You send four words. Something answers in half a second. By the end of today you can explain that whole trip — and know exactly where you need to start checking it.
You send four words. Something answers in half a second. By the end of today you can explain that whole trip — and know exactly where you need to start checking it.
Guess first. Test it. Then decide what’s actually true.
That loop is how we’ll learn all week.
Your phone ranked your notifications, unlocked with your face, fixed a typo, and suggested a video — before you even opened a chatbot. Most AI doesn’t look like a chatbot at all.
A chess engine. Face unlock. Autocorrect. A chatbot. A thermostat.
Funny thing: once something gets normal, people stop calling it AI. Autocorrect used to feel like magic. Now it’s just autocorrect.
Pick where you’d draw the line. Then say the rule out loud.
There’s no single right line. A chess engine, face unlock, and autocorrect all count as narrow AI; a plain thermostat usually doesn’t. The rule you named matters more than exactly where you drew it.
AI is the whole family. Machine learning is the kind that learns from examples instead of being told rules. Generative AI is machine learning that makes new things — pictures, sentences, songs. An LLM is a generative AI built just for language.
A chess program that just follows a rulebook still counts as AI — it never had to learn from examples.
Artificial intelligence
any machine doing something that usually takes a human brain
Machine learning
learns from examples instead of fixed rules
Generative AI
makes something new: words, pictures, sound
Large language model
a generative AI built for language
OpenAI is the company. GPT is the model. ChatGPT is the product wrapped around it. And this chat, right now, is just today’s conversation — it disappears once you close it.
Company
builds or hosts
Model
maps inputs to outputs
Product
adds interface, tools, rules
Conversation
your current exchange
ChatGPT is a product. GPT is a model family. The chat is your session.
A language model writes new sentences that sound right. A search engine finds pages that already exist. A chatbot can do both — search first, then write you an answer using what it found.
When a chatbot cites a website, it searched first, then wrote about what it found. Two different jobs, done back to back — you just see one answer.
generate a fitting continuation
find material that already exists
This is the whole trip in five words. We’ll stop at every one of these five spots today.
The machine doesn’t read your sentence like you do. It chops it into pieces called tokens — sometimes a whole word, sometimes just a chunk of one.
Different AI systems can chop up the same sentence differently. There’s no one correct way to slice it — just whatever the builders trained it to do.
A simplified illustration of the idea — real tokenizers use a trained method called BPE.
A common word usually stays in one piece. An unusual name, or a word from another language, often gets chopped into more pieces.
A simplified illustration of the idea — real tokenizers use a trained method called BPE.
One person says, “The cat sat on the…” Everyone else writes one possible next piece. Count the answers — you just built the same kind of list the machine builds, except it does it in a fraction of a second.
Watch what the room does without being told to. A handful of words like mat, rug, and chair usually crowd out the rest, long before anyone mentions probability.
Like your class just did, the machine keeps a few live options and usually picks the most likely one — but not always.
There’s a dial called temperature. Turn it down and the machine plays it safe, almost always picking the obvious word. Turn it up and it takes more risks — more surprising, sometimes weirder answers.
Just for illustration — real numbers vary by model.
Every word it picks becomes part of the question for the next word. A whole paragraph is really just this loop, run thousands of times in a row.
Low temperature: careful, predictable, a little boring. High temperature: creative, surprising, sometimes off the rails. Either way, it’s not checking facts — just picking differently.
This setting only changes how it picks between options it already had. It can’t add a fact that wasn’t already possible.
Low temperature (T = 0.3)
High temperature (T = 2.0)
Just for illustration — real numbers vary by model.
Training happens once, ahead of time, over weeks — that’s when it actually learns. Every chat you have afterward doesn’t change what it learned; it just hands the already-trained machine today’s question.
Typing to it doesn’t teach it anything in the moment — not unless the company later uses saved chats to train a future version.
Training room
billions of times, before you ever showed up
Answering room
just now, once
It guesses the next word, gets told the real answer, nudges itself a tiny bit toward being right next time — and does that guess-and-nudge billions of times before you ever meet it.
Training room
billions of times, before you ever showed up
Answering room
just now, once
Those numbers are called weights — like millions of tiny dials. No single dial holds a fact or a sentence. Together, all the dials just make some next words more likely than others.
A parameter is one of those dials. Modern models have billions of them — that’s what people mean when they talk about the “size” of a model.
Just a picture of the idea — not real measured numbers.
What’s on that desk? The rules it was given, your message, whatever you said earlier in the chat, and any files you handed over. If something isn’t on the desk, it simply isn’t part of the answer.
The desk isn’t infinite, and it’s not evenly good at using everything on it — something buried in the middle of a long chat can get missed even though it’s technically “there.” Being on the desk isn’t the same as being used.
What you put in front of it
The context window — its desk
nearly full
It only knows what fits here. Run out of room and the oldest bits drop off.
What you get
An answer built from exactly what's on the desk.
Nothing on the desk? It guesses — and that's when it goes wrong.
It might be pulling from what it learned in training, your prompt, earlier in the chat, saved memory, a file you gave it, or a tool like web search. Figure out which one — that tells you how hard to double-check it.
Ask for a dinner idea. Then add: one guest is allergic to nuts, there is no oven, and only twenty minutes remain.
A quick stir-fry, ready before anyone's actually hungry.
A worked example, not a live model response.
The allergy usually does the most work. It rules out whole dishes outright, where the oven and the time limit just narrow the field.
The machine’s whole job is to finish your sentence in a way that sounds right. Nobody built it to check if it’s actually true — that part’s still on you.
What continuation fits the patterns?
What evidence supports the claim?
It can sound totally sure and be completely wrong. It can sound hesitant and be totally right. Tone tells you nothing about whether it checked.
Just illustrating the four combinations — not real measured odds.
If it doesn’t have the real answer, it won’t necessarily say “I don’t know.” It can invent a believable title, date, or quote instead. The made-up detail sounds more convincing, not more true.
This is what people call a “hallucination” — a confident, made-up answer. It can happen for a few different reasons, but the fix is always the same: check it.
Ask it about a book that doesn’t exist, or a rule you made up. We’re not trying to trick it for fun — we’re practicing noticing when you need to double-check.
Before you even ask: what would count as proof?
A wrong answer that sounds sure is harder to catch than one that hedges. If nobody named a way to check before asking, that gap is the real finding.
Pick one fact you could actually check. Find who owns that fact. Check when it was true. Then look for someone else — who didn’t just copy the first source — saying the same thing.
Even a real-looking link can be broken, out of date, or not actually saying what it’s cited for. Always open it and read the actual sentence.
The same machine can help you brainstorm a cool name for your tank, and also confidently give you wrong medical advice. How much you check should depend on how bad it would be to be wrong.
ideas, drafts, translations, first tries
health, money, safety, facts that matter
Cover the labels and explain each stop yourself. If you can do that, you’ve got the whole trip — no more black box.
Hang onto one thing from today that surprised you. Tomorrow: we open it up further and ask how it turns numbers into actual language.