Learn
AI terms, answered directly
Plain definitions of the concepts behind modern AI work, each with the detail that matters and where it shows up in practice.
Concepts
The vocabulary of AI work, defined directly.
What is an AI workspace?The category between a chatbot and an IDE: models, files, agents, and tools working in one persistent environment.What is an AI agent?A model given a goal, tools, and permission to take steps toward it, instead of returning a single answer.What is MCP?The open protocol that lets AI tools and data sources plug into any compatible client, once.What is BYOK?Using your own provider API keys inside another product, so usage bills to your provider account.What is a context window?The amount of text a model can consider at once, measured in tokens, and what happens when you exceed it.What is an LLM gateway?A single API endpoint that routes requests to many model providers, with unified auth, billing, and failover.What is local inference?Running AI models on your own machine: private by construction, free per request, bounded by your hardware.What is agent memory?Persistent knowledge an AI agent keeps across conversations: preferences, facts, and decisions it can recall and update.What is a token in AI?The unit models read, generate, and bill by: roughly three-quarters of an English word.What is a tokenizer?The component that converts text into the token IDs a model processes, and why counts differ per model family.What is AI inference?Running a trained model to produce output: the phase you pay for per token, on a provider, a gateway, or your own hardware.What is LLM quantization?Storing model weights at lower precision so bigger models fit smaller hardware, with a measurable quality trade.What is a mixture of experts model?Models that route each token through a few specialized subnetworks, so total parameters exceed the compute spent per token.What is time to first token (TTFT)?The wait between sending a prompt and the first streamed token: the latency number that shapes perceived speed.What is a system prompt?The standing instructions a model receives before the conversation: role, rules, tone, and tool guidance.What is RAG (retrieval-augmented generation)?Fetching relevant documents at request time and placing them in context, so answers cite your data instead of guessing.What is prompt caching?Providers reuse the processed prefix of your prompt across requests and bill cached tokens at a discount.What is tool calling?Models emit structured calls to functions you define, then act on the results: the substrate of every agent loop.What is reasoning effort?The dial on reasoning models that trades think-time and cost against answer quality, set per request.What is multimodal AI?Models that accept or produce more than text: what input and output modality actually mean when picking a model.What are open-weights models?Downloadable model weights are not the same as open source: what each kind of license actually grants.What are AI agent autonomy levels?The graduated scale of what an agent may do without you: from point-only reading to unattended execution.
Benchmarks
What each benchmark measures, how it is scored, and how to read the numbers.
What is GPQA?198 PhD-level science questions built to resist search: the standard test of expert reasoning in AI models.What is Humanity's Last Exam (HLE)?2,500 questions at the edge of human expertise, designed so frontier models cannot saturate it: low scores are the point.What is MMLU-Pro?12,000 questions, 10 answer choices, 14 domains: the broad-knowledge benchmark rebuilt to resist guessing and contamination.What is SimpleQA?Short factual questions with one verifiable answer: the most direct public measure of whether a model hallucinates.What is LiveCodeBench?A rolling set of 1000+ coding problems dated after model training cutoffs, so scores measure skill rather than memory.What is SciCode?Coding problems drawn from real research workflows in physics, math, and biology, decomposed into verifiable subproblems.What is SWE-bench Verified?Models resolve real GitHub issues in real repositories, graded by the project's own tests: the agentic coding standard.What is the Aider Polyglot benchmark?225 Exercism exercises across 6 languages, passed only when edits apply cleanly and the tests pass: practical editing skill.What is MATH-500?500 competition problems spanning graded difficulty levels, checked by final answer: the standard mid-difficulty math test.What is the AIME benchmark?A 30-problem invitational math exam used to test AI reasoning, where every single problem moves the score by over 3 points.What is FrontierMath?Unpublished, expert-written research mathematics problems kept private to prevent contamination; frontier models solve few.What is Terminal-Bench?Models complete real tasks in a live terminal (installing, debugging, wiring systems): tool use measured end to end.What is ARC-AGI-2?Abstract grid puzzles humans solve easily and models find hard, scored with efficiency in mind: generalization, not knowledge.What is the METR time horizon?The length of task (in human working time) a model completes at 50% reliability: capability measured on the clock.What is the Epoch Capabilities Index (ECI)?One capability scale fit across 50+ benchmarks with confidence bounds: comparable scores where no single test suffices.