Walk into any conference room in 2026 and someone will say "we should add AI to this." The word is doing an enormous amount of work in that sentence — and most of the people nodding along mean different things by it. Before you can design with AI, ship with AI, or argue against using AI, you need a clear picture of what the word actually refers to today.
- Explain what "AI" means in 2026 — and what it doesn't.
- Distinguish narrow AI, general AI, and Artificial General Intelligence (AGI).
- Tell rules-based software apart from learned behavior, and know why that line defines "AI" today.
- Name the three ingredients — data, compute, and the transformer architecture — that made AI suddenly useful.
- Recognize the modern AI products you already use, visible and invisible.
What "AI" means in 2026
Three years ago, if you asked ten people to define AI you'd get ten answers: chess computers, self-driving cars, the dystopian robot from a movie, the algorithm Netflix uses to suggest shows. In 2026 the answers have collapsed. When someone says "AI" in a product meeting today, they almost always mean a system built on a large language model — ChatGPT, Claude, Gemini, or one of their cousins. The word has been quietly hijacked by one specific kind of technology, and that's the technology this course is about.
This matters because the word "AI" is doing two jobs at once. The dictionary definition — machines that mimic human cognition — covers everything from a thermostat to a Mars rover. The product definition — what people are buying, building, and arguing about right now — is much narrower: software that reads and writes text (and sometimes images or audio) with surprising fluency, by predicting what comes next based on patterns it has absorbed from enormous amounts of human-created data.
For the rest of this course, when we say "AI" without qualifiers, we mean the second thing. We'll be explicit when we mean the broader academic sense.
Narrow AI, General AI, and the AGI question
The AI you actually use is narrow AI. Narrow means it does one specific thing well: ChatGPT writes text, Midjourney generates images, your spam filter classifies email. Each of these is genuinely impressive at its task and useless outside of it. A spam filter has no opinion on what to cook for dinner, and ChatGPT cannot drive a car.
General AI, in casual usage, refers to a system that can flexibly handle many kinds of tasks without being purpose-built for each. Modern LLMs sit awkwardly here. They were built to predict text — narrow — but the patterns they learned are so broad that they can summarize a contract, write code, plan a trip, and translate a poem. That feels general, and in a practical sense it is, but they're still missing things humans take for granted: long-running memory, real-world action without help, and the ability to genuinely learn from a conversation.
Artificial General Intelligence (AGI) is the imagined further step: a system that matches or surpasses humans across the full range of cognitive work. As of 2026, AGI does not exist, no one credibly claims it does, and serious people disagree wildly about whether it will arrive in five years, fifty, or never. The honest position is: build what works today, take the AGI debate seriously, and don't let either hype or dismissal cloud the decisions in front of you.
Rules-based software vs learned behavior
Here is the most useful distinction in this entire lesson, and the one most professionals get wrong.
A 2005 email filter blocks messages that contain the word "Viagra" because an engineer wrote a rule that said if subject contains "Viagra" then mark as spam. The system is doing exactly what it was told and nothing more. It cannot adapt, it cannot generalize, and it cannot block a new variant unless someone updates the rule. That's rules-based software, and it has run the world for sixty years. It is excellent at problems where the rules are knowable and stable.
ChatGPT blocks the same email — but not because anyone wrote a rule. It read millions of examples of spam and not-spam during training, and from those examples it learned patterns about what spam tends to look like. When a new variant arrives, it can often recognize it as suspicious without ever having seen it before. That's learned behavior. It is excellent at problems where the rules are fuzzy, change constantly, or are simply too numerous for a human to encode.
The shift from rules-based to learned behavior is the substantive technical change underneath the word "AI". If a system follows handwritten rules with no learning component, it is not AI by today's definition, no matter how clever or automated it is. If a system's behavior emerged from data, it is.
Why AI suddenly feels useful
Researchers have been working on neural networks since the 1950s. The math behind modern AI was largely settled by 2017. So why did everything change in late 2022, when ChatGPT became the fastest-growing consumer product in history? Three things finally landed at the same time.
Data
Training a useful language model requires absurd amounts of text. The internet provided it: web pages, books, code, conversations, scientific papers — trillions of words scraped, cleaned, and fed into the training process. No prior decade had a corpus that large.
Compute
Training also requires enormous computational power. Specialized chips called GPUs, originally built for video games, turned out to be perfect for the math involved. The cost per unit of compute fell, the chips got faster, and what would have taken a decade became a multi-month run on a data center floor.
A breakthrough architecture
In 2017 a research team at Google published a paper introducing the transformer — a new way to organize neural networks so they could pay attention to relevant parts of a long input. Transformers scaled better than anything that came before. Throw more data and more compute at them and they got dramatically better, in ways that earlier architectures did not. Every modern LLM you know is a transformer underneath.
Data plus compute plus transformers is the recipe. Take any one ingredient away and 2026 looks very different.
AI products you already use
If you think you've never used AI, you almost certainly have. Modern AI is invisible far more often than it's visible.
The obvious ones are the chat surfaces: ChatGPT, Claude, Gemini, Perplexity, the AI features inside Notion and Google Docs. These are the products that put a chatbot in front of you and let you type.
The less obvious ones are everywhere else. Your phone's predictive keyboard runs a small language model. Your photo app uses computer vision to find pictures of your dog. Spotify's Discover Weekly is recommendation AI. Gmail's "Smart Reply" suggestions are an LLM. TikTok's For You page is one of the most powerful recommendation systems ever deployed. Even a tool you'd never think of, like your bank's fraud-alert text, is increasingly powered by AI underneath the rules.
A good exercise — and the one your activity below asks for — is to spend a single day noticing. You'll find AI in places you would have called "software" a year ago. That diffusion, more than any single product launch, is what makes 2026 different.
ChatGPT — the product that put AI into everyday life. When OpenAI released it in November 2022, no one inside the company expected the scale of the response: a million users in five days, a hundred million in two months, faster adoption than any consumer product before it. The model behind it (GPT-3.5 at launch) had existed for a year and a half. What changed was the interface — a simple text box, a chat history, free access, and the everyday miracle of asking a computer almost anything and getting a coherent answer.
The lesson for product people: the breakthrough was as much about packaging as it was about the model. The same underlying technology had been available through an API since 2020 and almost nobody used it. A chat UI changed everything. Keep that in mind whenever you read about an "AI breakthrough" — half the time the technology is older than the headline suggests, and the real story is about access and design.
- Calling any automated system "AI." A thermostat or a hand-coded "if-then" workflow follows fixed rules — if no behavior was learned from data, it isn't AI by today's definition.
- Treating today's LLMs as AGI. ChatGPT and its cousins are narrow systems that feel general; they still lack lasting memory, independent real-world action, and true learning from a conversation.
- Assuming an "AI breakthrough" headline means brand-new technology. The model behind ChatGPT existed long before launch — the leap was the chat interface and free access, not a new model.
- Believing you have to see a chatbot to be using AI. Most AI is invisible: keyboard predictions, photo search, recommendations, and fraud detection.
- Using "AI" loosely in a product meeting. Without saying whether you mean the academic sense or "an LLM-based feature," the team ends up building toward different things.
Spend 10–20 minutes, no code required. List five AI products you used this week. For each, guess: is it narrow or general AI? Is it generating something, predicting something, or recognizing something? Don't worry about being right — the goal is to notice how often AI is already in your day, and to start training your eye for what kind of AI is doing what kind of job.
Key Takeaways
- In 2026, "AI" in product contexts almost always means a large language model — ChatGPT, Claude, Gemini, or similar.
- All AI you actually use is narrow AI. AGI doesn't exist yet, despite the marketing.
- The technical heart of the shift is from rules-based to learned behavior. If the system's behavior wasn't learned from data, it isn't AI by today's definition.
- Three things made the 2022 takeoff possible: massive data, cheap compute, and the transformer architecture.
- Most AI you use is invisible — keyboard suggestions, recommendations, photo search — not chatbots.
"AI" in 2026 is shorthand for large-language-model products, all of which are narrow AI rather than the still-hypothetical AGI. The defining shift is from rules-based software to behavior learned from data, made possible when massive data, cheap compute, and the transformer architecture arrived together. Most of that AI is already woven invisibly into the tools you use every day.