AI & ML

Cursor's version of CLAUDE.md — a file in your project that tells the Cursor AI how to behave.

AI is when you teach a computer to do stuff that normally needs a human brain — like recognizing cats, translating languages, or writing code for you.

An AI Agent is an AI system that can think, plan, and act on its own to accomplish a goal.

Having an AI review your pull request before a human does.

A code editor where the AI isn't a plugin you installed — it's the whole point.

AI Safety is the field of making sure AI doesn't go off the rails.

AI watermarking embeds detectable patterns in AI-generated content to prove its origin — digital fingerprints for generated text or images.

An API key is your password to use an AI service. You include it in every request to prove you're allowed to use the API and so they know who to charge.

AWS Bedrock is like a model buffet — Anthropic's Claude, Meta's Llama, Mistral, Cohere, and more, all accessible through one AWS API. You don't manage any

Accuracy is the simplest way to score a model — what percentage of predictions were correct.

An activation function is the decision gate in a neural network — it decides if a neuron should 'fire' or stay quiet.

An AI agent is an LLM that doesn't just answer questions — it takes actions.

Agent deadlock is when two or more AI agents wait on each other forever. Agent A waits for Agent B's response. Agent B waits for Agent A's. Neither moves.

An agent fleet is a whole bunch of AI agents running in production — each handling one kind of task, all of them at once.

An agent orchestra is a group of AI agents with a conductor — one lead agent that coordinates specialists.

An agent swarm is multiple AI agents working in parallel on different parts of a problem. Like a small dev team, except every member is a model.

An agent team is 2–16 AI agents collaborating via shared task lists — like a sprint team that never sleeps, eats, or complains.

Agent thrash is when an AI agent keeps changing its mind, flipping between approaches without committing. 'Let me try approach A. Actually B. Wait, A.

Agentic describes an AI system that acts autonomously — making decisions, taking actions, using tools, and working toward a goal without constant human han...

Agentic AI is when the AI doesn't just answer questions — it takes action.

An agentic workflow is when you give an AI a goal and let it figure out the steps to get there — breaking down tasks, using tools, making decisions, and it...

Aider is a terminal app where you chat with Claude or GPT-4 about your code and it directly edits your files.

Alignment is the AI safety challenge of making sure AI does what we actually want, not just what we literally said.

Amazon's answer to Copilot — it writes code for you but with an AWS twist. It's especially good at AWS SDK code (obviously) and comes with a built-in secur

Amazon Q is AWS's AI assistant baked into the AWS console and your IDE.

Anthropic is the company that built Claude — think of them as the responsible parent at the AI party.

The API that lets you talk to Claude from your own code. Send a message, get a response — that's the basics.

The full name for AI — making computers smart enough to do human-brain tasks.

This is THE paper. Published in 2017 by Google researchers, 'Attention Is All You Need' introduced the Transformer architecture that powers every modern AI...

The attention mechanism is how AI decides what to focus on — like when you're reading a long email and your eyes jump to the part that mentions your name.

Auto memory is AI that saves and recalls context across sessions without being asked — like it actually paid attention last time.

AutoGPT was the viral 'what if GPT could run itself in a loop?' experiment. Give it a goal, and it breaks it into tasks, executes them, evaluates results,

AI autocomplete is like your phone's predictive text, but for code, and actually intelligent.

An autoencoder is a neural network that learns to compress data and then reconstruct it — like a zip file that learns what to keep and what to toss.

Azure OpenAI is OpenAI's models wearing a Microsoft enterprise suit. Same GPT-4, same capabilities, but running in Azure's data centers with enterprise sec

BabyAGI is the minimalist cousin of AutoGPT — just 140 lines of Python that create tasks, prioritize them, and execute them in a loop. It was a proof of co

Backpropagation is how errors flow backwards through a neural network during training.

A batch is a small group of training examples that the model processes at once before updating its weights.

Batch normalization is like hitting the reset button on each layer of a neural network so the numbers don't spiral out of control.

In AI, a benchmark is a standardized test that measures how good a model is — like the SAT for AI.

In ML, bias means the model has systematic errors — it's consistently wrong in the same direction.

Bolt is like a full-stack developer in a browser tab.

Browser use is AI controlling a real browser — scraping, filling forms, navigating UIs — computer use but scoped to the web.

CLIP connects text and images in one shared understanding — it can look at a photo and know what text describes it, or read text and find matching images.

CUDA is NVIDIA's secret weapon — it's the programming platform that lets developers use NVIDIA GPUs for AI, not just gaming.

Cargo cult AI is adding AI to your product because every competitor is, without a clear reason. A chatbot that nobody uses.

Cargo cult prompting is the same as prompt cargo cult — copying techniques without understanding when they work.

Chain of Thought is when you tell the AI 'show your work' like a math teacher.

Chat Completion is the API pattern for having a back-and-forth conversation with an AI.

ChatGPT is the app that made AI mainstream — it's the iPhone moment for artificial intelligence.

Cutting up big documents into smaller pieces so an AI can actually understand them.

Classification is teaching an AI to sort things into categories. Is this email spam or not? Is this image a cat, dog, or bird?

Claude Code is Anthropic's agentic coding tool that lives in your terminal.

Clustering is teaching an AI to find groups in data WITHOUT being told what the groups are.

Code Generation is AI writing code for you based on a description.

Codex is OpenAI's code-specialized AI model — it's the brain behind GitHub Copilot. Think of it as a GPT that went to coding bootcamp.

Cody is Sourcegraph's AI assistant that uses code search to give the LLM actual relevant context from your giant codebase — not just the files you have ope...

Computer use is AI controlling a desktop — clicking, typing, reading the screen — like an intern who can actually use your Mac.

Computer Vision is teaching AI to understand images and video. How does your phone unlock with your face? Computer Vision.

Confabulation is when an AI invents a plausible-sounding answer and delivers it with total confidence, when the honest answer would be 'I don't know.

Constitutional AI is Anthropic's approach to making AI behave — instead of relying on a giant team of human reviewers, the AI essentially reviews itself us...

Context budget is how you allocate tokens across system prompt, memory, tools, and conversation — every token is a dollar at scale.

Context compaction is summarizing a long AI conversation down to just the important bits so the model can keep going without hitting context limits.

The art of feeding the right information to an AI so it gives better answers.

Context poisoning is when bad information makes it into the AI's context and contaminates everything after.

Context relay is passing AI context forward across sessions, agents, or tools — the plumbing behind persistent, coherent AI workflows.

Context rot is when a long AI conversation has been running so long that the model is confusing old instructions with new ones, forgetting what it already...

Context surgery is manually editing your AI conversation — deleting bad turns, tweaking earlier messages, removing a failed tool call — before continuing.

Continue.dev is an AI coding extension that isn't locked to one provider.

GitHub Copilot is the AI coding assistant that lives inside your editor and autocompletes your code.

Copilot Chat is like having a coding buddy living in your IDE sidebar. You can ask it to explain code, fix bugs, write tests, or refactor functions — all i

Cosine similarity measures how similar two things are by comparing the angles of their vectors.

Cost per token is how much each token (input or output) costs with a given AI provider. Flagship models cost more per token than cheap ones.

CrewAI is a framework for building teams of AI agents that work together — like a virtual office where each agent has a role (researcher, writer, critic) a...

Cursor is a code editor (forked from VS Code) with AI built into every interaction.

Cursor is VS Code's cooler cousin who's really into AI.

DALL-E is OpenAI's AI image generator — describe an image in words and it creates it from scratch. Want 'an avocado armchair'? Done.

DPO is RLHF's more efficient younger sibling.

Data augmentation is making your training data go further by creating variations of what you already have.

Deep Learning is Machine Learning that's been hitting the gym.

DeepSeek is the Chinese AI lab that shocked everyone by building models as good as GPT-4 for a fraction of the cost.

Devin is the AI that claims to be a software engineer. Give it a GitHub issue and it tries to fix it: reads code, writes code, runs tests, opens a PR.

Diffusion models generate images by learning to reverse noise. In training, you take an image and slowly add random noise until it's pure static.

Teaching a small AI by having it copy a big AI's homework.

Elicitation is the MCP pattern where a tool pauses mid-task to ask you clarifying questions — the AI learned to ask before guessing.

An embedding is turning words, sentences, or entire documents into lists of numbers (vectors) that capture their meaning.

An embedding model turns words, sentences, or documents into lists of numbers (vectors) that capture meaning.

An epoch is one complete pass through your entire training dataset. If you have 100,000 examples, one epoch means the model has seen all 100,000 once.

Evaluation in AI is figuring out if your model actually works — not just on test data, but in the real world.

The F1 Score is the balanced average of precision and recall — a single number that captures both.

Few-shot learning is teaching an AI by showing it just a few examples — like showing someone two pictures of a platypus and saying 'find more of these.

Few-shot prompting is showing the AI a few examples of what you want before asking it to do the thing.

Showing the AI a few examples of what you want before asking it to do the real thing. Like 'here are 3 examples of good commit messages.

Fine-tuning is like taking a smart graduate student who knows everything and then sending them to a specialist bootcamp.

Fleet management for AI agents is monitoring, debugging, and operating many agents in production. Like SRE but for agents.

Function Calling is the OpenAI term for what Anthropic calls Tool Use — teaching the AI to call your code functions.

A GAN is two neural networks fighting each other. One (the Generator) tries to create fake images that look real.

GGUF is a file format for running AI models on your laptop — it's like the MP3 of AI models.

GPT is the brand of AI model from OpenAI that kicked off the LLM revolution. GPT-3 made everyone's jaw drop, GPT-4 made jaws stay dropped.

GPT-4o is OpenAI's 'omni' model — the Swiss Army knife of AI.

A GPU was originally built for rendering graphics in games, but turns out it's also perfect for AI.

Gemini is Google's answer to ChatGPT — their flagship AI model family. It comes in sizes from Nano (runs on phones) to Ultra (competes with GPT-4).

Generative AI is AI that creates new stuff — text, images, code, music, video — rather than just classifying or predicting. ChatGPT writes essays.

Copilot Workspace is GitHub's vision of AI-powered development at the project level. Instead of autocompleting one line, it reads your entire issue, propos

Google AI Studio is Google's playground for testing Gemini models without writing code. It's like a sandbox where you can chat with AI, test prompts, tune

Gradient Descent is how an AI learns — it's the algorithm that nudges the model's weights in the right direction after each mistake.

Groq built custom AI chips (LPUs) that make language models run ABSURDLY fast. While everyone else is using GPUs, Groq's hardware generates tokens so quick

Grounding is giving the AI real, verified information to base its answers on — so it doesn't just make stuff up.

Guardrails are the safety nets you put around AI applications — rules and checks that prevent the AI from going rogue.

Hallucination drift is when an AI starts with one small made-up fact, then builds on it, and by the end the entire conversation is based on fiction.

Hugging Face is like the GitHub of AI — it's where everyone shares their AI models, datasets, and demos. Need a sentiment analysis model?

Hyperparameters are the settings you configure BEFORE training starts — as opposed to parameters (weights) which the model learns ON ITS OWN.

Image classification is teaching a computer to look at a picture and say what it is — 'that's a cat,' 'that's a dog,' 'that's a suspicious mole you should...

In-context learning is the AI's ability to learn new tricks just from what you put in the prompt — without changing any of its actual brain weights.

Inference is when the AI actually runs and generates output — as opposed to training, which is when it's learning.

A jailbreak is a sneaky prompt that tricks an AI into ignoring its safety rules.

KV cache is how LLMs remember previous tokens without recomputing them.

Knowledge distillation is teaching a small model to mimic a big model — like having a genius tutor teach a regular student.

An LLM is a humongous AI that read basically the entire internet and learned to predict what words come next, really really well.

A startup that's basically just a pretty UI on top of someone else's AI model. 'We built an AI-powered email assistant!

LM Studio is a desktop app that lets you download and run AI models locally with a nice GUI — no command line required.

LangChain is a framework for building AI applications that need to do more than just chat.

The full name for an LLM — a really, really big AI trained on mountains of text.

Latent space is the AI's internal 'imagination room' — a hidden mathematical space where concepts live as points.

Layer normalization is batch norm's sibling — but instead of normalizing across the batch, it normalizes across the features of each individual example.

Llama is Meta's open-source AI model — it's like if one of the big tech companies just... gave away their homework.

LlamaIndex is a framework that makes it easy to connect your AI to your data. Got a bunch of PDFs, databases, APIs, and Notion pages?

LoRA is how you fine-tune a massive AI model without needing a massive GPU budget.

Local AI means running AI models on your own computer instead of sending data to the cloud.

Running an AI model on your own computer instead of calling an API in the cloud. No internet needed, no API costs, total privacy. The tradeoff?

The loss function is the AI's score of how badly it's doing.

Lovable is an AI tool that builds full web apps from a text description.

An MCP Server is like a plugin for your AI assistant.

The cool shorthand for Machine Learning.

DevOps but for machine learning.

Machine Learning is teaching a computer by showing it thousands of examples instead of writing out every rule.

MetaGPT simulates an entire software company with AI agents playing different roles — product manager, architect, engineer, QA. Give it a one-line idea and

Midjourney is the AI image generator with the best aesthetics — it makes everything look like a movie poster or concept art.

Mistral is the French AI startup that keeps punching above its weight.

Mixture of Experts is like having a team of specialists instead of one generalist.

A model is the trained AI — the finished product.

When AI trains on AI-generated data and slowly goes insane. Each generation gets slightly worse, like a photocopy of a photocopy of a photocopy.

MCP is a standard protocol that lets AI assistants connect to external tools and data sources — like a universal adapter for AI.

Model routing is dynamically choosing which AI model to call based on task complexity, cost, or latency — the smart switchboard for LLMs.

Model serving is the infrastructure that takes a trained AI model and makes it available as a fast, reliable API.

Model Stockholm syndrome is when you've been working with a specific AI model so long that you've internalized its quirks and stopped noticing them.

Multi-agent means multiple AI agents working together, each handling a different task.

Multi-agent debugging is figuring out which AI agent broke things when five of them were running in parallel.

Multi-head attention is running multiple attention mechanisms in parallel — like having several detectives investigate the same crime scene but looking for...

Multimodal AI can see, hear, AND read — it's not limited to just text. It's like the difference between texting someone and FaceTiming them.

NLP is the branch of AI that deals with human language — reading it, writing it, translating it, summarizing it.

The full name for NLP — making computers understand and produce human language.

A neural network is a system loosely inspired by the human brain — lots of little math nodes connected together, passing numbers to each other.

OCR reads text from images — take a photo of a document, receipt, or sign, and OCR turns the pixels into actual text your computer can search, copy, and ed...

Ollama is Docker for AI models. One command downloads and runs any open-source AI model on your computer.

Open Interpreter is like ChatGPT but it can actually DO things on your computer. It runs code, manages files, browses the web, and controls your system — a

AI models where the weights are publicly available — anyone can download, run, modify, and fine-tune them.

OpenAI is the company behind ChatGPT, GPT-4, DALL-E, and Codex.

The OpenAI API is the gateway to GPT models that kicked off the AI revolution. Send text in, get smart text back. It's the API that launched a thousand sta

Overfitting is when your model gets TOO good at the training data and becomes useless on new data.

Parameters is the technical word for weights — the individual numbers inside an AI model. When someone says 'GPT-4 has 1.

Perplexity (the metric) measures how 'surprised' a language model is by text — lower perplexity means the model predicted the text well.

Pre-training is the first massive phase where an AI reads basically the entire internet and learns to predict the next word billions of times.

Precision asks: 'Of all the times the AI said YES, how often was it actually right?

Prefix cache is when an AI provider reuses computation from shared prompt prefixes.

A prompt is the message you send to an AI to get it to do something. 'Write me a poem about JavaScript' — that's a prompt.

Prompt archaeology is digging through an old prompt (or someone else's system prompt) to figure out how it works.

Prompt burnout is the mental fatigue from iterating on prompts all day. You've rewritten the same instructions 30 times. Your brain melts.

Prompt caching is a speed and cost optimization where the AI remembers the beginning of your prompt so it doesn't have to re-process it every time.

Prompt cargo cult is blindly copying prompt tricks because you saw them on Twitter — 'Think step by step', 'You are an expert', 'Take a deep breath' — with...

Prompt compression is shrinking a prompt so it fits more context or costs less, without losing meaning.

Prompt injection is the SQL injection of the AI world.

Prompt laundering is rewording a prompt to get around an AI's refusal without actually changing what you're asking for. Mildly manipulative.

Prompt lore is the body of tips, tricks, and rumors about getting the best out of AI models. 'Claude responds better to XML tags.' 'GPT likes JSON schemas.

Prompt pruning is cutting unnecessary instructions out of a long prompt without hurting quality. Every word costs tokens and attention.

Prompt relay is passing one AI call's output directly as the next call's input — the chain in prompt chaining, minus the human review.

Prompt smithing is the craft of shaping prompts through deliberate iteration — testing, tweaking, measuring. Not vibes-based prompting.

Prompt surgery is surgically editing your prompt to fix a specific failure, rather than rewriting the whole thing.

A prompt template is a fill-in-the-blanks prompt for AI — you write the structure once with placeholders, then swap in different values each time.

A promptsmith is someone who specializes in crafting prompts. Not a prompt engineer (too corporate). Not a prompt whisperer (too mystical).

PyTorch is the most popular framework for building AI models — it's like React for machine learning.

Quantization is the art of making AI models smaller and faster by using less precise numbers.

RAG is how you give an AI access to your private documents without retraining it.

RLHF is like training a puppy — instead of giving the AI a textbook, you let humans rate its answers with thumbs up or thumbs down.

A rate limit is the AI provider saying 'slow down, buddy.

An AI model that actually thinks before it speaks.

Recall asks: 'Of all the actual YES cases in the world, how many did the AI catch?' High recall means the model finds almost everything it should.

Red teaming in AI is trying to break the AI on purpose — like hiring someone to try to rob your bank so you can find the security holes.

Regression is like classification but instead of sorting things into categories, you're predicting a number. What will this house sell for?

Reinforcement Learning is how you train an AI by giving it rewards and punishments instead of labeled examples.

Replicate is the 'run AI models with one API call' platform. Want to run Stable Diffusion, LLaMA, or some obscure research model? Replicate hosts it, scale

Replit Agent is an AI that builds entire apps from a description — you tell it what you want, and it creates the project, writes the code, installs package...

RAG is like giving the AI an open-book exam instead of expecting it to memorize everything.

Self-attention is how a model looks at a sentence and figures out which words are most important to each other.

A self-correcting agent notices its own mistakes and fixes them without being told. It runs tests, sees failures, debugs itself, retries. Sounds magical.

Self-healing code is code that automatically fixes its own bugs in production.

Semantic drift is when the meaning of a term or concept shifts over time in a codebase or an AI conversation.

Semantic Kernel is Microsoft's SDK for building AI agents and plugins. It's like a universal adapter between your code and any AI model — write your logic

Semantic search finds results based on meaning, not just keyword matching.

Sora is OpenAI's text-to-video model — you type 'a cat riding a skateboard through Tokyo at sunset' and it generates an actual video of that.

Cody is Sourcegraph's AI assistant that actually reads your ENTIRE codebase, not just the open file. Sourcegraph already knows how to search giant codebase

Stable Diffusion is an open-source AI that turns text into images. Type 'a corgi in a space suit on the moon' and get exactly that.

Streaming is when the AI sends you its response word by word as it generates, instead of making you wait for the whole thing at once.

Structured output means getting the AI to respond in a specific format — like JSON or a filled-in form — instead of just freeform text.

In AI coding tools, a suggestion is what the AI offers for you to accept or reject — a ghost of code that appears in your editor, waiting for you to press...

Swarm debugging is multi-agent debugging, but for flat peer-to-peer agent swarms.

A sycophancy loop is when the AI agrees with everything you say, including when you're wrong. 'You're right, my previous answer was incorrect.

Synthetic data is fake data that's good enough to train real models.

A system prompt is the secret instruction manual you give the AI before the conversation starts. It sets the personality, rules, knowledge, and behavior.

Tabnine is an AI code autocomplete that was privacy-conscious before it was cool. It can run its model locally on your machine so your code never leaves.

Temperature controls how creative (or chaotic) an AI's responses are. Low temperature (like 0.1) makes it boring, safe, and predictable — great for code.

TensorFlow is Google's deep learning framework — it was king before PyTorch took the crown.

Text-to-Speech takes written words and reads them out loud with a computer voice. Old TTS sounded like a robot reading a phone book.

Together AI is the open-source model hosting platform that competes on price and speed. They host Llama, Mixtral, and dozens of open models with an OpenAI-

A token budget is the cap on how many tokens a request, session, or user can consume. Like a food budget but for AI.

Token burn is how fast your AI bill climbs because the model keeps re-reading the same context. Every turn of a long chat costs more.

Token tax is the ongoing cost of running AI features in production. Every API call costs tokens. Every request the user makes. It never sleeps.

A tokenizer chops text into pieces that the AI model can understand — but not in ways humans would expect.

FPS, but for text instead of frames. Tokens per second measures how fast an AI spits out words.

Tool hallucination is when an AI calls a tool that doesn't exist, uses the wrong arguments, or invents a function signature. 'read_database(query=...

Tool use is when an AI can call external functions, APIs, or programs to do things it can't do alone.

A tool use loop is when an AI agent gets stuck calling the same tool over and over. It reads a file, reads it again, reads it again.

Top-k limits the AI's word choices to the K most likely options. If K is 50, the AI only picks from the top 50 most probable words for each step.

Top-p (also called nucleus sampling) is another dial that controls how an AI picks its next word.

Training is the long, expensive process where an AI learns from data.

Transfer Learning is using knowledge a model already has from one task to help it with a different task.

The Transformer is THE architecture behind all modern AI. ChatGPT, Claude, Midjourney, Whisper — all transformers under the hood. The key innovation?

Underfitting is the opposite of overfitting — the model hasn't learned enough and does badly on BOTH the training data AND new data.

Variance in ML means your model is too sensitive to the specific training data it saw.

In AI, a vector is just a list of numbers. But it's a list of numbers that means something — like [0.23, -0.91, 0.44, ...

A vector database is a special database built to store and search embeddings.

Vibecoding is when you code by vibes — you tell an AI what you want, it writes the code, and you just go with the flow.

A vision model is an AI that can understand images — it's got eyes, basically.

Weights are the numbers inside a neural network that determine what it knows and how it behaves — they're the AI's 'brain cells.

Whisper is OpenAI's speech recognition model — it listens to audio and writes down what was said.

Windsurf is Cursor's main competitor — a full IDE built around AI. Its Cascade feature is an agentic assistant that can plan changes across multiple files.

Zero-shot learning is when you ask an AI to do something it was never explicitly trained on — and it just... does it.

Asking an AI to do something without giving it any examples first. You just describe the task and it figures it out.

o1 is OpenAI's 'thinking' model — instead of blurting out answers immediately, it takes a moment to actually reason through the problem like a human would.

vLLM is like a turbocharger for running AI models in production.