A different kind of software

Every line of code in this curriculum shared one property. Written by a person, it does exactly what it says, every time. delivery_fee(19.00, True) returns 0, free delivery for the member, until the sun burns out.

Phase 8, the final stretch of the path, is about software that doesn’t work that way. And it’s about why that difference changes how products are built, tested, and trusted.

What is artificial intelligence?

Artificial intelligenceConcept · lights on your mapartificial intelligenceAny technique for getting computers to handle work that once needed a human mind. In current products the term nearly always points at one technology, models that learned their behavior from data, with LLMs the headline act. is, formally, the broad field of making machines do things that seem to require intelligence.

In today’s products, though, the word overwhelmingly means one family of technology: machine-learning models, and especially large language models. This module is about that family.

What is a machine-learning model?

A machine-learning modelConcept · lights on your mapmachine-learning modelA program nobody authored directly. Training shows it millions of examples and nudges its internal numbers until useful behavior appears, and what emerges behaves more like intuition than recipe. is software whose behavior was learned from data rather than written as rules. Here the recipe metaphor that has served you since Module 1 finally breaks.

The learning process itself is simpler than its results suggest. Show the model an example, a photo labeled cat, and it guesses. When it guesses wrong, nudge its billions of internal numbers so that guess would have been slightly less wrong. Then repeat, millions of times. The nudging is automatic, arithmetic run by the training software rather than judgment applied by a person.

No single number means anything readable. But together, after enough nudging, they hold something that behaves like a trained intuition, which is the better metaphor now that the recipe one is gone.

The code around a model, the part that accepts a request, feeds it in, and sends the answer onward, is ordinary software. The model itself is a different substance.

What is a large language model?

The kind that ate the world is the large language modelConcept · lights on your maplarge language modelAn ML model whose training data was a huge share of the written internet, and whose single skill is guessing the next fragment of text. At scale, that guess learns to answer, summarize, translate, converse, and write code, and it is the engine behind every AI assistant. (LLM): a model trained on oceans of text to do one thing, predict what comes next.

Here is what the box shows. The model takes the unfinished sentence and scores every word it knows on how well it would fill the blank. The percentages are those scores. A world of coffee-order text puts latte far ahead, leaves coffee a distant second, and buries pizza.

The answer you see is assembled by picking from the top of that list, one word after another. Nothing was looked up, and no database of orders was consulted. The model computed what usually follows.

The surprise of the decade is that at preposterous scale, predicting what comes next turns out to include answering questions, summarizing documents, translating languages, holding conversations, and writing Module 3’s Python.

Every AI assistant you have met is this engine, and “LLM” is the word in every meeting.

Training versus inference

The lifecycle splits sharply in two, between trainingConcept · lights on your maptrainingThe process of learning from data. At frontier scale it runs for months on warehouses of GPUs, at monumental cost, and it happens rarely, almost always inside an AI lab. Nearly no product team trains at that scale, and what products do instead is inference., the learning phase that builds the model, and inferenceConcept · lights on your mapinferenceThe act of answering one prompt with a model that already exists. This is the half a product team actually touches, usually by renting a provider’s API. It runs on every request, while training stays the rare, expensive half., the answering phase that puts it to work. The split decides who does what in the AI economy.

It is medical school versus seeing a patient, or years of culinary school versus cooking one meal. Training happens rarely and up front, at staggering cost, and inference happens constantly, per request.

The product consequence is worth saying early. Your company will almost certainly never train a model. Products do inference, typically by renting it through an API, a shape you have known since Module 5.

What is a GPU?

The economics of both halves run on the GPUConcept · lights on your mapGPUA chip built for drawing game frames, turned into AI’s engine room. Thousands of simple cores run arithmetic in parallel, which is the shape model math wants, and scarcity has promoted “compute” into boardroom vocabulary.: the graphics processor, repurposed as AI’s workhorse.

Module 2’s records office gains a wing. The CPU is one brilliant worker at a desk, while the GPU is thousands of simple workers moving in lockstep, each doing one small piece of arithmetic, all at once.

That shape happens to be exactly what model math wants. A model is billions of numbers, and answering a prompt means multiplying enormous grids of them, thousands of independent calculations at every step.

GPUs are scarce and expensive, since everyone training or serving models wants the same chips. That scarcity is why “compute”, the industry’s word for processing capacity itself, became a word CEOs say with feeling.

The mental model to remember

AI in today’s products overwhelmingly means machine-learning models, especially LLMs.

A machine-learning model’s behavior is learned, not written: examples in, billions of numbers nudged, behavior emerging. It is not a recipe but a trained intuition.

An LLM predicts what comes next in text, and at preposterous scale, that turns out to include answering, summarizing, and coding.

Training builds the model, rare and monumental, and inference answers with it, request by request. Products do inference, and almost nobody trains.

The GPU is thousands of simple workers in lockstep, the exact shape model arithmetic wants, and scarce enough to be a boardroom word.

You should now be able to hear “the model answered wrong” and know the strangest fact in this curriculum. No one can point to the line that did it, since no one wrote one.