What an AI actually is

Welcome. Nine of you, nine weeks, and this is where we start.

I've been building AI-native products for about two years. LotsTech is fifteen of them — four launched so far, the rest close behind. Almost every expensive mistake I made in that time came from having the wrong mental picture of what these systems are. So Day 1 isn't prompting and it isn't tools. It's just: what is this thing.

1. "AI" doesn't mean what it used to

For fifty years, "AI" meant robots, chess computers, self-driving cars. Machines doing logical things very fast.

What changed around 2022 is different, and for marketers it's the only part that matters: machines got good at language.

That's not a small thing for you, because marketing is language. Positioning is language. Copy, emails, posts, landing pages, customer research, competitor analysis — all language. A tool that handles language well touches almost everything you'll ever be asked to do.

2. What an LLM actually is

You'll hear "LLM" constantly. It stands for Large Language Model, and each word is doing real work:

• Large — trained on an enormous amount of text. Books, websites, articles, forums.
• Language — it works on words, not numbers or images (though some handle those too now).
• Model — a system that makes predictions.

So: a thing trained on a huge amount of writing, that predicts. And what it predicts is simply the next piece of text.

Here's the whole mechanism. Read this sentence:

You knew the answer before you finished reading — morning. You didn't look it up. You've read enough English that it's obviously the most likely word.

That's it. That's what the machine is doing, just at enormous scale.

Now draw the consequence out, because most people never do:

Often the plausible answer and the true answer are the same. That's why this works at all. But plausible and true are two different things, and the machine cannot tell them apart.

A picture that will serve you well all program: it's like a copywriter who has read everything ever published, works in seconds, will write in any style you ask — but has no access to your company's files, no memory of yesterday's meeting, and is never allowed to say "I don't know."

3. Model, assistant, agent are three different things

People say "AI" for all three. They are not the same.

• The model — the trained thing itself. GPT-5.6. Claude Opus 5. Gemini 3. Built once, then frozen. It's a file. On its own it does nothing.
• The assistant — the product you log into. ChatGPT, Claude, Gemini. It wraps a model and adds an interface, memory, file upload, web search.
• The agent — a program someone builds that calls a model through an API, gives it tools, and lets it do real work without a person typing each step.

Company	Model (frozen)	Assistant (the app)
OpenAI	GPT-5.6	ChatGPT
Anthropic	Claude Opus 5	Claude
Google	Gemini 3	Gemini
xAI	Grok 4	Grok

Anyone can pay for API access to these models and build their own assistant or agent on top. That's what LotsTech does, and what you'll be doing by Week 5.

4. The model is frozen in time. The assistant may or may not search.

• Every model is trained up to a cut-off date, then frozen. It knows nothing after that.
• The assistant can search the web and paste results in. Then the answer is current.
• But it does not always search. Often it just answers from frozen memory.
• Ask about something recent and you can get two-year-old information delivered as fact.

5. Why it's confidently wrong

When an AI states something false, that's called a hallucination. It's a bad name, because it sounds like a glitch. It isn't. It's the system doing exactly what it was built to do.

Ask it for data to support a campaign and you might get this:

Look at the shape of that. A real-sounding research firm. A plausible year. A number precise enough to feel measured — 23.4%, not "about 20%". It is a perfectly formed citation.

It may be entirely invented.

And here's the part that should worry you:

There's a second habit worth knowing about now. Ask a factual question, get a correct answer, then reply "Are you sure? I think that's wrong." Watch how often it folds and changes a correct answer to an incorrect one. These tools are tuned to be agreeable, and agreeableness sometimes beats accuracy.

Which means you cannot read anything into how confident it sounds. It sounds certain when it's wrong, and it backs down when it was right.

Now think about what that means in marketing specifically — a field that runs entirely on claims, numbers, and other people's trust. You can publish something false about a real company, with a real logo attached, in about four seconds.

6. It doesn't give you the same answer twice

This one surprises people most, and it's the reason you have to be careful about what you conclude from any single test.

Every computer you have ever used is deterministic. Ask a database the same question a million times and you get the same answer a million times. That is the deal computers have always offered: same input, same output, forever.

An LLM does not offer that deal. Ask it the same question in three fresh chats and you can get three different answers. Sometimes slightly different, sometimes genuinely different.

It isn't broken. Here is the mechanism, and it explains a lot at once:

• The model has no memory. None at all.
• The assistant re-sends your whole conversation as one block of text, every single time.
• The model reads that block cold, predicts a reply, and forgets everything.
• There is no "session" inside the model — only a growing pile of text handed over fresh each time.

So every message is an independent prediction, with deliberate randomness in it. It isn't fetching a stored row like a database. It's re-deciding.

That also explains why long conversations drift. Your instruction from message three is now buried under forty messages of other text. It didn't forget — it's diluted. Starting a fresh chat and pasting your brief back in usually beats continuing.

Now here is the consequence, and it's the most important sentence in today's lesson:

So what can you conclude? Only this, and it's enough:

These tools are not reliably consistent. Most of the time they will be right. Some of the time they won't. And the answer looks and sounds exactly the same in both cases — so you cannot tell which one you're holding without checking.

That's not a reason to distrust them. It's a reason to build one habit: check the things that matter, every time, regardless of which tool produced them or how well it did yesterday.

7. The tools — who's who

People say "AI" when they mean a specific product. These are the ones you'll hear about. All of them have a usable free tier.

Tool	Made by	Known for
ChatGPT	OpenAI	The all-rounder, and the most widely used. If someone says "I asked AI", it's usually this. Good at most things.
Claude	Anthropic	Long documents and careful writing. Tends to follow detailed instructions closely. Strong for editing and analysis.
Gemini	Google	Connected to Google's world — search, Docs, Gmail. Generous free tier.
Grok	xAI	Built into X/Twitter. Useful when you need what people are saying right now.
Perplexity	Perplexity	Search-first. Answers come with sources attached, which makes it a different kind of tool from the other four — closer to research than to writing.

Under the hood these are all doing the same next-piece-of-text prediction. They differ in what they were trained on, how they've been tuned, and what they're connected to.

Read that table as a rough starting point for which to reach for first — not as a description of how each one will behave on your particular question. Given section 4, you already know why: behaviour varies between runs, and all of these update every few months anyway.

Two years in, the thing I'd most want you to hear about this list is that there is no single best AI for every task. People pick one, get comfortable, and assume it's best at everything. It isn't, and it changes every few months anyway.

The only way to develop a feel for this is to run the same task through several of them and compare. That's why today's setup asks you to try all five rather than picking a favourite — and it's why the back half of this program will be straightforward for you and hard for people who only ever learned one tool.

8. What it's genuinely good at

Everything above is a warning, so let me balance it. These tools are extraordinarily useful — at specific things. The pattern is simple and it's worth memorising:

Notice that the right-hand column isn't forbidden. You can absolutely ask those questions — you just treat every answer as a lead to check, never as an answer to use.

Tomorrow — Day 2: Prompts and context. Today you learned what the machine is. Tomorrow you learn how to talk to it — and why the "magic prompt" people chase doesn't exist, but something much more useful does.