What do the AI words on an engineering job description mean?

What do the AI words on an engineering job description mean?

The AI words on a req name four different kinds of thing: models, the companies that train them, the products assembled on top of them, and the techniques used to do the assembling. Most of the confusion around AI hiring comes from those four getting mixed together. "Claude" is a product. "Anthropic" is a company. "RAG" is a technique. "LLM" is a category of model. Once each word lands on its shelf, a sentence like "we use Claude with RAG over our own docs" reads as plainly as "we use Postgres for the database."

This page is the map of the whole territory, and each section below has a full lesson behind it. What a model is and who makes GPT, Claude, and Gemini covers the models themselves. Who the AI companies are sorts the businesses. Building AI versus using AI draws the single most useful line in AI hiring. RAG, agents, evals, and MCP covers the parts an AI product is made from. What an AI company actually built decodes the phrase itself, and is AI overhyped gives the honest account of what the technology does today.

What is a model, and what is a large language model?

A model is software whose behavior was learned from data instead of written line by line. Every program covered so far in this course was built from instructions a developer typed out. A model is the exception. Its builders showed it an enormous number of examples, and a process called training adjusted billions of internal settings until the model's output matched the patterns in those examples. Nobody wrote its rules by hand, which is why even its builders get surprised by what it can and cannot do.

A large language model, or LLM, is a model trained on text. It reads text in and writes text out, and that one ability covers answering questions, summarizing documents, translating, and writing code, because a program is also text. The text sent to the model is called a prompt. When a job description says "AI" with no further detail, an LLM is almost always the thing being described.

The famous names sort cleanly into companies and their products. OpenAI makes GPT and ChatGPT. Anthropic makes Claude. Google makes Gemini. The relationship is the same one that holds between Amazon and AWS, a pairing covered back in the cloud lesson: a company, and the thing it sells. The full lesson unpacks training, tokens, and the context window, the terms that govern what these systems cost and hold.

Who trains models, and who builds on them?

OpenAI, Anthropic, Google, Meta, and a short list of others train large models. Everyone else builds on top of them. The group at the top is small for a money reason: training a frontier model takes months of computing time on thousands of specialized chips called GPUs, and the bill runs into the hundreds of millions of dollars. The companies that pay it are called labs, and the general-purpose models they produce are called foundation models, because everything else in the industry rests on them.

The rest of the industry rents. A product company pays a lab per use and reaches the model through an API, the same doorway one program uses to talk to another anywhere in software. Meta took a different route and releases its Llama models as open weights, meaning any company can download the finished model and run it on its own machines instead of renting access.

This split is the first thing to establish about any AI company in a search: lab or builder. The labs employ research scientists and compete for a small pool of specialists. The builders, which is nearly everyone, employ product engineers. Who the AI companies are goes company by company, including the cloud providers and the tooling businesses in between.

What is the difference between building AI and using AI?

Building AI means training models, and using AI means calling a model someone else trained. The phrase "AI experience" on a resume can point at either one, and they describe different people. Training work is research: rare, concentrated in the labs, and done by specialists with math-heavy backgrounds. Using a model is product engineering. The developer writes the software around the model, decides what to send it, handles what comes back, and plans for the answer being wrong. That is backend work with one unusual component in the middle, and the much larger pool of developers who have shipped ordinary backend systems can pick it up quickly.

Two middle terms refine the picture. Fine-tuning means running a short extra round of training on a team's own examples to specialize an existing model, which is closer to building than most AI work gets outside a lab. Inference means running a trained model to get an answer, and it is the metered cost every AI product pays on every request.

One more reading of "AI experience" sits alongside these: some candidates build AI into products, and others use AI tools, such as a coding agent, to write ordinary software faster. Both are real experience with different substance behind them. Building AI versus using AI treats the whole distinction properly, and it is the lesson to read first if only one gets read.

What are RAG, agents, evals, and MCP?

These four words name the parts an AI application is assembled from, and together they cover most of the technique vocabulary on current reqs.

RAG, retrieval-augmented generation, grounds a model's answers in real documents. A model has read the public internet and has never read a company's internal wiki, contracts, or manuals. RAG bridges that: the software searches the company's own content for the pieces relevant to a question, hands those pieces to the model inside the prompt, and the model answers from what it was handed. Most of the work in it is search engineering.

An agent is a system that uses a model to decide on actions and then takes them in other software: searching a database, filing a ticket, sending a message. A chatbot only produces text in reply, and an AI workflow runs a fixed sequence of steps with a model filling in the middle. Which of the three a candidate built changes the difficulty of what they did.

An eval is a repeatable test of output quality. Model output varies, so teams build suites of known inputs with graded answers and run them on every change, the way conventional teams run tests. Eval work is a strong sign of production experience, because only teams accountable for quality in front of users bother.

MCP, the Model Context Protocol, is the standard connector between an agent and outside tools, introduced by Anthropic in November 2024 and adopted across the industry within months. An MCP server is a small connector program exposing one tool or data source. The new plumbing covers all four in depth.

What did a company build when it calls itself an "AI company"?

The phrase names no single thing, and the honest range runs from deep to thin. An AI company can be a lab training foundation models. It can be a product company whose core feature is an agent completing real work. It can be an existing product that added an AI feature, a summarize button or a smarter search. It can be a workflow business with a model in the middle of an otherwise conventional pipeline. Or it can be an AI wrapper: a thin product layer over a rented model, where the model does nearly all the work and the company added an interface and a prompt.

Wrapper is the loaded word in that list, and it deserves a fair reading. A thin layer that reaches the right users at the right moment is a legitimate business, and some wrappers grew into substantial companies by deepening from that start. The word describes the engineering surface area, and for hiring that is the useful measurement: the depth of what a company built sets what its engineers spent their time on, which sets what a candidate from that company has actually done. What an AI company actually built walks through how each shape shows up in the work history of the people who built it.

What does AI reliably do today?

Today's models reliably draft, summarize, translate, answer questions over supplied material, and write working code for well-described tasks. Those capabilities are in production at most large companies, and the productivity gain in software work, where a coding agent writes and revises code under a developer's direction, is the most visible of them. Treating this as settled infrastructure is now the reasonable default.

The failures are equally specific. The best-known is hallucination: a model stating something false in the same confident voice it uses for true statements. Nothing in how a model produces text checks facts, so a wrong answer looks exactly like a right one. Techniques like RAG and eval suites exist largely to contain this, and containment is the accurate word. The rate has come down and the behavior remains.

Models also degrade on tasks needing knowledge past their training date, on long chains of precise reasoning, and on anything where being wrong is expensive and unreviewable. The industry argument about whether all of this amounts to a revolution or a bubble is real, and both sides hold evidence worth knowing. Is AI overhyped closes the module with that argument laid out straight, so the vocabulary from this page comes with a working judgment attached.