Meet the Machine — Party School

First, the plain truth

Underneath it all is a very good guesser.

The engine inside ChatGPT, Claude, and Gemini has a plain name: a large language model, or LLM. Here is the whole idea in one sentence. It read an enormous amount of text, most of the public internet plus a lot of books, and from all that reading it got very, very good at one trick: guessing the next few words. You give it some words, it predicts what would sensibly come next, then predicts the next bit after that, and keeps going until it has an answer.

That's it. That is the machine. It sounds almost too simple to be useful, and the surprising part is how far that one trick goes. Guess the next words well enough, over and over, and you get something that can draft your emails, explain a contract clause, or talk through a decision with you. Knowing it is a guesser, not a database, explains almost everything it does well and everything it gets wrong. You are not talking to a search engine that looks things up. You are talking to a pattern that fills in what usually comes next.

So the two habits that make it work follow directly. Give it plenty to go on, because it can only pattern-match against what you put in front of it. And read what comes back, because a confident guess and a correct answer look exactly the same on the screen. That's not a warning, it's just how a guesser works, and it's a skill you already have from every other part of your expert life.

Source: Anthropic Academy · OpenAI Academy

The whole machine · you type, it guesses the next words, you keep the good parts

The one model to remember

"AI" is really three modes wearing one name.

Here is the thing that untangles most of the confusion. When people say "AI," they mean one of three different setups, and they almost never say which. The three modes are chat, tool, and agent, and they line up like rungs on a short ladder: each one can do everything the one below it can, plus a little more. Learn these three and you'll always know what you're dealing with.

Mode one is chat. You type, it types back. That's the whole thing. It can talk about anything, explain, brainstorm, and draft, but it cannot touch the world. It can't open your files, can't look anything up, can't send an email. It only knows what it read during training plus what you paste into the conversation. A plain ChatGPT or Claude window with nothing switched on is chat.

Mode two is tool-using. Same machine, but now it can reach past the chat window. It can search the web for something current, read a document you upload, or write a draft into your inbox. You still ask for each thing, one at a time, but now it can actually go get things and do things instead of only writing about them. When you turn on web search or attach a file, you've moved into this mode.

Mode three is an agent. Now you hand it a whole goal instead of a single step, and it takes several steps on its own to reach it, checking as it goes. "Find three podcasts my ideal client listens to, and leave a draft pitch for each in my inbox" is a goal, not a single ask. An agent searches, reads, drafts, looks at its own weak draft, tries again, and only then surfaces the result. It's the same guesser from the first section, now allowed to take a few steps by itself before it checks back in with you.

That's the entire model: chat talks, a tool-user reaches, an agent runs a short errand. Everything else in AI is a detail hanging off one of those three.

Sources: Anthropic, "Building Effective Agents" · OpenAI, "A Practical Guide to Building Agents"

Scroll the diagram sideways →

Chat, tool, agent · each rung adds a little more the machine can do on its own

The judgment call

Which mode am I in, and which do I need?

Most "AI didn't work for me" moments are really a mismatch: someone was in chat mode and needed a tool, or reached for an agent when a plain chat would have done the job in ten seconds. The good news is that matching the task to the mode is a five-second read, and you already have the judgment for it from your own work. Find your task on the left, and the mode you actually need is on the right.

The mode map · what you're doing, and the mode that fits
What you're trying to do	The mode that fits	Why that one
Think out loud, brainstorm names, get a first rough draft, or understand something.	ChatJust talk to it.	No outside information needed. The answer lives in what it already knows plus what you tell it in the message.
Ask about something recent, a price, a news item, this year's rules.	Tool-usingTurn on web search.	Its training has a cutoff date. For anything current it needs to look, so it has to be able to reach the web.
Have it work from your actual document, contract, or spreadsheet.	Tool-usingAttach the file.	It can't see your files unless you hand them over. Uploaded, it can read and work from the real thing.
Get a draft written straight into your email or a doc, not just shown in the chat.	Tool-usingConnect the app.	Writing into your inbox is reaching out of the chat window, which is exactly what a tool connection is for.
Hand off a whole multi-step errand: research these five, draft each, leave them for me.	AgentGive it the goal.	It's several steps you'd rather not babysit. An agent takes the steps in order and shows you the finished stack.
Anything that sends, spends, publishes, or deletes for real.	YouKeep the last click.	Whatever the mode, the final yes stays yours. Let it draft and prepare; you're the one who presses send.

Notice how much lives in the first two rows. Most everyday wins are plain chat or a single tool switched on, not a full agent. Starting simple isn't the timid choice here, it's the correct one, and you can always climb a rung when the task genuinely asks for it.

The worksheet you leave with

Spot your three modes this week

Write down three real things you'd want help with this week, one per line, in your own words. Then, next to each, mark which mode it needs using the map above. That's the whole exercise, and it's the fastest way to make these three modes stick.

01A task you'd want help withwrite the real one, then mark it: chat / tool / agent
02A second real taskwrite it, then mark it: chat / tool / agent
03A third real taskwrite it, then mark it: chat / tool / agent
04Count your marksmost people find two chats and one tool, and that is normal, and it's exactly where to start this week

When you're ready: an agent actually has six named parts

You don't need this yet, and skipping it costs you nothing. But if you're the kind of person who likes to see the whole machine, here it is. When people who build agents draw one, they draw six parts: instructions (the standing rules you write once), context (everything it can see right now, like the files you attached), the model (the guesser that reasons and writes), tools (what it can actually touch, like search or your inbox), the loop (the plan-act-check-repeat that makes it take several steps), and you (the checkpoint, on purpose, not by accident).

The reason this is worth a peek and not a memorize: when an agent does something odd later, one of those six parts is usually the reason, and being able to point at the part is how people fix it in one move instead of retyping in a louder voice. The Understand-more lane takes all six apart in full. For now, the three modes are plenty.

Fresh from the lab

What changed this month

One AI company built a mode that splits itself into helpers. On July 9, OpenAI released a new version called GPT-5.6, and its top model, Sol, got an extra setting called Ultra. Instead of working through a hard problem step by step on its own, Ultra breaks the job into pieces and works on several at once, almost like sending out a few mini versions of itself to handle different parts before combining everything into one answer. On a tough coding test, that extra help pushed the score from 88.8% up to 91.9%. Using Ultra costs more, since it is doing more work at once, but the price for the regular version did not change.

The behind-the-scenes connector that lets AI reach your apps got a serious tune-up. There is a standard called MCP that lets one AI assistant plug into lots of different apps, your email, your files, your calendar, without a custom setup built for each one. On July 28, Anthropic updated it to run faster on lighter infrastructure, which means more companies can afford to offer the connection. It is catching on fast: Anthropic says it is now downloaded over 400 million times a month, four times as often as at the start of the year, and Claude alone can now connect to more than 950 different apps and services through it.

A free AI model just caught up on the "doing things" part. A company called Moonshot AI released a model called Kimi K3 on July 16 and made it free to download on July 26, no subscription required. What matters for this lesson: on tests that check whether an AI can actually finish a task using tools, not just talk about it, Kimi K3 came in first on four out of eight tests, including one that measures automating real work. It even nearly matched GPT-5.6 Sol on that same coding test from above. The free options are catching up fast, specifically on the skill this lesson is about: taking action, not just talking.

Sources: OpenAI, "GPT-5.6: Frontier Intelligence That Scales With Your Ambition", July 9, 2026 · Anthropic, "Bringing MCP 2026-07-28 to Claude", July 28, 2026 · MarkTechPost, "Moonshot AI Releases Kimi K3", July 16, 2026

Vocabulary

Six words, in plain English

These are the words you'll hear the most, translated into what they actually mean. Learn these six and the next AI conversation, pricing page, and settings menu gets easier to read. There is no quiz.

LLM

Large language model. The engine inside ChatGPT, Claude, and Gemini. It read a mountain of text and got very good at guessing the next words, which is how it writes and answers.

Prompt

Just a fancy word for whatever you type in. Your question, plus any background and examples you include. A better prompt gives it more to work from. That's the whole next lesson.

Chatbot

Mode one. The plain window where you type and it types back. It can talk about anything but can't reach out and touch your files, the web, or your email.

Tool use

Mode two. The machine reaching past the chat window: searching the web, reading a file you uploaded, or drafting into your inbox. Tools turn a talker into a doer.

Agent

Mode three. You give it a goal and it takes several steps on its own to reach it, checking its own work as it goes. You still say yes at the end.

Hallucination

When it states something wrong with total confidence. It's a guesser, so it fills gaps with plausible-sounding text. The fix is to give it the real source instead of asking from memory.

Plain-language guides, straight from the builders: Anthropic Academy · OpenAI Academy

Make it yours

Build your first simple helper

This is the smallest possible first win: one plain chat that already sounds like you and helps with the task eating the most of your week. No apps to wire up, no settings to hunt for. A free ChatGPT or Claude account is all you need, and about twenty minutes, once.

You'll stay in mode one, plain chat, the whole time, and that's on purpose. You're learning to hand it good materials and read what comes back. Those are the two habits from the top of this lesson, and everything you build later stands on them. Pick your line of work and follow the five steps.

For female founders

I'm the CEO, the marketer, the bookkeeper, and the intern, and the building only happens after 9pm.