# farez.ai

> Free, visual lessons on how AI and AI coding agents work, by Farez Vadsaria: a map of 549 AI concepts with prerequisite routes, 136 research papers in plain words, 19 Claude Code concepts verified against the official docs, and a 5-lesson course on coding-agent harnesses.

Every page on this site has a Markdown version. Request any URL with `Accept: text/markdown`, or add `.md` to the path (the homepage is https://www.farez.ai/index.md). The links below already point at the Markdown versions. Everything is public and read-only: no accounts or API keys. Unknown paths return a real 404.

Facts are written from primary sources (official documentation, papers, textbooks) and cite them; prefer the cited source when it matters. Claude Code facts are checked against code.claude.com and carry a review date.

## When to use this site

- [Explain an AI or machine-learning concept from the ground up](https://www.farez.ai/guide.md): use when someone asks what a term means (attention, RLHF, embeddings, overfitting…) or what to learn before it. Fetch /guide/<slug>.md: each term lists its prerequisites in reading order, related terms, and sources.
- [Explain a Claude Code feature](https://www.farez.ai/claude-guide.md): use when someone asks how hooks, subagents, skills, MCP, permissions, memory, worktrees, prompt caching or other Claude Code features work and how they relate. Each concept links its official-docs sources.
- [Summarize an influential AI paper](https://www.farez.ai/papers.md): use to find a plain-language summary of a well-known AI paper, why it matters, and which concepts to learn first.
- [Show how coding agents work inside](https://www.farez.ai/harnesses.md): use when someone wants to understand the agent loop, context, and tools of real open-source coding agents.
- [Explain why context drives AI agent cost](https://www.farez.ai/play.md): a worked example with real prices: one model alone costs $31 and fails; an orchestrator with subagents ships the same 12 jobs for $3.07.
- [Not a fit](https://www.farez.ai/developers.md): this site has no API, product or service to call; it is a reading resource. For anything else about Farez, see About and Contact.

## Start here

- [Learn](https://www.farez.ai/learn.md): every resource on the site, grouped by goal
- [Field Guide](https://www.farez.ai/guide.md): all 549 terms by region
- [Claude Guide](https://www.farez.ai/claude-guide.md): all 19 concepts
- [Papers](https://www.farez.ai/papers.md): 36 canon papers and the 100 most-cited of Aug 2025 – Aug 2026
- [Agentic Coding Harnesses](https://www.farez.ai/harnesses.md): 5 lessons
- [The Duel](https://www.farez.ai/play.md): context and token cost, as a game

## Claude Guide concepts

- [Checkpoints](https://www.farez.ai/claude-guide/checkpoints.md): Claude Code snapshots your code before every prompt, so you can rewind Claude's file edits, the conversation, or both with /rewind or a double Esc.
- [Claude Code in the cloud](https://www.farez.ai/claude-guide/claude-web.md): Claude Code sessions that run in an isolated cloud VM instead of on your machine. Start one from claude.ai/code, your phone or the terminal, then review the branch it pushes.
- [Context](https://www.farez.ai/claude-guide/context.md): The context window is everything Claude can see in a session: your messages, the files it reads, tool output and auto-loaded instructions. It fills as you work, and Claude Code compacts it when it gets full.
- [Slash Commands](https://www.farez.ai/claude-guide/slash-commands.md): Type / to run a command: built-ins that control the Claude Code session itself, plus prompts packaged as skills, plugin commands and MCP prompts, including ones you write yourself.
- [Goals (/goal)](https://www.farez.ai/claude-guide/goal.md): Give /goal a completion condition and Claude keeps taking turns toward it, while a separate small model checks after every turn whether the condition is met.
- [Hooks](https://www.farez.ai/claude-guide/hooks.md): Your own commands that Claude Code runs automatically at fixed points in its loop, so a rule always holds instead of depending on the model remembering it.
- [MCP](https://www.farez.ai/claude-guide/mcp.md): The Model Context Protocol connects Claude Code to outside tools and data, such as an issue tracker, a database or Figma, through servers that hand Claude new tools to call.
- [Memory](https://www.farez.ai/claude-guide/memory.md): CLAUDE.md files you write and auto memory notes Claude writes for itself: plain markdown on disk, loaded at the start of every session so Claude doesn't start from zero.
- [Models & Effort](https://www.farez.ai/claude-guide/model-and-effort.md): The two dials behind every answer: which Claude model does the work, and how much effort it spends thinking and checking before it replies, trading speed and tokens for depth.
- [Permissions](https://www.farez.ai/claude-guide/permissions.md): How Claude Code decides whether a tool call runs, asks you first, or is blocked: a permission mode sets the baseline, and allow, ask and deny rules carve out exceptions.
- [Prompt Caching](https://www.farez.ai/claude-guide/prompt-caching.md): Claude Code re-sends your whole conversation on every turn. Prompt caching lets the API reuse the unchanged start of it, so long sessions stay fast and cheap until something breaks that prefix.
- [Run it programmatically](https://www.farez.ai/claude-guide/agent-sdk.md): Run Claude Code with nobody at the keyboard: claude -p for scripts and CI, or the Agent SDK to drive the same agent loop, tools and context handling from your own Python or TypeScript code.
- [Subagents](https://www.farez.ai/claude-guide/subagents.md): Helpers Claude hands a side task to. Each works in its own fresh context window with its own prompt and tools, and only its summary comes back.
- [Worktrees](https://www.farez.ai/claude-guide/worktrees.md): A separate git checkout for each Claude Code session or subagent, with its own files and branch, so parallel work never edits the same files.
- [Dynamic Workflows](https://www.farez.ai/claude-guide/workflows.md): A JavaScript script Claude writes to orchestrate dozens of subagents at once and cross-check their results, run in the background and saved as a command you can rerun.
- [Sandboxing](https://www.farez.ai/claude-guide/sandboxing.md): An operating-system boundary around the shell commands Claude runs: they can write only where you allow and reach only the domains you approve, so you can approve fewer commands by hand.
- [Skills](https://www.farez.ai/claude-guide/skills.md): A SKILL.md file of instructions that Claude loads only when a task calls for it, or that you run yourself as a /command. Custom slash commands are now skills.
- [Plugins](https://www.farez.ai/claude-guide/plugins.md): A directory of skills, subagents, hooks and MCP servers that Claude Code installs and loads as one unit, usually from a marketplace, so a whole setup can be shared with one command.
- [Mods](https://www.farez.ai/claude-guide/mods.md): Plugins whose JavaScript or TypeScript functions run inside Claude Code on its events, so they can watch, rewrite or answer a tool call or prompt, add commands, and draw their own interface.

## Field Guide landmarks

- [Accuracy](https://www.farez.ai/guide/accuracy.md): The share of predictions a classifier gets right. Simple and intuitive, but misleading when one class is much rarer than the others.
- [Activation Function](https://www.farez.ai/guide/activation-function.md): A non-linear function applied to neuron outputs that introduces non-linearity, enabling networks to learn complex patterns.
- [Agent](https://www.farez.ai/guide/agent.md): In reinforcement learning, the learner and decision-maker that observes an environment, takes actions and adjusts its behaviour to maximize cumulative reward.
- [AI Agent](https://www.farez.ai/guide/ai-agent.md): A system where a large language model decides its own next steps in a loop: calling tools, reading the results, and continuing until the task is done.
- [Attention Mechanism](https://www.farez.ai/guide/attention-mechanism.md): A technique that lets a neural network weigh every part of its input when producing each output, focusing on the parts most relevant at that step.
- [Backpropagation](https://www.farez.ai/guide/backpropagation.md): The algorithm for computing gradients of the loss with respect to network weights, enabling training through gradient descent.
- [Classification](https://www.farez.ai/guide/classification.md): A supervised learning task where the model assigns each input to one of a fixed set of categories, such as spam or not spam.
- [Computer Vision](https://www.farez.ai/guide/computer-vision.md): The field of AI that gets computers to extract meaning from images and video: what is in them, where it is, and how it moves.
- [Context Window](https://www.farez.ai/guide/context-window.md): The maximum number of tokens a language model can take into account at once, counting both the input it reads and the output it writes. Also called context length.
- [Convolutional Neural Network](https://www.farez.ai/guide/cnn.md): A neural network that scans images with small learned filters, reusing the same weights at every position to build up from edges to whole objects.
- [Deep Learning](https://www.farez.ai/guide/deep-learning.md): A subset of machine learning that uses neural networks with multiple layers (deep neural networks) to learn hierarchical representations of data.
- [Diffusion Model](https://www.farez.ai/guide/diffusion-model.md): A generative model trained to remove noise a step at a time, so it can turn pure random noise into a new image, audio clip or other sample.
- [Dropout](https://www.farez.ai/guide/dropout.md): A regularization technique that randomly switches off units during training so the network can't lean on any one of them, which reduces overfitting.
- [Embedding](https://www.farez.ai/guide/embedding.md): A list of numbers (a vector) that represents a word, sentence, image or other item, learned so that similar items end up close together.
- [False Positive](https://www.farez.ai/guide/false-positive.md): A case the model labels positive that is actually negative, such as a legitimate email sent to spam. Statisticians call it a Type I error.
- [Feature](https://www.farez.ai/guide/feature.md): A single measurable property of an example, such as a house's floor area or how many links an email contains, used as an input to a model.
- [Fine-Tuning](https://www.farez.ai/guide/fine-tuning.md): The process of further training a pre-trained model on a specific dataset to adapt it for a particular task or domain.
- [Foundation Model](https://www.farez.ai/guide/foundation-model.md): A large model pre-trained on broad data at scale that can be adapted to many downstream tasks, such as GPT-style LLMs, BERT or CLIP.
- [Generalization](https://www.farez.ai/guide/generalization.md): A model's ability to perform well on new data it never saw during training, which is the whole point of learning from examples.
- [Gradient Descent](https://www.farez.ai/guide/gradient-descent.md): An optimization method that repeatedly moves a model's parameters a small step in the direction that most reduces the loss.
- [Hallucination](https://www.farez.ai/guide/hallucination.md): When a language model produces fluent, confident output that is false or unsupported by its sources, such as invented facts, quotes or citations.
- [Hyperparameter](https://www.farez.ai/guide/hyperparameter.md): A setting chosen before training, such as the learning rate or batch size, that controls how a model learns but isn't itself learned from the data.
- [Image Classification](https://www.farez.ai/guide/image-classification.md): Assigning one label from a fixed set of categories to a whole image, such as "cat" or "defective part"; a core computer vision task.
- [Image Generation](https://www.farez.ai/guide/image-generation.md): Creating new images, often from a text prompt, with generative models such as GANs, VAEs and diffusion models that learn to turn random noise into realistic pictures.
- [Inference](https://www.farez.ai/guide/inference.md): Running a trained model on new inputs to get predictions, with its weights frozen: the stage of a model's life that users actually interact with.
- [Intersection over Union](https://www.farez.ai/guide/iou.md): A score from 0 to 1 for how well a predicted region matches the true one: the area they share divided by the area they cover together.
- [Language Modeling](https://www.farez.ai/guide/language-modeling.md): Assigning probabilities to sequences of text, in practice by predicting each next token from the ones before it; the objective behind large language models.
- [Large Language Model](https://www.farez.ai/guide/llm.md): A neural network, almost always a transformer, trained on vast amounts of text to predict the next token, which lets it write, answer, summarize and follow instructions.
- [Learning Rate](https://www.farez.ai/guide/learning-rate.md): The hyperparameter that sets how big a step gradient descent takes on each update; too high makes training unstable, too low makes it crawl.
- [Loss Function](https://www.farez.ai/guide/loss-function.md): A function that scores how wrong a model's prediction is as a single number, which training then works to make as small as possible.
- [Machine Learning](https://www.farez.ai/guide/machine-learning.md): Building systems that learn patterns from data instead of following hand-written rules, getting better at a task as they see more examples.
- [Natural Language Processing](https://www.farez.ai/guide/nlp.md): The field of AI that lets computers read, interpret, translate and generate human language, from spam filters and search to chatbots.
- [Neural Network](https://www.farez.ai/guide/neural-network.md): A model built from layers of simple units, each taking a weighted sum of its inputs and applying a nonlinearity; training adjusts the weights so the whole stack maps inputs to useful outputs.
- [Object Detection](https://www.farez.ai/guide/object-detection.md): Finding every object of interest in an image and giving each a class label, a confidence score and a bounding box.
- [Overfitting](https://www.farez.ai/guide/overfitting.md): When a model fits its training data too closely, noise included, so it scores well on examples it has seen and poorly on new ones.
- [Parameter](https://www.farez.ai/guide/parameter.md): A number inside a model, a weight or a bias, that training adjusts to reduce the loss. Together, a model's parameters are everything it has learned.
- [Policy](https://www.farez.ai/guide/policy.md): The rule an RL agent uses to choose actions: a mapping from each state to an action, or to a probability distribution over actions.
- [Pre-training](https://www.farez.ai/guide/pre-training.md): Training a model on a large dataset (often self-supervised) before fine-tuning on specific tasks, enabling transfer learning.
- [Precision](https://www.farez.ai/guide/precision.md): Of everything a model flagged as positive, the fraction that really was positive: TP / (TP + FP). It answers "when the model says yes, how often is it right?"
- [Prompt Engineering](https://www.farez.ai/guide/prompt-engineering.md): Designing a model's input (instructions, context, examples and output format) to get better results without changing the model itself.
- [Recall](https://www.farez.ai/guide/recall.md): The share of actual positives a model correctly identifies, TP / (TP + FN); also called sensitivity or the true positive rate.
- [Recurrent Neural Network](https://www.farez.ai/guide/rnn.md): A neural network architecture with loops that allow information to persist, designed for sequential data like text and time series.
- [Regression](https://www.farez.ai/guide/regression.md): A supervised learning task where the model predicts continuous numerical values rather than discrete categories.
- [Regularization](https://www.farez.ai/guide/regularization.md): Any change to training meant to make a model generalize better rather than fit its training data more closely, such as weight penalties, dropout or early stopping.
- [Reinforcement Learning](https://www.farez.ai/guide/reinforcement-learning.md): Learning through interaction with an environment, receiving rewards or penalties to learn optimal behavior policies.
- [Representation Learning](https://www.farez.ai/guide/representation-learning.md): Letting a model learn its own internal description of the data, such as embeddings or layered visual features, instead of relying on features designed by hand.
- [Retrieval-Augmented Generation](https://www.farez.ai/guide/rag.md): Fetching relevant documents at question time and adding them to an LLM's prompt, so answers can draw on current, checkable sources without retraining.
- [Reward](https://www.farez.ai/guide/reward.md): The single number an environment sends back after each action, telling a reinforcement learning agent how good that step was; the agent learns to maximize its long-run total.
- [Self-Attention](https://www.farez.ai/guide/self-attention.md): An attention step in which every token in a sequence looks at every token in that same sequence and builds a new representation from a weighted mix of them.
- [Self-Supervised Learning](https://www.farez.ai/guide/self-supervised-learning.md): Training on unlabeled data by making the data supply its own answers: hide or alter part of an input and have the model predict or match it.
- [Semantic Segmentation](https://www.farez.ai/guide/semantic-segmentation.md): Labelling every pixel in an image with a class, such as road, car or sky, so the output is a map of what is where rather than a single tag.
- [Softmax](https://www.farez.ai/guide/softmax.md): A function that turns a list of scores (logits) into probabilities that are all positive and sum to 1; the standard output of classifiers and language models.
- [Supervised Learning](https://www.farez.ai/guide/supervised-learning.md): Learning from examples paired with the correct answer, so a model can predict answers for new inputs it hasn't seen.
- [Test Set](https://www.farez.ai/guide/test-set.md): Data locked away until the end of a project and used once to estimate how the finished model will perform on new, unseen examples.
- [Token](https://www.farez.ai/guide/token.md): The unit of text a language model reads and writes: a word, part of a word, or a character, drawn from the model's fixed vocabulary.
- [Tokenization](https://www.farez.ai/guide/tokenization.md): Splitting text into tokens, usually subword pieces, and mapping each to an integer ID so a language model can process it.
- [Tool Use](https://www.farez.ai/guide/tool-use.md): A language model asking the application around it to run a function, such as search, a calculator or an API, then using the result in its answer.
- [Training](https://www.farez.ai/guide/training.md): The process of fitting a model to data by repeatedly measuring how wrong its outputs are and adjusting its parameters to reduce that error.
- [Training Data](https://www.farez.ai/guide/training-data.md): The examples a model learns its weights from, kept separate from the validation and test data used to check how well it generalizes.
- [Transformer](https://www.farez.ai/guide/transformer.md): A neural network architecture, introduced in 2017, built from stacked self-attention and feed-forward layers; the basis of nearly every modern large language model.
- [Unsupervised Learning](https://www.farez.ai/guide/unsupervised-learning.md): Learning from unlabeled data to discover hidden patterns, structures, or relationships without explicit target outputs.
- [Validation Set](https://www.farez.ai/guide/validation-set.md): Data held back from training and used to choose hyperparameters and checkpoints, so those choices are judged on examples the model never fit.

## Harness course

- [Agentic Coding Basics: What a coding-agent harness actually is](https://www.farez.ai/harnesses/agentic-coding-basics.md)
- [DeepSeek Harness: Everything is a plugin — even the agent loop](https://www.farez.ai/harnesses/deepseek-harness.md)
- [opencode: The client/server harness](https://www.farez.ai/harnesses/opencode.md)
- [Kimi Code CLI: Subagents in disposable contexts](https://www.farez.ai/harnesses/kimi-code.md)
- [Pi: Four tools and a tiny prompt](https://www.farez.ai/harnesses/pi.md)

## About this site

- [About](https://www.farez.ai/about.md): who runs farez.ai and why
- [Contact](https://www.farez.ai/contact.md): email for questions and corrections
- [Privacy](https://www.farez.ai/privacy.md): what the site collects
- [For agents and developers](https://www.farez.ai/developers.md): every machine-readable resource, with examples

## Optional

- [Blog](https://www.farez.ai/blog.md): 2 posts
- [Projects](https://www.farez.ai/projects.md): free multilingual AI apps
- [Sitemap](https://www.farez.ai/sitemap-index.xml): every page
- [Field Guide JSON index](https://www.farez.ai/guide/search.json): every term: slug, name, region, tier, one-line definition
- [RSS](https://www.farez.ai/rss.xml): blog feed