Practice: What generative AI actually is
Self-check
Section titled “Self-check”Seven short questions. Answer each in your head before opening the collapsible. Active retrieval is where the learning sticks.
1. What does a generative language model actually do, and what are you watching when a reply appears word by word?
Show answer
Given some words, it predicts what word most likely comes next, then does it again, one word at a time, using everything written so far as its guide. When a reply appears word by word on your screen, that is not a typing effect. You are watching the machine actually work.
2. Generative AI existed before 2022. Why does November 30, 2022 still matter, and how should you treat that date?
Show answer
That day OpenAI released ChatGPT to the public as a free website, and for the first time a powerful language model reached millions of people who had never cared about AI. It crossed a reported million users within five days. Treat the date as a historical marker, like the launch of the web browser; the tools have changed many times since and keep changing.
3. Where do the model’s predictions come from, and why is it not memorization?
Show answer
The model’s makers feed it trillions of words of human writing from web pages, forums, and books. It cannot work by memorizing sentences, because most sentences have never been said before by anyone. Instead it learns the deep statistical patterns of how people write, so it can continue almost any passage in a way that sounds like us.
4. “That is just autocomplete.” What does that dismissal miss?
Show answer
What good prediction requires. To predict the last word of a sentence about a country’s capital, you have to know a fact about the world. To predict the next word of a sad story, you have to track feeling, grammar, and plot. Prediction done well enough soaks up a startling amount of knowledge. It is genuinely powerful, and it is still not a mind; both halves are true.
5. How is a generative model different from a search engine?
Show answer
A search engine retrieves pages that already exist. A generative model composes new text that sounds right. They are different machines with different failure modes, and asking one to be the other is where much frustration begins.
6. Name the three ways generative AI breaks from the older AI you already used.
Show answer
It is general (a chat assistant attempts almost any writing or thinking task, where a spam filter does one job). It is generative (it makes new things rather than labeling what exists). And it is approachable (you operate it in plain English). The framing comes from the Harvard course, which borrowed it from a 2023 Goldman Sachs analysis.
7. What two questions organize this track, and which lessons answer each?
Show answer
First, how do I use this well? Lessons 2 through 5. Second, what does it mean for my world? Lessons 6 through 9. The Harvard admissions story shows why you need both: the same tool that gives an applicant a real skill lands on the admissions office’s desk as a dilemma, and the admissions office is every office.
Try it yourself
Section titled “Try it yourself”Your first structured conversation
This is the gentlest practice in the whole track. The goal is not to accomplish anything. The goal is to watch the prediction machine work, so the core insight moves from something you read to something you saw. Open Clawless, the working environment we use across Clawdemy, and start a new conversation.
- Watch it write. Ask the model to continue a sentence five different ways, for example: Continue this sentence in five different ways: “The hardest part of my week is”. Watch each answer appear word by word. That is not an animation; it is the machine choosing the next word, one at a time.
- Prediction that knows things. Ask a factual question, such as what the capital of Australia is. Notice that answering required knowledge of the world. That is what good prediction quietly soaks up.
- Same question, twice. Ask a question with more than one reasonable answer, such as three tips for a better Monday morning. Then open a fresh conversation and ask the exact same question. Compare the two answers. Same question, different words. It is predicting likely text, not looking up a stored answer. You will handle this behavior deliberately later in the track.
- Make it about your week. Paste three or four rough bullet points about something real from your own week and ask for a short, clear note a colleague could read. Notice how confident and finished it sounds, and check it the way you would check a new coworker’s draft.
When you finish, say the core insight out loud once: it writes one word at a time, by prediction. Everything you just watched follows from that sentence. In the next lesson you start shaping these requests deliberately.
Flashcards
Section titled “Flashcards”Q. What does a generative language model do?
Given some words, it predicts what word most likely comes next, then does it again. One word at a time, until a whole answer has poured out.
Q. What are you watching when a reply appears word by word?
The machine actually working. It is not a typing effect; each word is a fresh prediction based on everything written so far.
Q. When did generative AI become a public phenomenon, and how fast?
November 30, 2022, when OpenAI released ChatGPT as a free website. It crossed a reported million users within five days.
Q. How should you treat the 2022 launch date?
As a historical marker, like the launch of the web browser. The tools have changed many times since and keep changing; the prediction insight is the part that stays true.
Q. Where do the model's predictions come from?
Trillions of words of human writing from web pages, forums, and books. It learns the deep statistical patterns of how people write, not memorized sentences.
Q. Why can the model not just memorize sentences?
Because most sentences you say have never been said before by anyone, ever. Patterns generalize; memorization would not.
Q. What did the Harvard classroom exercise demonstrate?
Students voted on the next word, one word at a time, and built the sentence “human language is surprisingly predictable.” Sharad Goel calls it “humans mimicking machines mimicking humans.”
Q. Why is good prediction more than autocomplete?
Predicting well requires facts, grammar, feeling, and plot, so it soaks up a startling amount of knowledge. As Goel puts it, “It’s a complicated prediction problem but it’s still really just prediction.”
Q. Is a generative model a mind?
No. It is a prediction machine that is genuinely powerful. Both halves are true, and both halves matter for how you use it.
Q. Name the three ways generative AI breaks from older, categorizing AI.
It is general, it is generative, and it is approachable. Framing from the Harvard course, borrowed from a 2023 Goldman Sachs analysis.
Q. Why is approachable the quiet revolution?
You operate the tool in plain English, the same language you use with a colleague. Clear thinking in ordinary words is the entry skill, so you are already qualified.
Q. How is a generative model different from a search engine?
A search engine retrieves pages that exist. A generative model composes new text that sounds right. Different machines, different failure modes.
Q. Why can answers sound confident even when wrong?
The machine predicts likely words; it does not look up verified facts. Confidence is a property of the writing style it learned, not of the truth.
Q. What two questions does this track answer?
How do I use this well (lessons 2 through 5), and what does it mean for my world (lessons 6 through 9).