Glossary

Accuracy is the fraction of rows you got right, and when one class is 99.8% of the rows, that fraction is almost entirely a measurement of the class balance.

From When accuracy lies, in Machine Learning, Foundations. Also taught in The Maths You Actually Need.

Activations: the intermediate tensors of the forward pass. Modest for a single short sequence, and they scale with batch size and sequence length.

From The memory arithmetic you can do in your head, in Hugging Face, End to End.

This is called an adversarial setting, and it changes everything about how the system is maintained.

From The models that have been working since 2002, in AI, Actually Explained.

A model with those abilities, running in a loop (look at the situation, decide on an action, take it, look again) is called an agent.

From When it starts to act, in AI, Actually Explained. Also taught in Building with AI, AI at Work and Agents and Automation.

An algorithm is a procedure: a finite sequence of steps that turns an input into an output.

From Five words that hide more than they say, in AI, Actually Explained. Also taught in Machine Learning, Foundations, Editing an Image and Making It Say Something.

Attention is the mechanism that does it, and it is the single idea that made everything since 2017 possible.

From Attention, without a single equation, in AI, Actually Explained. Also taught in AI, Safety and What Goes Wrong, Machine Learning, Foundations and How a Language Model Actually Works.

A human check counts only with time, independent information, authority and a counted override rate.

The key idea of Why people believe the machine, in AI, Safety and What Goes Wrong. Also taught in AI at Work.

This is called autoregressive generation, and everything below follows from it.

From One chunk at a time, and it cannot take it back, in AI, Actually Explained.

AWQ is activation-aware: it observes that roughly 1% of weight channels see large activations, and scales those channels up before quantising so they lose less.

From Quantisation, properly, in Running Models Yourself.

When we write [3, 4] we mean 3 lots of one reference direction plus 4 lots of another. Those reference directions are called a basis.

From The same vector, written two different ways, in The Maths You Actually Need.

Twenty examples from your own work predict your result better than any leaderboard.

The key idea of What a benchmark score leaves out, in AI, Safety and What Goes Wrong. Also taught in Knowing If It Works and The Open Model Ecosystem.

Bias is how far the average prediction across all those fits is from the truth.

From Bias and variance, with numbers attached, in Machine Learning, Foundations.

Blast radius is a count of irreversible actions: refunds issued, emails sent, rows deleted, files published.

From Budgets belong in the harness, not the prompt, in Agents and Automation.

Bleed is artwork extended past the trim, normally 3 mm in India and Europe or 0.125 in in the US, because a guillotine cutting 500 sheets drifts by a millimetre or two. Without it, that drift shows as a white sliver along one edge, so pull a background photograph out to the bleed line, never the trim line.

From The blue you chose on screen cannot be printed, in Vectors, Logos and Things That Get Printed.

A bus is a channel that several tracks feed into, so processing applied to the bus applies to all of them at once.

From A track layout that survives a revision, in Editing Video.

Calibration is a post-processing step, fitted on data the model did not train on.

From Does 0.7 mean seventy per cent?, in Machine Learning, Foundations. Also taught in AI, Safety and What Goes Wrong, Knowing If It Works and Fine-Tuning, and When Not To.

The common design is a cascade: a cheap model handles everything, and only the cases it cannot settle go to an expensive one.

From Spending money where it matters, in Choosing and Using the Tools.

The cast shadow is the shape thrown by the key light.

From Relighting a cut-out so it belongs, in Editing an Image.

A chunk must be small enough to be a precise retrieval unit and complete enough to be a usable answer, and those two demands pull in opposite directions.

The key idea of Chunking: the decision that quietly sets your ceiling, in How a Language Model Actually Works. Also taught in Python, from Zero, for AI, Building with AI and Context Engineering.

Two people labelling the same hundred items set the ceiling for any classifier, so measure human agreement before you measure the model.

The key idea of Pattern: one call, one label, then code, in Agents and Automation. Also taught in Making Things with AI and Fine-Tuning, and When Not To.

A server exposes capabilities; a client (an assistant, an IDE, an agent) consumes them.

From MCP: what a tool protocol buys you, and what it costs, in Building with AI.

The frame is a clipping path: an instruction that says render only the part of this object that falls inside this shape.

From Clipping hides; it does not remove, in Vectors, Logos and Things That Get Printed.

Closure means you rarely need to draw every boundary.

From When two grouping cues disagree, in Design, before Any Software.

The codec is how the video inside was compressed.

From Why the upload looks worse than the file, in Editing Video. Also taught in Editing Video.

Split every decision into the four outcome boxes and compute the rates separately per group: precision can look nearly equal while the true positive rate differs by a factor of three on the same numbers.

The key idea of Counting the four ways to be right and wrong, in AI, Safety and What Goes Wrong. Also taught in Machine Learning, Foundations and Knowing If It Works.

The container is the box that holds those compressed streams side by side, along with the timecode, the track names, the subtitle tracks and the metadata: MP4, MOV, MKV, MXF, WebM.

From The box and the codec are two different decisions, in Editing Video.

Context length: how many tokens fit in one call.

From What a language model actually is, in How a Language Model Actually Works.

Context window is how much text it can consider at once: its desk space.

From Five words that hide more than they say, in AI, Actually Explained. Also taught in AI, Actually Explained, Building with AI and AI at Work.

Continuity means a list of items sharing a left edge reads as one column even when the gaps between them are large.

From When two grouping cues disagree, in Design, before Any Software.

async def declares a coroutine: a function that can pause.

From asyncio: twenty calls in the time of one, in Python, from Zero, for AI.

Cosine similarity is the cosine of the angle between two vectors.

From Cosine, dot product, and the normalisation that connects them, in Hugging Face, End to End.

Cross-validation averages several held-out estimates so a score depends less on where the split fell, but the best score from a hyperparameter search is biased upward and needs an outer loop or an untouched test set to correct.

The key idea of Cross-validation: using every row twice, legally, in Machine Learning, Foundations.

Data poisoning means inserting text into a training corpus so that the resulting model behaves as the attacker wants.

From Attacks on the model itself, in AI, Safety and What Goes Wrong.

The database is where things survive.

From What a web app is actually made of, in Building Apps with AI.

A de-esser is a compressor that only responds to a narrow high band (typically 5 to 8 kHz) turning down the s sounds and leaving everything else.

From Four EQ moves that make speech intelligible, in Editing Video.

Verify synthetic media by origin and a second channel, not by squinting at pixels for artefacts.

The key idea of Deepfakes and synthetic media, in AI, Safety and What Goes Wrong. Also taught in AI, Actually Explained and AI for Teachers.

The number that decides whether this was worth building is the deflection rate: the fraction of requests served locally without escalation.

From Local first, API behind it, in Running Models Yourself.

Density is the spacing between stitches, typically 0.35 to 0.45 mm for a satin.

From Embroidery, which does not use vectors at all, in Vectors, Logos and Things That Get Printed. Also taught in Making It Say Something.

Generative fill initialises a masked region with noise and denoises it towards something consistent with the surrounding pixels and your prompt, which is why filled areas are cleaner than photographs, softer than the original file, and wrong wherever text or regular geometry is involved.

The key idea of Noise in, picture out: inpainting explained, in Editing an Image.

A small model trained to match a large model's full next-token distribution learns from the ranking of every wrong answer as well as the right one, which transfers behaviour and reasoning habits well and long-tail knowledge poorly.

The key idea of Distillation: how small models got good, in How a Language Model Actually Works. Also taught in Running Models Yourself and Fine-Tuning, and When Not To.

Dither is a tiny amount of random noise added before the truncation.

From Two numbers doing two different jobs, in Motion, Sound and Music.

The empirical finding, named double descent, is that test error follows the U up to roughly the point where the model can exactly fit the training data, and then falls again as capacity increases further.

From Bias and variance, with numbers attached, in Machine Learning, Foundations.

A model name is a pointer that providers move, so behaviour changes have to be separated into jitter, version drift and input drift by test rather than by guess, and only a pinned snapshot with a scheduled canary set makes a change attributable.

The key idea of Drift: the same name, a different model, in How a Language Model Actually Works. Also taught in Making Things with AI and Knowing If It Works.

The mechanism behind that is called an embedding, and it is worth twenty minutes of your attention even if you never write a line of code, because it is what search, document Q&A, recommendation and clustering are all built on, and it explains several behaviours that otherwise look like magic or malfunction.

From Why it finds "refund" when the customer typed "money back", in AI for a Small Business. Also taught in AI, Actually Explained, Machine Learning, Foundations and How a Language Model Actually Works.

An epoch is one full pass over the training data.

From Batches, epochs, and the two knobs that decide everything, in Machine Learning, Foundations. Also taught in Fine-Tuning, and When Not To.

The cache matches on an exact prefix: the request has to begin with byte-identical text.

From The two discounts almost nobody uses, in Choosing and Using the Tools.

F1 is the harmonic mean of precision and recall.

From When accuracy lies, in Machine Learning, Foundations.

A model can produce a fabricated citation: a quote that reads plausibly and is not on the page.

From Making the gaps visible, in AI for a Small Business.

Examples are the most expensive way to give an instruction, so choose two to four that differ from each other, include the refusal case, put the most representative one last, balance the labels, and re-test periodically whether they still beat the instruction that could replace them.

The key idea of What examples cost, and what they buy, in Context Engineering. Also taught in AI, Actually Explained and Running Models Yourself.

Fine-tuning is good at teaching form and poor at adding facts, and confusing the two is the commonest expensive mistake in applied work.

The key idea of Fine-tuning: what it changes and what it cannot, in How a Language Model Actually Works. Also taught in The Open Model Ecosystem and Fine-Tuning, and When Not To.

The idea, in its 2023 form, replaces attention with a fixed-size state (a block of numbers of constant size) that is updated once per token.

From Beyond the transformer: what the alternatives trade, in How a Language Model Actually Works.

The front matter (the YAML block at the top) is structured, is used by tooling, and is the hardest part to make vague. license, base_model, language, pipeline_tag, library_name, datasets, tags.

From Reading a model card for what it does not say, in The Open Model Ecosystem.

GGUF (llama.cpp) is a file container plus a family of quant schemes.

From Quantisation, properly, in Running Models Yourself.

A GOP (group of pictures) is one I-frame and the dependent frames that follow it.

From Frames that stand alone, and frames that do not, in Motion, Sound and Music.

The nudging procedure is called gradient descent, and the arithmetic behind it is first-year calculus, but you do not need the arithmetic to hold the shape: guess, compare, adjust, repeat.

From Nearly all of it is prediction, in AI, Actually Explained. Also taught in Machine Learning, Foundations.

Grading is what you do after that, deliberately moving the colour away from neutral for effect.

From A look is a mapping, and a LUT is a table of it, in Editing an Image.

A colour page applies them as a graph: a set of connected operations where you choose what feeds what.

From Why grading happens in a graph, in Editing Video.

The second is called grounding, or retrieval-augmented generation when people want it to sound technical.

From Grounding: answering from your material rather than its memory, in AI for a Small Business.

Open weights ship with trained refusals and nothing else, so serving them to others means choosing deliberately among model alignment, classifier passes that cost a real forward pass each, and deterministic quotas, and treating prompt injection as unsolved, defended by architecture rather than by instructions.

The key idea of You are now the moderation team, in Running Models Yourself. Also taught in The Open Model Ecosystem.

CorelDRAW and several plotter drivers offer a width called hairline.

From A path has no thickness, in Vectors, Logos and Things That Get Printed.

Every ten steps the value falls by roughly the same factor, about 0.35. That is exponential decay, and it has a half-life: the number of steps to halve.

From Exponentials, decay, and the memory of a moving average, in The Maths You Actually Need.

A model produces wrong answers in exactly the same confident voice as right ones; check consequences, not tone.

The key idea of Hallucination, and how to check, in AI, Safety and What Goes Wrong. Also taught in AI, Actually Explained, How a Language Model Actually Works and Building with AI.

The model does this many times in parallel, with separate sets of weights called heads.

From Attention, without a single equation, in AI, Actually Explained.

Flex / PVC banner: the standard hoarding material. Needs eyelets every 500 mm or so and a hem (a folded, welded edge) for strength. Add 100 mm of extra material all round for the hem, and keep artwork out of it.

From Signage, viewing distance and the seams, in Vectors, Logos and Things That Get Printed.

To aggregate correctly you need the underlying distribution, which is why metrics systems record latency as a histogram (a set of counters, one per bucket) and compute percentiles from the summed buckets at query time.

From Four numbers that describe any service, in Shipping It.

Imposition is arranging the pages on the sheet so they come out in order after folding.

From Page counts, creep and the spine, in Vectors, Logos and Things That Get Printed.

A convolution reverb needs an impulse response: a recording of how the room responds to a very short sound.

From Making a replaced line sound like it belongs, in Editing Video.

A large model shown three examples inside your message often handles the fourth correctly. This is called in-context learning, and it looks like fast learning.

From Why "learning" is a borrowed word, in AI, Actually Explained.

Indemnification is what Adobe, Getty and Shutterstock offer on business plans: if a claim arrives about the training data behind their generator, they will stand behind it.

From What you can sell, and what you must disclose, in Image Generation, in Practice.

CPU generation speed is set by memory channels and memory speed rather than cores: populate every channel, thread to physical cores, and expect cores to help prompt processing while doing almost nothing for generation.

The key idea of CPU inference, honestly, in Running Models Yourself. Also taught in Shipping It, The Open Model Ecosystem and Hugging Face, End to End.

Inpainting is the right tool for a wrong region.

From Masks, feathering and the context window, in Making Things with AI. Also taught in Image Generation, in Practice.

SFTTrainer renders chat templates, masks the prompt, packs and truncates conversations for you, so the remaining decisions are about data: a few hundred diverse, well-written examples in the messages format beat tens of thousands of scraped ones, and the register of your responses becomes the model's register.

The key idea of TRL, and training on conversations, in Hugging Face, End to End.

The robust approach is to compose for the intersection: the region that survives every crop you will need is a tall-ish central box, and anything load-bearing goes in it.

From Something is always sitting on top of your picture, in Making It Say Something.

An iterable is anything iter() accepts: a list, a string, a dict, a file, your Conversation with its __iter__.

From The iterator protocol: what for actually does, and why a generator is empty the second time, in Python, from Zero, for AI.

iter(things) calls things.__iter__() and gets back an iterator: an object with a __next__ method. next(it) calls it.

From The iterator protocol: what for actually does, and why a generator is empty the second time, in Python, from Zero, for AI.

Refusal is a trained behaviour with a distribution rather than a rule, so capability survives inputs that are far enough from the training examples, which means a product's real boundaries must live in tool permissions and data scoping, never in the prompt.

The key idea of Why safety training is a surface, not a wall, in AI, Safety and What Goes Wrong. Also taught in How a Language Model Actually Works.

Kerning is the space between one specific pair.

From Why the printer asked you to outline the fonts, in Vectors, Logos and Things That Get Printed.

A kill fee is an agreed payment if the client terminates before completion.

From Silence, holds and kill fees, in Working for Yourself.

The part that surprises people is the KV cache: the model's working memory for the conversation so far.

From Working out what your machine will run, in Choosing and Using the Tools. Also taught in How a Language Model Actually Works, The Open Model Ecosystem and Running Models Yourself.

When you upload, the file is decoded and re-encoded into a ladder: a set of renditions at different resolutions and bitrates, so a viewer on a weak connection gets 360p and a viewer on fibre gets 2160p, and the player can switch between them mid-stream without stopping.

From What the platform does to your upload, in Editing Video.

Lanczos is a windowed sinc function, usually over 36 or 64 samples. It is the sharpest of the classical filters and it rings: a bright rim on the light side of a hard edge and a dark rim on the dark side, exactly the shape of an over-sharpening halo. On photographs this is usually welcome. On a logo against flat colour it is a defect.

From Nearest, bicubic, Lanczos: what the resize dialogue is computing, in Editing an Image.

A language model is a very large set of numbers, called weights.

From What the model computes, in AI, Safety and What Goes Wrong. Also taught in How a Language Model Actually Works, Building with AI and AI for Students.

Large language models are a slice of that innermost ring.

From Four rings, one inside the next, in AI, Actually Explained.

Latency is how long one person waits.

From Throughput and latency are different problems, in Hugging Face, End to End. Also taught in How a Language Model Actually Works, Shipping It and Knowing If It Works.

Diffusion happens in a compressed latent grid roughly sixty-four times smaller than the image, and the autoencoder that compresses and restores it is the source of a specific family of defects.

The key idea of The small space the picture is built in, in Making Things with AI. Also taught in Video with AI.

Leakage is a feature that contains information which will not exist at prediction time.

From Features beat models, in Machine Learning, Foundations. Also taught in Machine Learning, Foundations, Data, SQL and Getting to the Answer and Working in AI.

Mini-batch size trades gradient noise against step count, and the learning rate must match the scale of the features, which is why standardising inputs fixes more convergence problems than raising the iteration limit.

The key idea of Batches, epochs, and the two knobs that decide everything, in Machine Learning, Foundations. Also taught in The Maths You Actually Need and Fine-Tuning, and When Not To.

A limiter is a compressor with a very high ratio and a very fast attack, used as a ceiling rather than as shaping.

From What a compressor does, and what it does to the gaps, in Editing Video.

Lineart is a learned edge extractor that produces drawing-like lines.

From Canny, lineart, depth, softedge, pose, normal, in Image Generation, in Practice.

Serving 90% of people quickly and telling 10% to come back is a better outcome than serving nobody. This is called load shedding and it is the single most useful overload behaviour to configure in advance, because you will not be configuring anything thoughtfully at the time.

From Finding your ceiling before your users do, in Shipping It.

Log loss is the mean of -ln(p_assigned_to_the_truth) and punishes confident mistakes savagely: being wrong at 0.9 costs -ln(0.1) = 2.30, being wrong at 0.6 costs -ln(0.4) = 0.92.

From Making a 0.8 mean eighty per cent, in Knowing If It Works.

A LoRA is a small set of weight adjustments layered onto a checkpoint: a character, a style, a concept.

From A model is four kinds of file, and mixing them up wastes an evening, in Image Generation, in Practice. Also taught in The Open Model Ecosystem, Fine-Tuning, and When Not To and Image Generation, in Practice.

Average error: ₦22,500. That single number is called the loss.

From What a machine actually learns, in Machine Learning, Foundations.

A loss function is a scalar.

From The loss is where you state what a mistake costs, in Machine Learning, Foundations.

LUFS (Loudness Units relative to Full Scale) is a measurement that applies a filter modelling the ear's frequency response and then averages energy over time.

From Loudness, measured rather than guessed, in Editing Video. Also taught in Motion, Sound and Music.

Luminance noise is the grain-like variation in brightness.

From Sharpening adds no detail, and it goes last, in Editing an Image.

Machine learning earns its cost only when the decision rule is real but unwritable; when the rule is known, exact, or the data is too thin to tell a good model from a lucky one, code the rule.

The key idea of The problems where the right answer is a rule, in Machine Learning, Foundations.

A capture device or a piece of software creates a manifest: a structured record: what made this, when, with what settings, and what actions were applied.

From Signed provenance, and where the chain breaks, in Making Things with AI.

A mascot (a drawn character) is sometimes asked for and is a different commission.

From Five kinds of mark, and what each one costs, in Vectors, Logos and Things That Get Printed.

A bug report with a minimal reproduction: the version numbers, the smallest failing input, the expected and actual behaviour. Maintainers can tell instantly whether somebody did the reduction work, and reducing a bug is the diagnosis skill in miniature.

From What a repository proves, and what it stopped proving, in How Software Development Changed.

A sparse mixture-of-experts model has a large parameter count but activates only a fraction per token, which decouples what it knows from what it costs to run.

The key idea of Mixture of experts: big model, small bill, in How a Language Model Actually Works. Also taught in The Open Model Ecosystem and Running Models Yourself.

MKV (Matroska) is the open box that holds anything: any codec, unlimited subtitle and audio tracks, chapters.

From The box and the codec are two different decisions, in Editing Video.

The model card is a file called README.md inside it.

From The Hub is git, and that explains everything else, in Hugging Face, End to End. Also taught in Choosing and Using the Tools, The Open Model Ecosystem and Running Models Yourself.

Motion blur means the subject's edge is semi-transparent across several pixels.

From Masks, and the honest cost of rotoscoping, in Editing Video.

After the if, the checker knows row is a dict (this is called narrowing), and the error disappears. Optional[dict] from typing means the same as dict | None; the bar form needs Python 3.10 or later.

From Running a type checker: the bugs it finds that tests do not, in Python, from Zero, for AI.

Net 30 means payment is due thirty days after the invoice date.

From The document that has to survive an accounts system, in Working for Yourself.

A neural network learns its own features; the loss, the splits, the metrics and the thresholds do not change.

The key idea of Neural networks as the natural next step, in Machine Learning, Foundations.

NF4 is a 4-bit format whose levels are spaced to match the roughly normal distribution of neural network weights, rather than spaced evenly.

From QLoRA: training a 7B on the card you have, in Running Models Yourself.

Non-destructive editing is the discipline of never modifying original pixels: only stacking instructions on top of them.

From Every edit should still be undoable next week, in Editing an Image.

Normal, or any other blend mode, makes the group isolated.

From Why an adjustment stopped working when you put it in a folder, in Editing an Image.

On is what you want for a logo. Shrink the mark to 10 mm and the outline shrinks with it.

From A path has no thickness, in Vectors, Logos and Things That Get Printed.

One-shot is a single demonstration.

From Zero, one, or five: how many examples, in Getting Real Work out of a Model.

When a lab calls a model open it almost always means the weights can be downloaded, not that the data, code or licence are open, and the difference decides what you can audit, run and rely on.

The key idea of Open weights, closed weights, and what "open" is hiding, in AI, Actually Explained. Also taught in AI, Safety and What Goes Wrong, Choosing and Using the Tools and The Open Model Ecosystem.

Optical size is the idea that a typeface should be drawn differently for different sizes.

From Letter spacing, and the sizes type was drawn for, in Design, before Any Software.

An orphan is a lone first line of a paragraph stranded at the bottom of a column.

From Setting a paragraph somebody can read, in Vectors, Logos and Things That Get Printed.

An output intent: a named characterisation of the intended printing condition, such as FOGRA39 or Japan Color 2001 Coated.

From PDF/X, and what each flavour forbids, in Vectors, Logos and Things That Get Printed.

Compare training error with validation error; the gap between them, not the score itself, names the problem.

The key idea of Overfitting, underfitting, and how to tell, in Machine Learning, Foundations.

The inner one is the package: a folder Python can import as a unit. run.py at the top is the entry point.

From Laying out a project so the imports keep working, in Python, from Zero, for AI.

Parameters are the numbers in the file, fixed after training: the model's size.

From Five words that hide more than they say, in AI, Actually Explained. Also taught in Machine Learning, Foundations and The Open Model Ecosystem.

Pass Through means the group is purely organisational.

From Why an adjustment stopped working when you put it in a folder, in Editing an Image.

A port is a numbered door on a machine. localhost:3000 means door 3000 on this computer. That is why, when two projects start at once, the second either refuses to start (a plain Node server does) or quietly moves to the next port, as next dev does to 3001, and you can end up looking at the wrong project.

From What a web app is actually made of, in Building Apps with AI.

Pre-delay is the gap between the direct sound and the start of the reverb.

From Reverb is information about where, in Motion, Sound and Music.

Precision, or positive predictive value: of the people the model approved, who actually repaid.

From Counting the four ways to be right and wrong, in AI, Safety and What Goes Wrong. Also taught in Python, from Zero, for AI, The Maths You Actually Need and Running Models Yourself.

Preference training optimises a learned proxy for what humans approve of, and every characteristic behaviour of chat assistants (the hedging, the length, the agreeableness) comes from optimising that proxy rather than truth.

The key idea of Preference training, from reward models to DPO, in How a Language Model Actually Works.

Pretraining is where knowledge comes from; the later stages mostly shape behaviour, not facts.

The key idea of Pretraining, fine-tuning, and preference training, in How a Language Model Actually Works. Also taught in The Open Model Ecosystem, Running Models Yourself and Fine-Tuning, and When Not To.

Developer instructions, user requests and processed content share one text channel with no reliable separator, so any content a model reads can carry commands, and unlike SQL injection, there is no parameterisation that fixes it.

The key idea of Instructions hidden in the content, in AI, Safety and What Goes Wrong. Also taught in AI, Actually Explained, Agents and Automation and AI for a Small Business.

Provenance is a record of where a file came from and what was done to it.

From Provenance, watermarks and what to tell people, in Motion, Sound and Music.

Quantisation works because rounding errors average out across long dot products, which is exactly why it fails where they cannot (small models, multi-step arithmetic, long-context recall and weakly represented languages) and why perplexity is the wrong number to choose a level by.

The key idea of Quantisation: what the numbers lose, in How a Language Model Actually Works. Also taught in AI, Actually Explained, The Maths You Actually Need and Choosing and Using the Tools.

The result of all this is what editors call a radio edit: the entire piece cut for sound only, with the pictures ignored or left as whatever happened to be under the words.

From Editing the words before you touch a picture, in Editing Video.

Measure retrieval, groundedness and relevance separately, and always include questions your corpus cannot answer so you find out whether the system says so.

The key idea of Three questions, not one, for a retrieval system, in Knowing If It Works. Also taught in Vectors, Logos and Things That Get Printed.

A matrix has a number of rows and a number of columns. It also has a rank: the number of linearly independent rows, which always equals the number of linearly independent columns.

From Rank, and why a huge matrix can be two small ones, in The Maths You Actually Need.

Supercardioid and hypercardioid. Narrower front, but they have a rear lobe: a zone of renewed sensitivity directly behind. Point one at a speaker in a room and you are also recording whatever is behind you. People place monitors and noise sources by cardioid logic and then wonder where the bleed came from.

From Pressure, gradient, and everything that follows, in Motion, Sound and Music.

Reasoning models genuinely improve multi-step accuracy by generating working before answering, but experiments show the written trace can omit what actually changed the answer, so it is not an audit trail.

The key idea of Models that think out loud, and what that does not prove, in AI, Actually Explained. Also taught in Getting Real Work out of a Model.

True positive rate, also called recall or sensitivity: of the people who would have repaid, who got approved.

From Counting the four ways to be right and wrong, in AI, Safety and What Goes Wrong.

When correctness can be verified by a program, the reward is exact rather than a learned proxy, and models trained this way learn to spend more tokens working before answering.

The key idea of Training against answers you can check, in How a Language Model Actually Works.

Rendering is where the HTML is assembled. On the server, before it is sent, or in the browser afterwards by JavaScript. Both are normal, and mixing them badly is where a page turns up empty for one second and full the next.

From What a web app is actually made of, in Building Apps with AI.

The repository is the chain of commits already made.

From What a commit actually is, in Building Apps with AI.

Retrieval-augmented generation: the same search, with the results pasted into a prompt.

From A list of numbers where near means similar, in Hugging Face, End to End.

A RIP (raster image processor) is the software that turns your PDF into the pattern of dots a press, platesetter or large-format printer lays down.

From What a RIP actually does with your file, in Vectors, Logos and Things That Get Printed.

The safe area (the part of the frame that is reliably visible) is much smaller than the frame, and it was never in your editing preview.

From Something is always sitting on top of your picture, in Making It Say Something.

Saturation: how full the constrained resource is. Usually not CPU. Usually the database connection pool, the queue depth, or memory.

From Four numbers that describe any service, in Shipping It.

If nothing highlights, it is a scan: a photograph of a page wearing a PDF wrapper.

From PDFs, scans and the black rectangle that hides nothing, in AI at Work.

The search area is where the tracker looks.

From What a tracker can and cannot see, in Motion, Sound and Music.

A selects timeline (also called a string-out) is a timeline containing only the good bits, in the order they were shot, with nothing removed but the dross.

From Watching an hour so you can find any second of it, in Editing Video.

Put reverb on a send: the dry voice goes to the output, a copy goes to a reverb bus, and you mix the wet return underneath.

From Reverb is information about where, in Motion, Sound and Music.

The settings page is where the defaults live, and defaults beat policies.

From Reading a privacy policy in ten minutes, in AI, Safety and What Goes Wrong.

Larger presses print 8, 16 or 32 pages per sheet, and a 16-page folded sheet is called a signature or a section.

From Page counts, creep and the spine, in Vectors, Logos and Things That Get Printed.

Any recorded song carries at least two copyrights: the composition (melody, harmony, lyrics) and the sound recording (the specific captured performance).

From When a tune becomes somebody else's, in Making Things with AI.

In video compositing it is called spill suppression and every keyer has one.

From Three equations, seven unknowns: why hair is hard, in Editing an Image.

A spot colour is an ink mixed to a recipe before it goes on the press, and printed from its own unit.

From A spot colour is a plate, and a plate is money, in Vectors, Logos and Things That Get Printed. Also taught in Design, before Any Software.

Do it again and again, and for most graphs the vector settles to one that no longer changes, the stationary distribution: the fraction of time a long random walk spends at each node.

From A graph is a matrix, and its paths are its powers, in The Maths You Actually Need.

A stem is a bounce of one bus on its own: the whole film's dialogue as a single file, the whole film's music, the whole film's effects.

From A track layout that survives a revision, in Editing Video.

A subclip is a named fragment of a longer clip that behaves like its own clip in the bin.

From Watching an hour so you can find any second of it, in Editing Video.

Preference training rewards agreement and confidence, so models validate premises and reverse correct answers under doubt: ask for the strongest case against your position, and never tell it which side is yours.

The key idea of Systems that are optimised to please you, in AI, Safety and What Goes Wrong. Also taught in How a Language Model Actually Works.

Everything in the context is the same kind of thing (tokens) so text from a document or a web page can act on the model exactly as an instruction does.

The key idea of Why there is no boundary between instructions and data, in How a Language Model Actually Works. Also taught in AI, Actually Explained, Building with AI and Agents and Automation.

A tangent is where two edges just touch, or barely overlap, or come within a hair of each other.

From Why a near-touch is worse than an overlap, in Design, before Any Software.

There is also a targeted adjustment tool: the hand icon in the Curves panel.

From Brightness/Contrast is one knob for a problem with five, in Editing an Image. Also taught in Editing an Image.

The setting that controls how adventurous that sampling is is called temperature: near zero it almost always takes the top choice and answers are nearly repeatable; higher and it wanders.

From Run it five times, in Choosing and Using the Tools. Also taught in How a Language Model Actually Works.

Every channel in this module has a threshold: a minimum quantity of sustained effort below which it returns nothing at all, not proportionally less.

From Thresholds, and one at a time, in Working for Yourself.

Throughput is how many items you finish per minute.

From Throughput and latency are different problems, in Hugging Face, End to End.

Time to first token is what an interactive user feels; tokens per second is what a batch job cares about.

From Renting open weights by the token, in The Open Model Ecosystem.

The chunks are called tokens.

From What it is actually doing, in AI for Students. Also taught in AI, Actually Explained.

Tokens per second is how fast text then appears.

From Streaming, and the difference between fast and feeling fast, in Building with AI.

Top-p, or nucleus sampling, takes a different approach: keep only the smallest set of tokens whose probabilities add up to p, commonly 0.9, and draw from those.

From Why the same question gives different answers, in AI, Actually Explained.

Tracking, or letter-spacing, applies the same adjustment uniformly across a run of text.

From Letter spacing, and the sizes type was drawn for, in Design, before Any Software.

Training data is a stack of human choices, and each one shows in the answers.

The key idea of Where the data came from, in AI, Safety and What Goes Wrong. Also taught in AI, Safety and What Goes Wrong, How a Language Model Actually Works and The Open Model Ecosystem.

In 2017 a group of researchers published a design for handling sequences of words called the transformer.

From What changed around 2022, in AI, Actually Explained. Also taught in How a Language Model Actually Works and Hugging Face, End to End.

A trend: the slow direction over a year or more.

From Patterns you can name, and the promotion that poisons the history, in AI for a Small Business.

True positive rate, also called recall or sensitivity: of the people who would have repaid, who got approved.

From Counting the four ways to be right and wrong, in AI, Safety and What Goes Wrong.

That overall greyness is called typographic colour, and the word colour here has nothing to do with hue.

From The grey of a paragraph, in Design, before Any Software.

Compare training error with validation error; the gap between them, not the score itself, names the problem.

The key idea of Overfitting, underfitting, and how to tell, in Machine Learning, Foundations.

Uniformity means everything looks the same.

From Consistency is not uniformity, in Design, before Any Software.

Diffusion-based upscalers re-generate rather than sharpen, so never point one at text, a logo, a product detail or a face you need to stay recognisable.

The key idea of Upscaling either restores detail or invents it, in Image Generation, in Practice. Also taught in Making Things with AI and Editing an Image.

The VAE is the decoder that turns the small latent image into pixels.

From A model is four kinds of file, and mixing them up wastes an evening, in Image Generation, in Practice.

The best of many noisy validation estimates is biased upward by roughly the standard error times the square root of twice the log of the number of comparisons, so the count of experiments belongs in the report alongside the score.

The key idea of A validation set you have used two hundred times, in Machine Learning, Foundations.

Variance is how much the prediction moves between fits.

From Bias and variance, with numbers attached, in Machine Learning, Foundations.

Below about a million vectors a brute-force dot product on a CPU is exact and fast enough, and above it an approximate index buys speed by giving up a measurable fraction of recall, which you must measure against exact results rather than assume away.

The key idea of Vector search: when NumPy is enough, and what an index trades away, in Building with AI.

Text-to-speech has solved pronunciation and timbre but not interpretation, so synthetic voice holds up for short informational lines and degrades over long-form narration where a listener starts noticing that nothing is being meant.

The key idea of Synthetic voice, and where it stops being good enough, in Video with AI.

Voice mode is a spoken conversation: you talk, it talks back, you can interrupt.

From Voice, screen readers and hands-free use, in Choosing and Using the Tools.

A generator can tilt its own random choices (which pixel value, which token) toward a secret pattern that a paired detector can count, which proves the generator's origin when it cooperates and proves nothing when it does not.

The key idea of How a watermark is hidden inside the output, in AI, Actually Explained. Also taught in Editing an Image.

WebM is the browser box for VP9 and AV1.

From The box and the codec are two different decisions, in Editing Video.

The numbers are called weights, or parameters.

From It is a file, and here is how big, in AI, Actually Explained. Also taught in AI, Actually Explained, AI, Safety and What Goes Wrong and How a Language Model Actually Works.

A widow is a lone last line carried to the top of the next column, or, in the stricter sense, a single word alone on the last line of a paragraph.

From Setting a paragraph somebody can read, in Vectors, Logos and Things That Get Printed.

A window is a drawn shape (circle, rectangle, polygon, or a freehand curve) with a softness setting.

From Changing one part of the frame, in Editing Video.

A hundred hyperparameter configurations, scored on a validation set with a standard error of 1 point, will produce a winner about 2.5 points better than it is. This is the winner's curse: the act of choosing the best guarantees that its score overstates it.

From Regression to the mean, and the winner's curse in model selection, in The Maths You Actually Need.

The working tree is the folder you can see.

From What a commit actually is, in Building Apps with AI.

What you actually perceive as size is the x-height: the height of a lowercase x, and therefore of most of the letters in any word.

From Your font is fine, your line is too long, in Design, before Any Software.

Zero-shot is a prompt with no examples.

From Zero, one, or five: how many examples, in Getting Real Work out of a Model. Also taught in Making Things with AI.