformers in plain words.

Encoder. A transformer in which every token can see every other token, before and after it. It reads rather than writes, and is the natural shape for classifying. Taught in: Transformers in plain words.

Entropy. A measure of how spread out a probability distribution is: 0 when all weight is on one answer, highest when it is even. Taught in: How decision models work.

ERA5. A free, decades-long, gap-free reconstruction of past weather (a “reanalysis”) that AI weather models learn from. Taught in: Vertical foundation models.

Expected cost. Each outcome’s cost times its probability, added up. Pick the action with the lowest expected cost. Taught in: Probability is the product.

Expected value. The average you would get over many repeats: each value times its probability, all added up. Jev’s Score returns the expected level this way. Taught in: Probability is the product.

Extraction. Turning a raw web page, PDF or log into the clean text your model will read, with tools such as trafilatura (HTML) or Docling (PDF). Taught in: Sourcing from the internet.

Fair use. A US legal doctrine that allows some unlicensed use of copyrighted work; courts weigh factors such as how transformative the use is and whether it harms the market for the original. Taught in: Data rights.

Fallback rate. The share of cases a cheap model hands to an expensive one. It usually dominates the bill. Taught in: Serving for a fraction of a cent.

False positive and false negative. A false positive is acting when you should not have; a false negative is not acting when you should have. They usually cost different amounts. Taught in: From probability to automation.

Few-shot. Handling a task from a handful of worked examples placed in the input, with no training run. Taught in: Three ways to teach a model.

Fine-tuning. Continuing to train an already pre-trained model on your own examples so it gets good at your task. Full fine-tuning updates every parameter; LoRA trains only a small add-on. Taught in: Three ways to teach a model.

FLOP. One arithmetic operation, such as a multiply or an add. Model cost is counted in FLOPs. Taught in: What a model is.

Forward pass. One run of a network over its input, from input to output scores. A decision model answers in one. Taught in: How decision models work.

Foundation model. A model trained once on broad data and then reused for many tasks, instead of training a new model per task. Taught in: What a model is.

FTBFS. “Fails to build from source”: the distro word for a package that no longer builds. Taught in: Where the time goes.

Function. A rule that turns an input into an output. A model is a function whose rule was learned rather than written by hand. Taught in: What a model is.

Generalization. Doing well on inputs the model never saw in training. It is the only thing that counts. Taught in: What a model is.

GIFT-Eval. Salesforce’s time-series benchmark: 23 datasets and 144K series across 7 domains and 10 frequencies, with leakage flags on the board. fev-bench (Amazon, 100 tasks, 46 with covariates) is the other main forecasting benchmark. Taught in: TabPFN, TimesFM and friends.

Gold set. A sample labelled carefully, by you or by a trusted outcome, used to measure every other label source and model. Taught in: Distill a decision model.

GPAI model. A “general-purpose AI model” under the EU AI Act, indicatively one trained with more than 10^23 FLOP that can generate text, images or video. Its provider has extra duties; a small decision model is very likely not one. Taught in: Data rights.

Gradient boosting. Building many small decision trees, each one correcting the mistakes of the ones before; LightGBM and XGBoost are common libraries. It is still a strong baseline on tables. Taught in: Building the co-pilot.

Gradient descent. The standard training method: for every parameter, work out which way lowers the loss, nudge it a little that way, and repeat. Taught in: What a model is.

Gross margin. Revenue minus the direct cost of serving it, divided by revenue. For a decision API it falls as model size grows. Taught in: Transformers in plain words.

Hard and soft labels. A hard label is just the teacher’s answer; a soft label is the teacher’s full probability for each answer, which carries more information for a student. Taught in: Distill a decision model.

Hard negative. A near-miss: a small edit to a real item that flips its correct label. Synthetic data is most useful for these. Taught in: Labels from AI and from the world.

Head. The small final layer that turns a model’s internal vectors into a typed answer, such as one probability per option. Taught in: How decision models work.

In-context learning (ICL). A model picking up a new task from examples placed in its input, with no change to its parameters. Taught in: In-context learning and PFNs.

Inference. Running a trained model, with its parameters frozen, on a new input. Over a model’s life it is where most of the cost sits. Taught in: What a model is.

Interruptible (spot) rental. A cheaper GPU rental that the host can take back at any moment, so your training script must save checkpoints (snapshots of the training state) and resume from them. Taught in: The bootstrapper’s stack.

Jaccard similarity. Shared items divided by all distinct items. Dedup uses it on shingles (runs of n consecutive words) to find near-duplicate pages. Taught in: Sourcing from the internet.

Jev. TypeSafe AI’s hosted decision model, opened in early access on 2026-09-15. You send a state and typed questions, and it returns a probability distribution for each one; it never writes text. TypeSafe calls this class “System One models”. Its question types are Choice, Score and Noul, plus a Box preview for images. Taught in: Anatomy of Jev.

Jev Decision Index. A community leaderboard on Hugging Face, started in September 2026, that ranks open and hosted decision models across many tasks, with calibration error per task. Taught in: Anatomy of Jev.

Jury. Several models from different families labelling the same items independently, then voting. Panels of smaller models have beaten one large judge at a fraction of the cost. Taught in: Labels from AI and from the world.

KL divergence. A measure of how far one probability distribution is from another, such as a student’s from its teacher’s. 0 means identical. Taught in: Distill a decision model.

KV cache. Stored attention data from tokens a model has already read, so later tokens can reuse it instead of recomputing. It lets many options share one reading of the input. Taught in: Generating vs deciding.

Labelling function. A small program that votes for a label, or abstains, on each item; weak supervision combines many of them. Taught in: Labels from AI and from the world.

Layer. One round of attention followed by a feed-forward network (a small block of learned multiplications per token). A transformer stacks many, each refining every token’s vector. Taught in: Transformers in plain words.

Leakage. Test data, or something close to it, sneaking into training, or a feature that already contains the answer. It inflates scores; when it makes a “zero-shot” score partly memory it is also called contamination. Split by time and by entity to avoid it. Taught in: What a model is.

Licence. The terms that say what you may do with a model or dataset. Many 2026 models are “open for research, paid for production”. Taught in: The market map.

llama.cpp. An open engine for running language models on ordinary hardware, including CPUs and Macs. Its file format is GGUF, where Q8_0 is the 8-bit type and Q4_K_M a 4-bit one. Taught in: Serving for a fraction of a cent.

Log Detective. A Fedora service that explains failed RPM build logs with a large language model, aimed at newer packagers. Taught in: The decision map.

Log loss. A score for probability forecasts: minus the log of the probability given to what actually happened, averaged. Confident mistakes cost the most. Taught in: Calibration.

Logit. The raw score a model gives each possible answer before softmax: any number, where bigger means “more likely”. Taught in: Generating vs deciding.

Logprob. The log of the probability a model gives a token. Reading logprobs is how you score options with an LLM. Taught in: How decision models work.

LoRA. Low-rank adaptation: freeze the original model and learn a small correction beside each big weight matrix, so fine-tuning is cheap and the add-on is tiny. The width of those small matrices is the rank. Taught in: Three ways to teach a model.

Loss. One number that says how wrong a model is on a set of examples. Training makes it small. Taught in: What a model is.

Mass rebuild. Rebuilding every package in a distro against a new toolchain, such as a new compiler. Taught in: Where the time goes.

Mean squared error. Square each prediction’s error, then average. A common loss for numbers. Taught in: What a model is.

MinHash and LSH. Tricks that make near-duplicate search cheap at web scale: MinHash compresses each document into a short signature, and locality-sensitive hashing (LSH) avoids comparing every pair. Taught in: Sourcing from the internet.

MLIP (universal interatomic potential). A neural network trained on results of density functional theory (DFT), the accurate but expensive quantum calculation of a material’s energy. It predicts energies and forces for almost any atoms, thousands to millions of times faster. Taught in: Vertical foundation models.

MLX. Apple’s array framework for its own chips; mlx-lm is its library for running and fine-tuning language models on a Mac. Taught in: The bootstrapper’s stack.

Moat. Whatever stops a well-funded rival from copying you next month, such as data rights, private evals or distribution. Taught in: Wedges, moats and proof.

Model. A function whose rule was learned from data instead of written by hand. Taught in: What a model is.

Model collapse. The decay that happens when models train on model outputs. It shows up when synthetic data replaces real data, and not when it is added alongside. Taught in: Labels from AI and from the world.

Model-output terms. A provider’s rules on who may train on what its model writes. Anthropic, Google and OpenAI bar using their outputs to train competing models; open weights under MIT or Apache-2.0 carry no such clause. Your labeller’s terms follow your labels into your product. Taught in: Data rights.

Natural language inference (NLI). Deciding whether one sentence entails, contradicts or is neutral towards another. Since 2019 it has been used as a zero-shot classifier: “Does this text entail ‘This text is about sport’?” Taught in: How decision models work.

Naturally labelled outcome. A label the world produces anyway, such as “this build failed” or “this retry passed”, so no annotator is needed. Also called an outcome label. Taught in: The decision map.

Near-duplicate. Two documents that differ only slightly, such as the same article on two sites. Finding them needs a measure of overlap, unlike exact duplicates. Semantic dedup goes further and removes documents that say the same thing in different words, by comparing embeddings. Taught in: Sourcing from the internet.

Noise floor. The calibration error a perfectly calibrated model still shows by chance on a finite test set; about 0.03 ECE on 1,000 items. Taught in: Calibration.

Non-dilutive funding. Money, such as a grant or cloud credits, that costs no equity in your company. Taught in: Wedges, moats and proof.

“None of the above” and “Unsure”. Two different escape answers: “none of the above” says the item fits no category; “unsure” says the annotator, human or model, cannot tell. Taught in: Designing the questions.

Noul. Jev’s yes/no question type, short for Bernoulli (the name for a yes/no random outcome). It returns one probability of “yes”. Taught in: Anatomy of Jev.

Numerical weather prediction (NWP). Forecasting by solving the physics equations of the atmosphere on a supercomputer, as ECMWF’s IFS system does. Taught in: Vertical foundation models.

Odds. A probability written as “for : against”. A probability of 0.2 is odds of 1 : 4. Taught in: Probability is the product.

ONNX and ONNX Runtime. ONNX is an open file format for trained models; ONNX Runtime is an engine that runs it fast on CPUs and GPUs. Taught in: Serving for a fraction of a cent.

Open for research, paid for production. The common 2026 licence pattern: weights are downloadable under a non-commercial licence, and commercial use goes through a paid API or licence. TimesFM-3 and TabPFN-2.5 to 3.5 ship this way. Taught in: The market map.

Open weights. A model whose trained parameters you can download and run yourself; the licence still decides what you may do with it. Taught in: The market map.

Oracle. A program that gives the right answer every time, such as a compiler deciding whether code builds. Where a fast one exists, use it instead of predicting. Taught in: Designing the questions.

Outcome reward and process reward. An outcome reward scores only the final answer; a process reward scores each step of the reasoning. Taught in: Verifiable rewards and evals.

Overconfident and underconfident. Overconfident means the model’s stated probability is higher than its real hit rate; underconfident means lower. Chat-tuned LLMs tend to be overconfident. Taught in: Calibration.

Overfitting. Fitting the training examples, noise and all, so tightly that the model does worse on new ones. Taught in: What a model is.

Over-training. Training a smaller model on far more tokens than the Chinchilla rule suggests, because a small model is cheaper to run for its whole life. Taught in: The big debates.

Package. One piece of software as a distro ships it, with a build recipe that says how to compile it and where its files go. Taught in: Where the time goes.

Parameters. The numbers inside a model that set its behaviour, learned in training. Their count is the model’s size. Taught in: What a model is.

Patches. Blocks of consecutive time steps that a time-series model reads as one token each. Taught in: TabPFN, TimesFM and friends.

Perturbation prediction. Forecasting how a cell’s genes respond when one gene is knocked down. Foundation models have so far failed to beat linear baselines at it. Taught in: Vertical foundation models.

PFN (prior-data fitted network). A transformer pre-trained on millions of synthetic datasets drawn from a prior, so that at use time it predicts on your data from examples in its context, in one forward pass. TabPFN is the best-known one. Taught in: In-context learning and PFNs.

PII. Personally identifiable information: anything that identifies a person, such as names, email addresses or IP addresses. Remove it before training. Taught in: Sourcing from the internet.

Platt scaling and isotonic regression. Two calibration fixes: Platt scaling fits a two-parameter sigmoid from score to probability and suits yes/no outputs; isotonic regression fits any non-decreasing step function and needs more data. Taught in: Calibration.

Posterior. What you believe after seeing the evidence, from Bayes’ rule. Taught in: Probability is the product.

Posterior predictive distribution (PPD). The Bayesian answer to “given these examples, how likely is each label for this new input?”, averaged over every explanation weighted by how well it fits. A PFN learns to output it. Taught in: In-context learning and PFNs.

Precision, recall and F1. Precision: of the cases the model flagged, how many were right. Recall: of the cases that should have been flagged, how many it caught. F1 balances the two in one score, and macro-F1 averages it over classes so rare ones count equally. Taught in: The decision map.

Prefill. The phase where a model reads its whole input (its input tokens) in one parallel pass. Decision models only prefill, which is why they are fast and priced per input token. Taught in: Generating vs deciding.

Pre-training. The first, large training run on broad data that gives a model its general skills, before any task-specific step. Taught in: Three ways to teach a model.

Prior. What you believe before seeing the evidence. For a PFN, the prior is also the recipe that generates its synthetic training datasets: the model’s worldview. Taught in: Probability is the product.

Probability. A number from 0 to 1 that says how often something happens in the long run, among cases like this one. Taught in: Probability is the product.

Probability distribution. A probability for every possible answer, adding up to 1. It is what a decision model returns. Taught in: Probability is the product.

Proper scoring rule. A scoring rule gives a forecast a number once you know what happened. A proper one, such as Brier or log loss, gives its best expected value only when you report what you truly believe. Training with one rewards honest probabilities. Taught in: Probability is the product.

Protein language model (PLM). A model trained to predict masked or next amino acids in protein sequences; its internal representations end up encoding structure and function. Taught in: Vertical foundation models.

QLoRA. LoRA on top of a base model stored in 4 bits, so a big model’s fine-tune fits on one GPU. Taught in: Three ways to teach a model.

Quantiles. Points of a forecast’s spread: the 90th-percentile forecast is a value the outcome should fall below 90% of the time. Taught in: TabPFN, TimesFM and friends.

Quantization. Storing a model’s weights in fewer bits (8 or 4 instead of 16), like rounding prices to whole dollars, so it is smaller and often faster. GPTQ and AWQ are 4-bit methods mostly used on GPUs. Taught in: Serving for a fraction of a cent.

Query, key and value. The three vectors attention makes for each token: the query says what it is looking for, the key what it contains, and the value what it hands over when matched. Taught in: Transformers in plain words.

Reinforcement learning (RL). Training by letting a model try, scoring each try with a reward, and nudging it toward the tries that scored higher. Taught in: Verifiable rewards and evals.

Relational foundation model. A foundation model that predicts across a database’s linked tables, not just one table; Kumo’s KumoRFM is the example, and Nvidia bought Kumo in 2026. Taught in: The market map.

Reliability diagram. A chart that groups predictions by stated confidence and plots how often each group was actually right. Points on the diagonal are calibrated. Taught in: Calibration.

Request for Startups (RFS). Y Combinator’s public list of ideas it wants founders to apply with. The course’s opening quote is Diana Hu’s “Vertical Foundation Models” entry in the Winter 2027 list. Taught in: The market map.

Reward. A score that tells a training process how good an output was. Taught in: Verifiable rewards and evals.

Reward hacking. A model maximizing the checker instead of doing the task, for example by editing the test. It is common, not an edge case. Taught in: Verifiable rewards and evals.

Reward model. A model that outputs one number for “how good is this answer”, learned from people’s preferences between pairs of answers. Taught in: How decision models work.

Risk-coverage curve. Sort cases by confidence, accept from the top, and plot the error rate at each coverage. It shows how much you can automate at a given error rate. Taught in: From probability to automation.

RLCD. “Reinforcement Learning for Calibrated Decisions”, the name TypeSafe AI gives Jev’s training method; details are unpublished. Taught in: Anatomy of Jev.

RLCR. Reinforcement learning with a reward of correctness minus the Brier score, so a model learns to be right and honest about its confidence at once. Taught in: Calibration.

RLHF. Reinforcement learning from human feedback: training a model toward answers people prefer, using a reward model built from their choices. Taught in: Anatomy of Jev.

RLVR. Reinforcement learning with verifiable rewards: the reward comes from a program that checks the answer, not from a learned reward model. It made today’s reasoning models strong. Taught in: Verifiable rewards and evals.

robots.txt. A file at a website’s root that states crawl rules for bots. It is a request, not access control, but honouring it is part of collecting data with clear rights. Cloudflare’s content signals (2025-09-24) add lines saying whether pages may be used for AI training. Taught in: Data rights.

Router. A component that picks one model per query up front, instead of trying a cheap one first as a cascade does. Taught in: From probability to automation.

rpmeta. A small, typed model inside Fedora’s build system that predicts how long an RPM build will take, trained on 705,766 Copr builds. Taught in: What a model is.

RPM and spec file. RPM is Fedora’s package format; a spec file is its build recipe. Arch’s equivalent recipe is a PKGBUILD, and the AUR (Arch User Repository) held 123,268 user-submitted ones on 2026-10-28. Taught in: Where the time goes.

Score. Jev’s question type for “where on this ordered scale?”, with 2 to 10 levels. It returns the probability-weighted average level plus a probability per level. Taught in: Anatomy of Jev.

Seasonal naive. The “do nothing clever” forecast: predict each future step as the value one season earlier. It is the yardstick forecasting benchmarks divide by: MASE is a model’s error over seasonal naive’s, so below 1 beats it. Taught in: TabPFN, TimesFM and friends.

Secure Boot. A firmware feature that only loads kernel code signed with a trusted key. On Fedora, akmods builds out-of-tree modules such as NVIDIA’s driver and signs them with a local key, which you enrol at the MOK (Machine Owner Key) screen at the next boot. Taught in: Where the time goes.

Selective prediction. Letting a model abstain on cases it is unsure about, so it answers fewer cases with fewer errors. Selective risk is the error rate on the cases it does answer. Taught in: From probability to automation.

Serverless. Hosting where a container starts when a request arrives, so no server sits waiting all day. The catch is the cold start: the delay while a fresh container boots and loads the model. Taught in: Serving for a fraction of a cent.

Shadow mode. Running a new model next to the current process so it predicts but does not act, letting you measure it on live traffic safely. Taught in: Building the co-pilot.

Sharpness and discrimination. Sharpness: forecasts commit (3% or 97%, not 50%). Discrimination: scores rank right answers above wrong ones. Both matter alongside calibration. Taught in: Calibration.

Sigmoid. The one-logit version of softmax for a single yes/no answer: p = 1 / (1 + exp(−z)). Taught in: Generating vs deciding.

Single-cell foundation model. A model pre-trained on millions of single-cell gene-expression profiles, meant to transfer to tasks like cell typing. Taught in: Vertical foundation models.

Softmax. The formula that turns a list of logits into probabilities that add up to 1: exponentiate each, then divide by the total. Taught in: Generating vs deciding.

SOTA. State of the art: the best published result on a named test at a named date. Without the test and the date, the claim means little. Taught in: Talking SOTA.

State. What you hand Jev to judge: a string, a JSON object or a JSON array. Questions are asked about the state. Taught in: Anatomy of Jev.

Structural causal model (SCM). A recipe that generates data from a random graph of variables, each a random function of its parents plus noise. TabPFN’s synthetic tables come from SCMs. Taught in: In-context learning and PFNs.

Synthetic data. Training data written by a model. It helps most for edge cases and hard negatives, not as the backbone of a dataset. Taught in: Labels from AI and from the world.

System 1 and System 2. Kahneman’s names for fast, intuitive judgment and slow, deliberate reasoning. In the course, a decision model plays System 1 and an LLM agent System 2. Taught in: Is 5x real?.

TabArena. A live leaderboard on Hugging Face that ranks tabular prediction methods by Elo, with timing and leak warnings. Taught in: TabPFN, TimesFM and friends.

Tabular foundation model. A model trained once on millions of small, mostly synthetic tables that predicts on your table from the rows you hand it, with no training run. Taught in: TabPFN, TimesFM and friends.

TabPFN. The best-known tabular foundation model, from Prior Labs, pre-trained only on synthetic tables. SAP agreed to buy Prior Labs on 2026-05-04; TabPFN-3.5 (2026-09-15) handles up to 1M rows. Taught in: TabPFN, TimesFM and friends.

Taxonomy. The list of allowed answers to a question. Getting it wrong is the costliest data mistake, because every later fix assumes it is right. The answer type is its shape: one of several options, a place on an ordered scale, or yes/no. Taught in: Designing the questions.

TDM opt-out. Under Article 4 of the EU’s DSM Directive, text and data mining of lawfully accessible works is allowed unless the rightsholder reserves the right in a machine-readable way. Your crawler must honour those reservations. Taught in: Data rights.

Temperature. One number that divides all logits before softmax: above 1 it softens the probabilities, below 1 it sharpens them. Taught in: Generating vs deciding.

Temperature scaling. The standard calibration fix: keep the model frozen and fit one temperature on held-out labelled data. Taught in: Calibration.

Time split. Testing only on cases that happened after the training cases (ideally also after every model’s training cutoff), so the test cannot leak the future. Taught in: Verifiable rewards and evals.

Time-series foundation model. A model that forecasts a new series zero-shot from its history placed in context, such as TimesFM-3, Chronos-2 or Toto 2.0. Taught in: TabPFN, TimesFM and friends.

TimesFM. Google’s time-series foundation model. TimesFM-3 (announced 2026-08-31) has 330M parameters and forecasts related series at once in one forward pass; its downloadable weights are non-commercial. Taught in: TabPFN, TimesFM and friends.

Token. A chunk of text, often a word piece; roughly three-quarters of a word. Models read and write tokens. Taught in: What a model is.

Token probabilities and verbalized confidence. Two ways to get confidence from an LLM: read the probabilities it gives the answer tokens, or ask it to say a number. A third is the agreement rate across several samples. Taught in: Calibration.

Tokenizer. The part that splits text into tokens and gives each one a number, its ID. Taught in: Transformers in plain words.

Training. The search for parameter settings that make the loss small; one round of nudging them is a training step. Taught in: What a model is.

TypeSafe AI. The company behind Jev. It opened Jev with a $40M seed and on 2026-10-09 announced an $870M Series A at a $7.5B valuation. Taught in: Anatomy of Jev.

Utilisation. How busy your serving hardware is. An idle GPU still costs money, so low utilisation can make each decision many times dearer. Taught in: Serving for a fraction of a cent.

Verifiable reward. A training or evaluation signal checked automatically against ground truth (a resolved forecast, a passing build, a lab assay) rather than by opinion. The course’s opening quote calls getting these “the hard part”. Taught in: Verifiable rewards and evals.

Verifier. Whatever tells you an answer was right: a test suite, a compiler, a sensor reading or a later outcome. Taught in: The market map.

Vertical foundation model. A model pre-trained once for a single industry or science domain, such as weather, proteins or crystals, then reused across that domain’s tasks. Taught in: Vertical foundation models.

Visual anomaly detection. Flagging defective parts in images, ideally from a few or no examples of defects, since defects are rare by design. Taught in: Vertical foundation models.

Vocabulary. The fixed list of all tokens a model knows. A generating model picks its next token from it. Taught in: Generating vs deciding.

WARC, WAT and WET. Common Crawl’s three file types: WARC holds the raw HTTP responses, WAT metadata such as headers and links, and WET plain text Common Crawl already extracted. Taught in: Sourcing from the internet.

Weak supervision. Labelling data with many cheap, noisy rules (labelling functions) instead of hand labels, then combining their votes into a probabilistic label. Taught in: Labels from AI and from the world.

Wedge. The narrowest product you can own end to end: one typed question, one data source, one verifier and one buyer who pays per decision. Taught in: Wedges, moats and proof.

Zero-shot. Handling a task with no examples of it, from only an instruction or a list of label names. Check for contamination before believing a zero-shot score. Taught in: Three ways to teach a model.