The Kinds of AI and What Each Is For

Picture a hardware store that sells everything with a blade under one sign: SHARP THINGS. Chisels, lawnmowers, scissors and a chainsaw, all in one aisle. That is roughly how the word AI gets used. This lesson sorts the aisle.

The first cut is simple. Predictive AI looks at data and gives back a number or a label. Will this customer cancel? Is this tumor likely cancer? How many kilowatt-hours will the town use at 6 p.m. tomorrow? Generative AI produces something new: a paragraph, a picture, a voice, a song, a video clip, a block of code. Predictive AI has been running quietly in banks, hospitals and warehouses for decades. Generative AI is the loud newcomer.

Most real products combine several kinds. Your phone's voice assistant uses speech recognition to hear you, a language model to figure out what you meant, maybe a search tool to find the answer, and speech synthesis to say it back. Knowing the parts helps you see which part broke.

The field guide

Kind	What it does	Learns from	Everyday examples	Typical failure
Rule-based systems	Follows if-then logic people wrote	Nothing; humans write the rules	Tax prep software, benefit eligibility screeners, phone menus ("press 2 for billing")	Breaks on cases nobody wrote a rule for
Classic machine learning (prediction and classification)	Predicts a number or sorts into categories	Tables of past examples with known answers	Spam filters, credit scores, fraud alerts, insurance pricing, power demand forecasts	Copies bias in old data; fails when the world changes
Recommender systems	Ranks items for each person	Clicks, watches, purchases, skips	YouTube and streaming queues, "customers also bought," social media feeds	Narrow loops; boosts whatever holds attention, true or not
Computer vision	Finds, labels and tracks things in images and video	Millions of labeled images	Phone photo search ("dog"), license plate readers, face recognition to open your phone, crop disease apps	Misidentifies under bad light or unfamiliar faces and scenes
Speech recognition and synthesis	Turns speech into text, and text into speech	Thousands of hours of recorded speech with transcripts	Auto captions, voicemail transcripts, dictation, audiobook narration, voice cloning	Mangles names, accents and crosstalk; cloned voices used for scams
Large language models (LLMs)	Predicts and writes text, answers questions, writes code	Huge amounts of text, then human feedback	ChatGPT, Claude, Gemini, writing assistants, customer service bots	States false things confidently (hallucination)
Image and video generators	Makes pictures, video or music from a text prompt	Hundreds of millions of images or clips with captions	Midjourney-style art tools, Gemini's image models, video tools like Veo; ImageStudio on this site	Garbled text and hands, invented details; deepfakes
Reinforcement learning	Learns a strategy by trial, reward and penalty	Its own attempts, scored by a reward signal	AlphaGo, warehouse robot routing, data center cooling control, tuning chatbots to be helpful	Games the reward: finds a shortcut that scores well but misses the point
Robotics and autonomous vehicles	Senses the physical world and moves in it	Sensor data, driving logs, simulation	Robot vacuums, warehouse robots, Waymo robotaxis, lane-keeping assist	Rare situations (odd road work, a person in a costume) and real physical harm
Agents that use tools	Plans steps and takes actions: searches, runs code, clicks, files things	Usually an LLM plus tools and instructions	Coding assistants that edit and test code, research assistants that browse and cite, booking bots	Small errors compound across steps; acts on bad instructions hidden in web pages

A closer look at a few

Classic machine learning

This is the workhorse, and the least glamorous. A model looks at rows in a table (age of account, size of purchase, distance from home, time of day) and outputs a score. It is cheap to run and can be checked against reality: did the flagged charges turn out to be fraud? That checkability is its great strength. Lesson 3 walks through how one learns, with numbers.

Recommender systems

Every feed you scroll is ranked by one. It predicts what you'll click, watch or buy, using your history and the history of people who behave like you. The model's goal is whatever the company tells it to maximize, usually time spent or purchases. It has no idea whether a video is accurate. That explains a lot about the internet.

Computer vision and speech

These turn messy signals (pixels, sound waves) into labels and text. They got good fast after 2012. Auto captions on a city council livestream are speech recognition. A plate reader on a highway pole is computer vision. Both make errors at rates that sound tiny until you multiply them by millions of reads a day.

Robotics and autonomous vehicles

The US government uses the SAE scale from Level 0 (warnings only) to Level 5 (drives anywhere, any conditions, no driver). Most "self-driving" features sold on new cars are Level 2: the car steers and brakes, but you must watch the road every second. Waymo runs Level 4 service: no driver, but only inside mapped areas. As of October 2026 its FAQ lists 13 metro areas for its own ride-hailing app, including Phoenix, the San Francisco Bay Area, Los Angeles, Dallas and Miami, plus Austin and Atlanta through Uber. No company offers Level 5.

Agents

An agent is the newest kind, and mostly a combination of the others. Take a language model, give it tools (a web browser, a calculator, a code runner, your email) and let it work in a loop: plan, act, look at the result, act again. In November 2024 Anthropic released the Model Context Protocol, an open standard for connecting AI assistants to outside tools and data, and other companies adopted it. Lesson 8 covers agents in more depth.