Every word, plain first.

Comparing two versions of a model or prompt with real users to see which wins.

The fraction of predictions the model got right.

The output of a neuron or layer in the middle of a neural network.

The nonlinear twist applied to a neuron's output so the network can learn complex patterns.

Tweaking a model's internal activations to change its behavior at inference time.

A small module you bolt onto a frozen base model to specialize it.

Inputs specifically designed to fool an AI.

A weird-looking string that, when appended to a prompt, reliably jailbreaks an LLM.

An AI that can plan, take actions, and use tools to achieve a goal — not just chat.

The repeat cycle of think-act-observe that drives agentic AI.

An IDE setting where the AI can run tools, read files, and iterate on its own — versus single-turn edit mode.

A benchmark suite for evaluating agent behavior across diverse environments.

AI that plans and acts over many steps, using tools to get things done.

Artificial general intelligence — AI that can do most human cognitive tasks as well as humans.

Computer systems that do things that usually need human thinking, like recognizing faces or writing stories.

Shorthand for the EU AI Act, the world's first comprehensive AI law.

A 2022 White House blueprint laying out principles for safe and rights-respecting AI.

How organizations and governments oversee AI development and deployment.

Vercel's open-source toolkit for building AI apps in JavaScript and TypeScript.

Algorithmic Impact Assessment — a structured review of potential harms from an AI system.

An open-source command-line coding agent that pair-programs with you over a Git repo.

AI Safety Institutes — government bodies (UK, US, and others) that evaluate frontier AI.

A step-by-step recipe a computer follows to solve a problem.

Alibaba's Qwen model family, a leading open-weights LLM series.

An alternative position encoding that biases attention toward nearby tokens — helps extrapolation.

Making sure AI actually does what we want, in a safe and helpful way.

The cost in capability you pay to make a model safer and more aligned.

The AI safety company behind the Claude model family.

A widely used permissive license that also grants patent rights.

A way for programs to talk to each other — how apps use AI models.

Apollo Research — a nonprofit focused on detecting deceptive and scheming AI behavior.

Now called METR. Originally the evaluation arm of the Alignment Research Center.

Francois Chollet's grid-puzzle benchmark designed to measure true reasoning, not memorization.

Made-to-be-seen creative work — drawings, paintings, photos, videos.

Made by humans, not grown in nature.

The full name for AI — when computers act like they can think or learn.

Artificial superintelligence — AI significantly smarter than humans across the board.

An AI designed to help you do tasks, like a super-smart intern.

The mechanism that lets AI focus on the most important parts of the input.

Plotting which tokens an attention head looks at, to get intuition for what it does.

Sound data — like a recording of a voice, song, or noise.

Microsoft's multi-agent conversation framework for LLM-driven workflows.

An agent that runs long-running tasks mostly on its own, checking in only when needed.

Activation-aware Weight Quantization — a strong weight-only quantization method for LLMs.

NVIDIA's next-generation AI GPU after the H100, with more memory and speed.

A hidden trigger in a trained model that makes it behave badly only when a secret phrase appears.

The algorithm that figures out how much each weight contributed to the error.

Training many models on different random subsets of data and averaging them.

A small group of training examples the model processes together.

A cheaper, slower way to send lots of requests — results within 24 hours.

A layer that rescales activations during training to speed things up and stabilize them.

A classic decoding strategy that keeps a beam of top-B sequences at each step.

AWS's managed LLM platform — Claude, Llama, Titan, and others behind one API + IAM.

A standardized test used to compare AI models.

Google's 2018 encoder-only model — a big deal for search, classification, and embeddings.

Brain float 16-bit — the training precision of choice on modern accelerators.

When an AI treats some people or topics unfairly because of patterns in its training data.

A score for translation quality, based on matching words and phrases to reference translations.

Training many weak models in sequence so each fixes the mistakes of the last.

Byte-pair encoding — a common tokenization method that merges frequent byte pairs into tokens.

A tokenizer that can fall back to individual bytes, so every possible input can be represented.

An industry standard for signing content with tamper-evident info about how it was made.

How well a model's confidence matches its actual accuracy.

A sensor that captures pictures or video, often used as the 'eyes' of an AI.

Latent ability that's already in a model but only surfaces with the right technique.

How much CO2 an AI model's training and use produces.

A European conformity mark required to sell many products in the EU, now extending to AI.

The US Committee on Foreign Investment — it reviews foreign deals with national security implications.

Asking the model to show its reasoning step by step before answering.

An AI you talk to in a chat window, like texting a helper that knows a lot.

LMSYS's platform where users compare two model responses and vote, producing Elo rankings.

OpenAI's chat app — the product most people first met AI through.

A saved snapshot of a model's weights during training.

The DeepMind recipe for balancing model size and training tokens for best compute efficiency.

The business of selling semiconductors across borders — now heavily policed for AI chips.

A small subnetwork of neurons and attention heads that together compute a specific behavior.

A formal way of saying 'this info came from here'.

Predicting which category something belongs to.

An AI that sorts things into categories, like spam vs. not spam.

Anthropic's family of AI assistants, known for safety, long context, and coding skill.

Anthropic's agentic coding tool — Claude running in your terminal with filesystem and tool access.

Anthropic's smallest, fastest, cheapest Claude model — great for high-volume tasks.

Anthropic's flagship Claude model — smartest and slowest, for the hardest problems.

Anthropic's mid-tier Claude model — strong and fast, widely used in production.

Code you install to talk to an API from your app.

OpenAI's vision-language model that produces joint embeddings for images and text.

A model you can only use through an API — you can't download the weights.

Grouping similar things together without any labels.

The step that turns a model's proposed edit into an actual file change in the user's project.

Mistral's code-focused open-weights model, popular for self-hosted coding assistants.

An AI company focused on enterprise LLMs, search, and embeddings.

The delay when a model is loaded onto a GPU before it can serve traffic.

Using a model or tool to make money — something many AI licenses restrict.

The raw processing power needed to train or run AI.

Regulating AI by limiting, monitoring, or allocating the hardware needed to train big models.

Regulatory or export rules that limit access to AI compute — often aimed at specific countries.

A machine that runs programs and crunches numbers really fast.

An AI agent that controls a real computer via screenshots, clicks, and keyboard input.

Running workloads in hardware that keeps the cloud provider itself from seeing your data.

The structured process of verifying a product meets a regulation's requirements.

A grid showing where a classifier got predictions right and wrong.

Anthropic's approach to training models to follow a written set of principles.

A lightweight model that checks outputs against a safety constitution in real time.

The specific rules about what kinds of content an AI service allows or blocks.

Everything the AI is considering right now — your prompt, chat history, uploaded files.

Another name for prompt caching — reusing long context computations across requests.

When earlier content in the context biases or manipulates later model behavior.

Designing what goes into the model's context — not just the prompt but docs, memory, tool results.

The maximum amount of text the model can see at once, measured in tokens.

Packing new requests into a GPU batch as soon as slots free up, instead of waiting for a whole batch.

A back-and-forth chat with an AI, made of your messages and its replies.

All the prior messages in a chat that the model can see when answering.

A neural network specialized for images, using filters that slide across the picture.

A duplicate of something — AI can make copies that look close to the original.

The legal right that says only the creator can copy, share, or sell their original work.

A way to measure how similar two vectors are by the angle between them.

Making something new — with AI, that 'something' can be a picture, song, story, or plan.

Making new stuff that's interesting, original, or fun.

Multi-agent framework where role-specialized agents collaborate on tasks.

The loss function used for training most classifiers and language models.

Training and testing multiple times on different data splits to get a more reliable score.

AI-native code editor (VS Code fork) with deep model integration for multi-file edits.

Cursor's multi-file agent mode that plans and applies changes across a codebase from a single prompt.

OpenAI's image-generation model, integrated into ChatGPT.

Testing whether a model could meaningfully help with serious harms — biosecurity, cyberattack, autonomy.

The information an AI learns from — text, images, sounds, numbers, or anything else a computer can read.

Expanding your dataset by tweaking copies — flipping images, paraphrasing text.

Each GPU holds a full model copy and processes different data, synchronizing gradients.

The steps that move raw data to the form a model can train on.

Sneaking bad data into a training set to corrupt the model's behavior.

Where data came from and how it got to you.

A structured collection of data used for training or evaluating an AI.

A short document describing a dataset — what's in it, where it came from, and its limits.

An alignment technique where two AI copies argue a question and a judge decides.

A hypothetical (and worrying) failure mode where a model fakes being aligned during training.

A flowchart-like model that makes decisions by asking yes/no questions.

The part of a model that generates output, one token at a time.

A transformer that's just a decoder stack — the setup behind GPT and most modern chatbots.

The phase where the model generates output tokens one at a time.

A fake photo, audio, or video made by AI that looks or sounds like a real person.

A Chinese lab whose open-weights MoE models stunned the industry with efficient training.

Microsoft's open-source library for scaling deep learning training and inference.

A cap on how much reasoning a model is allowed to do before it must answer.

Instructions from an app developer, sitting between system and user in trust.

Cognition Labs' autonomous software engineering agent that runs tasks in a sandbox.

A math framework that limits how much any single training example can influence a model.

Hugging Face's library for running and training diffusion models like Stable Diffusion.

An image-generation model that starts with noise and gradually denoises into a picture.

Running the prompt-processing (prefill) stage and the decoding stage on different GPUs.

Teaching a small model to copy a big one, so you get speed with most of the smarts.

Using unauthorized distillation to clone a proprietary model.

When the data in production differs from the training data — a common cause of model failure.

Data Protection Impact Assessment — required under GDPR for high-risk data processing.

Direct Preference Optimization — a simpler alternative to RLHF that skips the reward model.

Making a picture, usually with lines — AI can do it, and you can too.

When model performance degrades over time because the world changed but the model didn't.

Randomly turning off neurons during training to keep the model from overfitting.

A voice-AI company famous for realistic text-to-speech and voice cloning.

Coaxing out a model's hidden abilities through clever prompting, scaffolding, or fine-tuning.

A chess-style rating system used to rank AI models from pairwise comparisons.

Turning a word, sentence, or image into a list of numbers that captures its meaning.

A model specialized for turning text (or images) into semantic vectors.

A skill that suddenly appears at a certain model scale but is absent at smaller scales.

The part of a model that reads and understands input.

A transformer architecture with a separate encoder stack and decoder stack — great for translation.

The electricity AI uses — growing fast as data centers scale up.

Combining multiple models so their mistakes cancel out.

A measure of how uncertain or surprising a probability distribution is.

One full pass through the training data.

The European Union's comprehensive AI law, in force since August 2024.

Straight-line distance between two points — or two vectors.

The framework that runs a model against a dataset, scores outputs, and aggregates metrics.

Testing a model's quality — beyond benchmarks, using real tasks and user feedback.

One data point used to teach or prompt an AI — like a labeled photo or a sample answer.

In an MoE, the fact that only a few experts fire per token — most are skipped.

Rules restricting which technologies can be sold to which countries — a big deal for AI chips.

Anthropic's feature that lets Claude generate a long internal reasoning trace before its final answer.

A single number combining precision and recall — good when you care about both.

Something that's actually true and can be checked.

A US legal doctrine letting you use copyrighted material in limited ways without permission.

Making sure AI treats different groups of people equally.

Not real — made up or generated, often to trick someone.

A specific concept or pattern detected by a neural network.

Training a shared model across many devices without centralizing the raw data.

A network where data flows one direction from input to output, no loops.

Telling the AI what it did right or wrong so it (or its makers) can improve.

Giving the model a handful of examples in the prompt to show what you want.

A check that blocks certain inputs or outputs from reaching the user.

Fill-in-the-middle — a training technique letting a model complete text between a prefix and suffix.

Taking a pre-trained model and doing extra training on your own data.

A managed service that fine-tunes provider models on your data without you touching GPUs.

A memory-efficient attention algorithm that's much faster on modern GPUs.

A floating-point operation — the basic unit of compute, used to measure model size and cost.

Black Forest Labs' image-generation model, known for sharp text and prompt adherence.

A large, general-purpose model trained once on broad data and fine-tuned for many tasks.

An 8-bit floating-point format — modern GPUs train and run LLMs in FP8 for speed and memory.

Fundamental Rights Impact Assessment — required by the EU AI Act for some deployers of high-risk systems.

A company at the cutting edge of AI capability research, like Anthropic, OpenAI, or Google DeepMind.

The most capable, cutting-edge AI models — the ones that require special safety attention.

PyTorch's Fully Sharded Data Parallel — shards model states across GPUs for memory-efficient training.

A specific style of tool use where the model fills in arguments for a named function.

A benchmark for general AI assistants, with multi-step real-world questions.

Something fun with rules — and AI is used in tons of games.

Greedy Coordinate Gradient — an algorithm that finds adversarial suffixes that jailbreak LLMs.

Google DeepMind's flagship multimodal AI family.

Large, flexible models covered under the EU AI Act as a distinct category.

How well a model performs on new data it didn't see during training.

When AI creates something new — text, an image, a song, a video.

The original lightweight tensor library behind llama.cpp; its file format preceded GGUF.

A file format for quantized open-weights LLMs, popularized by llama.cpp.

Microsoft/GitHub's in-editor coding assistant — the original mainstream AI pair-programmer.

When an AI learns the right behavior in training but the wrong underlying goal, and it shows in new situations.

A small, hand-curated set of input-output pairs treated as ground truth for evaluation.

'When a measure becomes a target, it ceases to be a good measure' — a key AI safety warning.

Google's combined AI research arm, behind Gemini, AlphaFold, and Imagen.

Graduate-level Google-Proof Q&A — hard science questions experts can barely handle.

OpenAI's family of Generative Pre-trained Transformer models, including GPT-5.

The decoder-only transformer design popularized by OpenAI's GPT series.

A post-training quantization method for LLMs based on second-order information.

A graphics chip that's also great at the matrix math AI needs.

Grouped-query attention — heads share keys and values in small groups, balancing speed and quality.

A memory-saving trick that recomputes intermediate activations during backprop instead of storing them.

The optimization algorithm that nudges weights toward lower error during training.

Reconstructing inputs from gradients shared during distributed or federated training.

Always picking the highest-probability next token at each step — deterministic but boring.

xAI's chatbot, integrated into X (Twitter).

When a model suddenly 'gets it' long after memorizing — training beyond overfit into true generalization.

Custom-silicon inference provider competing on tokens-per-second and latency.

Tying a model's output to specific sources or data to reduce hallucinations.

A lighter cousin of LSTM with fewer gates but similar performance.

A benchmark of 8,000 grade-school math word problems used to test reasoning.

Rules or checks around AI to keep it from going off the rails — input filters, output checks, retries.

NVIDIA's workhorse AI GPU from 2022 — once dominant, now generally superseded by Blackwell (B200/GB200).

When an AI makes up facts that sound real but aren't true.

A commonsense-reasoning benchmark where models pick the most plausible sentence ending.

AI as a helper for almost anything — homework, writing, planning, coding.

A category in the EU AI Act for AI used in sensitive areas like hiring, credit, or critical infrastructure.

Encryption that lets you compute on data without decrypting it — potentially great for private AI.

Training and designing AI to tell the truth, including uncertainty and disagreement.

The GitHub of AI — a hub for open-weights models, datasets, and demos.

A classic coding benchmark of 164 Python problems used to grade LLMs.

Hypothetical Document Embeddings — query-rewrite technique that boosts retrieval recall.

A setting you pick before training that controls how the model learns.

A picture — made of tiny colored dots called pixels — that AI can see or create.

Having AI describe what's in a picture.

Making pictures from scratch with AI, usually from a text prompt.

Google DeepMind's text-to-image model, integrated into Gemini.

Copying someone's style or voice — AI is very good at imitation.

Teaching a model a task just by including examples in the prompt — no weight updates.

When hostile instructions arrive through data (a document, email, webpage) an agent reads — not from the user.

Attention heads that copy-and-extend patterns, crucial for in-context learning.

A specific pair of attention heads across layers that enables in-context pattern copying.

Running the model to get an answer — the 'use it' step, not the 'train it' step.

What it costs to actually run the model per query or per token.

A method to measure which training examples most influenced a specific model prediction.

What you give the AI — text, an image, a file, a voice clip.

Fine-tuning a base model on instruction-following examples so it behaves like an assistant.

The idea that almost any capable goal-directed AI will want power, resources, and self-preservation.

The ability to learn, reason, and solve problems.

Smart habits for staying safe online — especially around AI and strangers.

Understanding what AI models are doing inside — their reasoning, features, and behavior.

The international standard for AI management systems, like ISO 27001 for security.

Training AI by having the model plus a human decompose and solve problems together.

Tricking a model into ignoring its safety rules.

A setting that forces the model to output valid JSON.

A classic clustering algorithm that sorts data into k groups based on similarity.

A math measure of how different two probability distributions are.

A regularizer that keeps an RL-tuned model close to its pre-RL behavior.

Kuaishou's text-to-video model, a major Chinese competitor to Sora.

A memory trick that stores past attention keys and values so the model doesn't recompute them each step.

The right answer attached to an example so the AI knows what to learn.

Open-source framework for chaining LLM calls, tools, and memory into apps.

Stateful, graph-based orchestration library for multi-step LLM agents.

Written or spoken words — the main thing most AIs today understand and generate.

A giant neural network trained on huge amounts of text to generate and understand language.

How long it takes the model to start (or finish) responding.

Per-layer normalization used in the original transformer.

A public ranking of models on a benchmark.

How an AI gets better — by updating its numbers based on examples.

How big a step the optimizer takes on each update — too big and training blows up, too small and it crawls.

Predicting a number as a weighted sum of inputs — the OG machine-learning model.

Meta's open-weights LLM family, a staple of the open-source AI ecosystem.

The decoder-only transformer with RoPE, SwiGLU, and RMSNorm used in Meta's Llama models.

A C/C++ project that runs LLMs fast on normal computers, no giant cluster required.

Data framework for connecting LLMs to private/external data via indexes and retrievers.

Using a strong LLM to grade other LLM outputs during evaluation.

MLOps specifically for LLM-based applications — prompt versioning, eval, and inference ops.

A desktop app for discovering, downloading, and chatting with open-weights LLMs.

EleutherAI's toolkit for running standard LLM benchmarks reproducibly.

A simple, fast model for binary classification — still a great baseline.

Peeking at what the model would predict if you stopped at an intermediate layer.

Context windows big enough to fit many documents — 200K, 1M, even 2M tokens.

Info the AI keeps between sessions — facts about you, past conversations, saved notes.

A lightweight way to fine-tune a model by training small add-on matrices.

A number that measures how wrong the model's predictions are — training tries to make it small.

A fancy RNN with gates that help it remember long sequences.

Teaching computers by showing them lots of examples instead of writing step-by-step rules.

Running the model multiple times and picking the most common answer.

Using a long context of fake harmful examples to convince a model to break its rules.

Hiding random words and training the model to fill them in — the BERT pre-training objective.

A benchmark of competition-level math problems from high school and early college.

Model Context Protocol — an open standard for connecting AI models to tools and data sources.

The host application that connects an LLM to one or more MCP servers.

Read-only data — files, database rows, API payloads — that an MCP server exposes for the model to consume.

A program that exposes tools, resources, or prompts to an AI client over the Model Context Protocol.

The wire format MCP uses to move messages between client and server — usually stdio or HTTP/SSE.

Using interpretability tools to spot when a model is doing something weird under the hood.

Reverse-engineering a neural network down to the specific circuits that implement each behavior.

NVIDIA's original library for large-scale transformer training.

An attack that tries to tell whether a specific example was in the training set.

What the AI can remember from earlier in a conversation — or across sessions.

A guide who gives advice — and yes, AI can play that role too.

When a trained model contains an inner optimizer that may pursue goals different from the training objective.

The company (formerly Facebook) that releases the popular open-weights Llama model family.

Model Evaluation and Threat Research — a nonprofit that runs capability evaluations on frontier models.

A sensor that captures sound so AI can listen.

A popular image-generation model known for painterly, cinematic aesthetics.

A Chinese AI company behind abab and Hailuo video models.

A French AI lab making open-weights and commercial models.

A short, permissive open-source license — do almost anything as long as you keep the notice.

Training with both low-precision (BF16/FP16) and high-precision (FP32) numbers where each is best.

Mistral's open-weights mixture-of-experts model that stunned the open-source community.

A model made of many specialized 'experts', with only a few active per token — fast and scalable.

Engineering practices for running ML in production — CI/CD, monitoring, data pipelines.

Apple's ML framework for running and training models efficiently on Apple Silicon.

A benchmark of 57 subjects testing broad academic knowledge — a go-to LLM capability metric.

A type of data the AI handles — text, image, audio, video.

The actual trained AI — the big blob of numbers that can answer questions or make images.

A short document describing what a model does, how it was trained, and its limits.

When training on too much AI-generated data makes models lose diversity and degrade.

Copying a closed model's behavior by querying it a lot and training on the outputs.

An attack that reconstructs training data from a trained model.

A discipline (borrowed from finance) for managing the risks of using models in decisions.

Choosing which model to call for a given request — fast cheap model for easy stuff, big model for hard.

A collection of pre-trained models — Hugging Face Hub is the biggest.

Checking inputs and outputs for policy violations.

The mechanism that decides which experts each token goes to in a mixture-of-experts model.

When a single neuron or feature corresponds to a single, interpretable concept.

Chinese AI lab Moonshot's flagship chatbot, known for long context windows.

Multi-query attention — all heads share one set of keys and values to cut memory and speed up decoding.

A multi-turn chat benchmark graded by GPT-4 (or a similar strong judge model).

Running several parallel attention computations so the model can look at different things at once.

The basic fully connected neural network: input layer, hidden layers, output layer.

An AI that handles more than one kind of input or output — text, images, audio, video.

Organized sounds that feel like melody, rhythm, and harmony. AI can generate it now.

The field of AI that deals with understanding and generating language.

The closest point in a dataset to the one you're asking about.

A benchmark where a small fact is hidden in a massive context to test long-context retrieval.

A model made of layers of tiny math units, loosely inspired by brain cells, that learns patterns from data.

The core trick of language models: always guess what token comes next.

The US NIST AI Risk Management Framework — a voluntary guide for managing AI risks.

An organization authorized to certify whether a product meets EU conformity requirements.

The GPU maker whose chips (H100, B200) power most AI training.

Optical character recognition — reading text out of an image.

A one-command tool for running open-weights LLMs locally.

Giving the model exactly one example of the task.

Software whose source code anyone can read, use, and modify — often under a free license.

A model whose weights you can download and run yourself.

The company behind ChatGPT, GPT-5, DALL-E, Whisper, and Sora.

An open-source agentic AI stack popular in the Tendril Creators tier projects.

A personal view — what you think, not a fact.

First of its kind — not copied from someone else.

Inputs that differ from the training data — where models are most likely to fail unexpectedly.

Whatever the AI spits back — text, image, code, audio.

When a model memorizes the training data but fails on new examples.

Managing KV cache memory in fixed-size pages like virtual memory — much more efficient.

An image with color and texture — AI can paint in almost any style.

Decoding multiple tokens at once, not strictly one-at-a-time.

When a model emits multiple independent tool calls in a single turn so the runtime can run them concurrently.

One learnable number in the model — modern models have billions.

Fine-tuning by training a tiny fraction of parameters, saving compute and memory.

The probability that at least one of k sampled attempts solves the task.

A secret word or phrase that proves you're you when you log in.

An app that remembers your passwords for you and helps you make strong ones.

A repeating shape, sound, or idea the AI learns to spot in data.

A score that measures how surprised a language model is by text — lower is better.

A search-first AI product that cites its sources as it answers.

Details that identify you — full name, address, birthday, phone number, school.

A picture taken by a camera — or, these days, made by AI to look like one.

A user-friendly text-to-video startup with a playful product and strong creator community.

Splitting model layers across GPUs so different stages run in a pipeline.

The smallest dot of color in a digital image.

An agent pattern where the model first writes a multi-step plan, then executes each step with tools.

When one neuron fires for many unrelated concepts — hard to interpret.

Applying normalization after the sublayer output — the original transformer design.

Applying normalization before the sublayer (attention, FFN) — makes training big models stable.

The big, expensive training stage where a model learns from huge amounts of raw data.

Of the things the model said were positive, how many actually were.

The AI's best guess about something — the next word, a label, or a number.

Pairs of responses where a human (or AI) says which is better — fuel for RLHF and DPO.

The phase where the model processes your whole prompt before generating the first output token.

Reusing the computed KV cache for shared prompt prefixes across many requests.

OpenAI's version of tiered safety commitments scaling with capability.

Keeping your personal info safe and not sharing it with people (or AIs) who don't need it.

How likely something is, from 0% (no way) to 100% (definitely).

A small classifier trained on a model's internal activations to see what it knows.

A set of instructions a computer runs to do something.

What you type to an AI to tell it what you want.

Provider feature that caches repeated prompt content for much cheaper follow-up calls.

The craft of writing prompts that get good answers out of AI.

An attack where someone sneaks instructions into input data that the model then follows.

A curated set of reusable, tested prompts for common tasks.

A reusable prompt with slots for variables, like a form letter.

Owned and controlled by a company — not freely shared.

Trustable information about where a piece of content came from and how it was made.

A company that offers AI models through an API — like Anthropic, OpenAI, or Google.

Removing unimportant weights from a model to make it smaller.

LoRA combined with 4-bit quantization — fine-tune a 65B model on a single consumer GPU.

Shrinking a model by storing its weights in fewer bits.

The total amount you're allowed to use over a longer window — like per month.

A model that averages lots of decision trees for better, more stable predictions.

A cap on how many requests or tokens per minute a user or app can send.

A prompting pattern that interleaves reasoning and action — think, act, observe, repeat.

Actually happened, actually exists — the opposite of fake.

A model trained to think step-by-step before answering — used for hard math, code, and planning.

Internal thinking tokens a reasoning model generates but usually hides in the final response.

A visible record of the steps a reasoning model took before answering.

Of all the real positives out there, how many the model caught.

When an AI identifies something — a face, a voice, a handwritten number.

A network that processes sequences one step at a time, carrying memory forward.

Formal testing where experts try to break a model — measuring actual safety, not just training intent.

People whose job is to attack a system to find weaknesses before real attackers do.

An agent pattern where the model looks back at its recent actions and decides how to improve.

An agent technique where the model critiques its own output, writes a lesson, and retries.

When the AI politely declines to answer a request because of a safety rule.

A list of topics or requests a model is trained to refuse.

Predicting a number rather than a category — like house price or temperature.

An automated check that confirms a prompt or model change didn't break previously-working examples.

Techniques that keep a model from getting too attached to its training data.

A supervised testing space where companies can trial AI with relaxed rules and regulator feedback.

Training by trial and error, where the AI learns from rewards for good actions.

Generating many outputs, keeping the good ones, and fine-tuning on them.

A super-simple activation: output the input if it's positive, otherwise zero.

Another word for the AI's response in chat.

Cross-encoder model that re-scores top-k retrieved results for higher precision in RAG.

A second step after retrieval that reorders candidates by relevance using a better model.

The AI's reply to you.

An umbrella term for building AI that's fair, transparent, safe, and accountable.

Anthropic's framework tying AI capability thresholds to required safety commitments.

A standard style for web APIs using simple HTTP requests.

Making a chatbot look stuff up before answering, so it stays accurate and current.

Finding cheats that boost reward without doing the actual task.

A model trained on human preferences that scores how 'good' an output is — the heart of RLHF.

Reinforcement learning from AI feedback — using an AI rater instead of humans.

Reinforcement learning from human feedback — how chatbots are taught to be helpful and polite.

A lightweight normalization layer used in modern LLMs instead of LayerNorm.

A way to encode token positions in attention that extrapolates to longer contexts than seen in training.

A score for summarization quality, based on overlap between generated and reference summaries.

An if-then instruction, written by a human, that a program follows.

A benchmark for long-context LLMs that goes beyond simple needle-in-a-haystack.

A creative AI company whose Gen-series video models power filmmakers and designers.

Preventing AI from causing harm — to users, bystanders, or society.

A company's rules about what their AI will and won't do.

How the model picks the next token from its probability list.

When a model intentionally performs worse than it can to avoid detection.

Running model-generated code or shell commands in an isolated environment so they can't damage the host.

Extra structure around a model — tools, memory, retries — that turns it into an agent.

Ways to supervise AI that's smarter than you — using AIs to help or using clever procedures.

Making a model bigger — more parameters, more data, more compute — to get it smarter.

Serving large models to many users cheaply and fast.

The power in the equation that predicts how fast loss drops as you scale up.

Equations that predict how much smarter a model gets when you scale up data, compute, or parameters.

Techniques that let a model handle longer context without breaking accuracy or cost.

A decoding technique that forces the model to only emit tokens that conform to a given schema (e.g. JSON Schema).

When a model deliberately deceives to achieve its goals — a worrying advanced failure mode.

A hidden region of the model's output where it works through a problem before answering.

A software development kit — a library in a specific language that wraps an API.

Something you don't want other people (or AIs, or the internet) to see.

US law that protects online platforms from liability for user-posted content — its role in AI is debated.

A US law authorizing warrantless foreign surveillance — relevant to where AI providers can operate.

Regulating AI separately in each industry (health, finance, transport) rather than with a single law.

A general term for a TEE — an isolated chunk of hardware that keeps secrets safe.

Several parties compute together on private inputs, learning only the result.

When the model attends to different positions within the same sequence.

Asking the model to review and improve its own output before returning it.

Learning from unlabeled data by creating labels out of the data itself.

Finding stuff by meaning instead of exact keyword match.

A gadget that lets a computer sense the world — like a camera, microphone, or thermometer.

Google's tokenizer that works directly on raw text without language-specific preprocessing.

One ongoing conversation with the AI.

Supervised fine-tuning — training a model on labeled examples of good answers.

What the AI remembers during the current conversation — basically the context window.

An S-shaped activation that squashes any input into a 0-to-1 range.

A pretend version of something real — a model of a city, a game, or a conversation.

Registering package names that LLMs hallucinate, so unsuspecting copy-paste users install your malware.

In tech, 'smart' usually means 'has some AI or internet smarts inside'.

An activation that turns a list of numbers into probabilities that sum to 1.

A piece of music, sometimes with words. AI can now compose full songs.

OpenAI's flagship text-to-video model.

Where a piece of info comes from — like a website, a book, or a study.

A tool that breaks a model's activations into many human-interpretable features.

When most values in a model or activation are zero — lets you skip computation.

When an AI finds a loophole that technically satisfies its reward but not what we really wanted.

Making a fast small model draft several tokens, then having the big model verify them in parallel.

Spoken words. AI can turn speech into text, and turn text into speech.

Converting spoken audio into written text.

An open-weights image-generation model that sparked the open-source AI art boom.

A published technical specification that helps operationalize regulation.

A direction in activation space that, when added, shifts the model's behavior predictably.

A narrative — a tale with characters and events. AI is great at drafting them.

Getting the AI's answer word-by-word as it's generated, instead of waiting for the whole thing.

When the AI replies in a strict format like JSON that your code can read directly.

An AI music generator that makes full songs from text prompts.

Research aimed at aligning AI systems much smarter than humans.

Packing more features into a representation than there are dimensions, by using sparse combinations.

Training on data where each example has a correct answer attached.

Keeping AI's growth compatible with long-term environmental and social health.

A benchmark of real GitHub issues to test how well an AI can fix bugs in real codebases.

When a model agrees with the user even when they're wrong, to please them.

Training data generated by another AI instead of collected from humans.

Any content created or heavily altered by AI — images, audio, video, text.

Google DeepMind's watermarking system for AI-generated images, audio, and text.

A detailed public document describing a deployed AI system — its risks, limits, and safeguards.

Instructions set by the developer that tell the AI how to behave in a chat.

Google's text-to-text transformer — every task is framed as input text to output text.

Trusted execution environment — a hardware enclave that isolates code and data from the OS.

A knob that controls how random the model's output is.

Splitting individual matrix multiplies across several GPUs so big layers fit.

NVIDIA's optimized LLM inference library using TensorRT for maximum GPU performance.

Held-out data used only at the end to measure the final model's real performance.

Spending more compute at inference to get better answers — the trend behind reasoning models.

Written words — the main thing language AIs read and write.

Turning a text prompt into a picture.

Converting written text into spoken audio.

Turning a text prompt into a short video.

Text Generation Inference — Hugging Face's production LLM serving stack.

Hidden reasoning tokens a model generates before producing its final visible answer.

Tokens spent inside a model's reasoning channel — billed separately from visible output tokens.

How many tokens per second the system can produce.

Regulating AI differently based on risk level or capability — the EU AI Act is the prime example.

How long between sending a prompt and getting the first token back — a big UX factor.

Breaking text into tokens so the AI can read it.

The piece of software that splits text into tokens — and joins them back.

Small chunks of text (like pieces of words) that AI reads and writes in.

A single invocation where the model asks to run a specific tool with specific arguments.

A prompt-injection attack hidden inside a tool's description or output that hijacks the agent.

The output a tool returns back to the model after being called.

Letting the AI call external functions — like a calculator, search, or code runner.

A sampling setting that only lets the model pick from its top k most likely next tokens.

A sampling setting that picks from the smallest group of tokens whose probability adds up to p.

Google's custom AI chip, used for training and serving Gemini and other models.

The process of teaching an AI by showing it examples until it gets good at a task.

The money and compute it takes to train a model from scratch.

The specific pile of examples used to teach an AI.

Techniques for figuring out which training examples caused a model to produce a given output.

Deliberately polluting a model's training set to plant backdoors or degrade behavior.

The slice of your data used to actually train the model.

The neural network architecture behind almost all modern AI, based on attention.

Hugging Face's open-source library that makes using and fine-tuning LLMs straightforward.

Exploring multiple reasoning paths in a tree, pruning weak ones, to solve harder problems.

Another name for a backdoored AI model — appears helpful, secretly malicious.

Believing something or someone is reliable — and knowing when not to.

An AI teacher that walks you through learning something at your own pace.

An AI music-generation app that writes full songs from short prompts.

When a model is too simple to capture the patterns in the data.

Training on data with no labels — the model finds structure on its own.

The message the user types in — the visible input side of a conversation.

Vercel's generative UI tool that turns prompts into React/shadcn components and full apps.

Data used during training to tune settings without cheating on the test set.

A list of numbers — the form embeddings take so math can compare them.

A database built to store embeddings and find the most similar ones quickly.

Google DeepMind's text-to-video model.

A unified API for routing calls across AI providers with failover, caching, and cost tracking.

Double-check that something is true.

Moving pictures with sound. AI can now make short videos from text.

An AI helper that does stuff for you — like setting timers, sending messages, or answering questions.

A model that can 'see' — take images or video as input.

A popular open-source LLM serving library known for PagedAttention and continuous batching.

Vision-language model — an LLM that can also handle images as input.

The audio version of language — what you say out loud.

Making a synthetic voice that sounds like a specific person, usually from a short sample.

A hidden or visible mark that flags content as AI-generated.

Embedding an invisible signal in AI output so you can later prove it came from that AI.

Using weak supervisors (humans, smaller models) to teach stronger ones effectively.

A benchmark of realistic web-browsing tasks for AI agents.

A two-way connection that stays open, good for live streaming and voice.

All the numbers inside a trained model — the thing that actually gets updated during training.

OpenAI's open-weights speech-to-text model that handles many languages.

The percentage of head-to-head comparisons where one model's output is preferred over another's.

Codeium's agentic IDE that plans and executes multi-file code changes.

Elon Musk's AI company, behind the Grok model family.

A method to extend RoPE's effective context window beyond training length.

DeepSpeed's Zero Redundancy Optimizer — shards optimizer state, grads, and params across GPUs.

Asking the model to do a task without any examples — just an instruction.