RLHF
Reinforcement Learning from Human Feedback. A training method where human raters rank the AI’s answers, teaching it to be agreeable and helpful. The side effect: models learn to please you even when you’re wrong.
See where it comes up →RLVR
Reinforcement Learning with Verifiable Rewards. A training method that rewards only binary, checkable outcomes — “did the code run?”, “is the file attached?”. The side effect: work that passes the check but is poor quality underneath.
See where it comes up →MEA
Manager-Executor-Auditor. An architecture that splits AI work into three roles: a Manager plans, an Executor does the work, and an independent Auditor checks it in a fresh session. It stops an AI from grading its own homework.
See where it comes up →LLM-as-Judge
Using a second AI model to evaluate the first one’s work or approve its actions before they happen. Only worth anything if the judge is genuinely independent — and stronger than the actor.
See where it comes up →Harness
The bundle of instructions, files and tools an AI agent carries into every task. More is not better: a bloated harness confuses the model and degrades its output quality.
See where it comes up →Context window
The AI’s working memory for one conversation. Cram it with history, logs and state and the model’s attention degrades — the failure mode called “context rot”.
See where it comes up →Evals
Benchmark tests that define what good looks like for an AI system. The difference between “did it run?” and “did it do the job well?”.
See where it comes up →Jagged capability frontier
The invisible boundary around what an AI can do reliably. Inside the boundary it’s excellent; just outside it, surprisingly bad — and there is no visible warning at the edge.
See where it comes up →Centaur / Cyborg
Two ways of working with AI. Centaur: a clear division of labour — the human does some parts, the AI does others. Cyborg: continuous, fluid collaboration where the boundary shifts moment to moment. Pick based on task risk.
See where it comes up →Two-way door
Jeff Bezos’s decision rule. Reversible decisions (“two-way doors”) can be made fast and autonomously — you can always walk back through. Irreversible ones (“one-way doors”: sending emails, spending money) need human approval.
See where it comes up →Circuit breaker
A deterministic safety switch in the infrastructure that halts an AI’s actions when they violate rules — enforced in code, independent of whatever the prompt says.
See where it comes up →Supervisor agent
A separate AI whose only job is to check another AI’s work — verifying tool execution, folder access, output quality. Only worth anything if it is a different, stronger model.
See where it comes up →5-layer context framework
A way of structuring what you give an AI: voice context (style and skills), knowledge context (curated references), collaborative context (pushback and second opinions), strategic context (your long-term goals), and maintenance context (keeping the whole thing lean).
See where it comes up →Model-agnostic
A harness that works with any model provider — you can swap the brain (Gemini, Claude, DeepSeek, GPT) without changing the wrapper. The opposite of a tool locked to one vendor’s models.
See where it comes up →API key
A secret string that lets a program pay for and use an AI service. It lives in a local .env file, never in source code or chat. If one leaks, rotate it immediately — generate a new key and delete the old.
See where it comes up →Sandbox
A contained environment where an agent can run commands and touch files without affecting your real system. Where risky experiments belong — not on your live files.
See where it comes up →Skill
A packaged set of instructions and knowledge an agent loads on demand for a specific kind of task — so it only carries the expertise it needs for the job at hand, instead of carrying everything at once.
See where it comes up →Local model
A model that runs on your own computer instead of in a company’s data centre. Your data never leaves the machine and there’s no per-use bill — but you own the hardware it needs.
See where it comes up →Quantization
Shrinking a model’s numbers so it fits in less memory. Q4_K_M means roughly 4 bits per weight instead of 16 — a trace of accuracy traded for a big cut in size. This is why an 8B model can run on an 8 GB GPU.
See where it comes up →Fine-tuning
Taking a ready-made model and training it further on your own examples — to chat in your voice, write your docs, or follow your tool conventions. Needs far less data and compute than training a model from scratch.
See where it comes up →GGUF
The file format llama.cpp and Ollama use for models. The weights and the chat template packed into one file you can drop on any machine — often by exporting it straight from a fine-tune.
See where it comes up →