What Is AI, Really?

The term artificial intelligence gets thrown around so loosely that it has almost lost its meaning. Your email spam filter is AI. The recommendation engine behind your streaming service is AI. The chatbot you used last week to draft an email is also AI. These are wildly different things, so let us sort them out.

Narrow AI (also called weak AI) is software that does one specific task well. It cannot do anything outside its lane. The algorithm that detects fraud on your credit card is narrow AI: brilliant at spotting suspicious charges, completely useless if you ask it to write a poem.

General AI (sometimes called AGI, for artificial general intelligence) would be a system that can learn and perform any intellectual task a human can. This does not exist yet. No product you can buy or sign up for today is general AI, despite what some marketing pages might imply.

Generative AI is the category that exploded into the mainstream starting in late 2022. These systems generate new content (text, images, audio, code, video) based on patterns learned from massive datasets. When people talk about "AI" in everyday conversation now, they almost always mean generative AI. ChatGPT, Claude, Gemini, Midjourney, and Stable Diffusion all fall into this bucket.

The important takeaway: AI is not one thing. It is a broad family of technologies, and the generative AI tools you will use in this course are just one branch of that family.

The AI assistants you interact with (ChatGPT, Claude, Gemini) are all built on top of something called a Large Language Model, or LLM. An LLM is a type of neural network that has been trained on enormous amounts of text data. We are talking hundreds of billions of words pulled from books, websites, academic papers, code repositories, and more.

At its core, an LLM does something deceptively simple: given a sequence of text, it predicts what comes next. If you type "The capital of France is," the model assigns high probability to the word "Paris" appearing next. It does this not because it "knows" geography, but because it has seen that pattern thousands of times during training.

This next-token prediction, repeated thousands of times in rapid succession, is what produces the flowing paragraphs and detailed answers you see when you chat with an AI assistant. Each word (or piece of a word) is generated one at a time, left to right, with each new prediction informed by everything that came before it.

What makes modern LLMs remarkable is scale. These models have billions of parameters, the adjustable numbers inside the neural network that get tuned during training. GPT-4, Claude, and Gemini each have parameter counts in the hundreds of billions. That scale is what allows them to handle such a wide variety of tasks, from writing code to explaining quantum physics to drafting a birthday card for your aunt.

Training an LLM happens in two major phases, and understanding them helps explain why these models behave the way they do.

Phase 1: Pre-training. The model reads a massive corpus of text: a significant slice of the publicly available internet, plus books, code, and other sources. During this phase, the model learns grammar, facts, reasoning patterns, coding conventions, and a lot of general knowledge. It does this entirely through self-supervision: the model tries to predict the next word in a sentence, checks whether it was right, and adjusts its parameters slightly to do better next time. This process repeats trillions of times across the entire dataset. Pre-training is astronomically expensive, often tens or hundreds of millions of dollars in compute costs for a single run.

Phase 2: Fine-tuning with human feedback. A pre-trained model is impressive but not very useful as a conversational assistant. It might ramble, produce harmful content, or ignore your actual question. To fix this, developers use a process called RLHF, short for Reinforcement Learning from Human Feedback. Human reviewers rate the model's responses (is this helpful? is this safe? does it follow instructions?), and the model is further trained to produce outputs that score well on those ratings.

This two-phase approach is why AI assistants are both remarkably knowledgeable and surprisingly easy to talk to. The pre-training gives them breadth, and the fine-tuning gives them manners.

Here is something that surprises most people: LLMs do not process text word by word. They break everything down into tokens, which are chunks of text that might be a whole word, part of a word, or even a single character.

For example, the word "understanding" might be split into two tokens: "under" and "standing." Common words like "the" or "is" are usually a single token. Rare or technical words get broken into smaller pieces. Even spaces and punctuation are tokens.

Why does this matter to you? Three practical reasons.

First, context windows. Every model has a maximum number of tokens it can process in a single conversation, known as its context window. If a model has a 128,000-token context window, that is roughly 96,000 words. Anything beyond that limit gets dropped. When you paste a very long document into a chat, you are using up tokens fast.

Second, cost. If you ever use an AI API directly, you pay per token. Knowing that a typical English word is about 1.3 tokens helps you estimate costs.

Third, behavior at the edges. When you are near the context limit, the model may start "forgetting" things you said earlier in the conversation. It is not being careless. It literally cannot see those tokens anymore.

One of the most important things to understand about LLMs is that they can be confidently wrong. The AI community calls this hallucination, which is when a model generates information that sounds plausible and is stated with full confidence, but is factually incorrect.

This happens because of how the model works at a fundamental level. Remember, it is predicting the most likely next token based on patterns. It is not retrieving facts from a verified database. If the patterns in its training data suggest that a certain combination of words is probable, the model will produce it, whether or not it is true.

Common hallucinations include inventing academic papers that do not exist, attributing quotes to people who never said them, generating plausible-sounding but completely fabricated statistics, and confidently describing events that never happened.

The tricky part is that hallucinated content looks identical to accurate content. There is no warning label, no asterisk, no change in tone. The model does not know it is wrong because it does not "know" anything in the way you do. It has no internal fact-checker.

This is why verification is non-negotiable when using AI for anything that matters. If an AI assistant tells you a legal statute, a medical fact, or a historical date, treat it as a starting point and verify it through a reliable source. The model is an excellent first draft generator, not a reference library.

AI discourse tends to swing between two extremes: breathless hype ("AI will replace all jobs within five years") and dismissive skepticism ("it is just autocomplete"). The truth is more interesting and more useful than either extreme.

What AI can genuinely do well in 2026: Draft and edit text at a high level. Write functional code in most programming languages. Summarize long documents accurately. Translate between languages. Brainstorm creative ideas. Analyze data when given clear instructions. Answer factual questions about well-documented topics. Generate images, audio, and video from text descriptions.

What AI still struggles with: Tasks that require real-world physical interaction. Truly novel reasoning that goes far beyond its training data. Maintaining perfect factual accuracy without external verification. Understanding your specific personal context without being told. Long-term planning and multi-step problem solving that requires common sense about the physical world. Anything that requires genuine empathy or emotional understanding.

What AI fundamentally cannot do: Think, feel, want, or understand in any human sense of those words. It processes patterns. That is powerful, more powerful than most people expected, but it is not consciousness, and it is not general intelligence.

The most productive mindset is to treat AI as an extraordinarily capable tool that amplifies what you can do. It will not replace your judgment, your expertise, or your ability to verify what is true. But it will make you dramatically faster at the parts of your work that involve producing, transforming, or analyzing text and code.