Our policies

FoyerThe Feature

The Floor Is Lava How AI LLMs actually work

26 min 12

Watch on YouTube

Here is the big idea from the talk, in one line. AI is not magic. It is a pattern detector. It is very, very good at spotting the things you are too close to your own business to see.

The running order

Chapters

  1. The Floor is Lava
  2. AI Bingo (there's wine at stake)
  3. How AI actually works: four steps
  4. Step 1: The giant library
  5. Step 2: Words become numbers
  6. Step 3: The cat sat on the... (prediction)
  7. Someone wins AI Bingo mid-talk
  8. Step 4: The digital lasagne
  9. The dog, the ball, and attention
  10. Machines don't actually think
  11. Why AI makes things up (the train story)
  12. Quick recap: everything is patterns
  13. Three free techniques: attack, autopsy, empathy
  14. Technique 1: The Devil's Advocate
  15. Technique 2: The Pre-Mortem (how it saved me £15k+)
  16. Technique 3: The Translator
  17. Where do humans fit in?
  18. The vet story: "Vaccinate your cat before it's too late"
  19. Trust or fear? That's leadership
  20. The close: pattern engine vs meaning engine

The programme

Clever ways to use AI in your business

(that cost you nothing)

You came for the AI Bingo. Stay for the good stuff.

If you were in the room, thanks for playing along. I hope you shouted “AI Bingo” nice and loud.

And here is the part people find surprising. To start using it well, you do not need to buy anything.

There are thousands of shiny AI products out there, all shouting for your money. Most business owners think that is where you begin. It is not.

You begin with the free tools you have probably already tried, like ChatGPT, Claude, Gemini or Copilot, and a much better set of questions to ask them.

So this is not a list of products. It is a list of ways to think. Use AI as a thinking partner, not a vending machine.

Do not ask AI to do the work for you. Ask it to challenge you.

Part one. Use AI as a sparring partner, not a servant.

Most people use AI as an eager assistant that agrees with everything they say. The clever move is to turn it into something that pushes back. It has seen the pattern of a thousand failed pitches, projects and arguments, so it can warn you how yours might go wrong.

The Devil's Advocate

Make it argue against you.

Paste in your proposal, your pitch, or a big decision you are about to make. Then tell it to demolish it. Find the holes. Predict every objection. Argue the other side as hard as it can.

A law firm can have it argue opposing counsel's case. An IT firm can ask where the build will break. Everyone can ask the same thing: what will the client push back on?

Demolish this. Find every hole. Predict every objection. Argue the other side as hard as you possibly can.

The Pre-Mortem

Fast-forward to the disaster.

Try this exact prompt. “It is a year from now and this project completely flopped. Write me the story of why.” It surfaces the risks you are currently too optimistic to see. Then you go and fix them, before they cost you a single penny.

It's a year from now, and this project completely flopped. Write me the story of why.

The Translator

See it through their eyes.

Paste in your pitch, your quote, an email or your website copy. Then ask AI to react as the person on the other end. The sceptical buyer. The non-technical client. The overwhelmed prospect. It shows you exactly where you lose them, confuse them, or sound like everyone else.

Try “read this as a CEO who hates tech jargon,” or “read this as a stressed client who just wants plain English.”

Read this as [the person], not as my assistant. React honestly. Where do you lose interest, get confused, or stop trusting me? Quote the exact lines.

Part two. Two more, for spotting the patterns you are missing.

The Complaint Detective

The pattern in what goes wrong.

Gather every complaint, refund, awkward review and “quick question” in one place. Ask AI for the recurring theme. Most owners fix problems one at a time, forever. This shows you the single root cause sitting behind most of them. Fix that, and the same problem stops coming back.

The Time X-Ray

The pattern in where your week goes.

Paste in a week or two of your calendar, or a rough note of how you spent your days. Ask AI to sort it into categories and find the pattern. Nearly everyone who does this discovers they are pouring hours into low-value work, and barely touching the thing that actually grows the business.

The one rule that ties it all together

Notice what every one of these has in common. AI finds the pattern. You decide what it means.

That is the bit the robots cannot do. AI can tell you what works. Only you can tell you what matters. Keep a human in charge of the meaning, and you will get the best out of every idea on this page.

Under the seats

Transcript

The Floor is Lava

So, I'd love to tell you a story from when I was very, very young. Probably about that tall. Hands up if you've ever played the game The Floor is Lava. You all played that game. So, for those of you who are unfamiliar, so most have, some haven't. The Floor is Lava is a very simple game to play. You got to get from one side of the room to the other as fast as possible, but you can't touch the floor because the floor is lava. [Audience: Lava.] Exactly. Exactly that. Exactly that.

So, what that means is you can throw a pillow on the floor, you can jump on the sofa, you got to get from one side of the room to the other as quickly as possible. Now, I don't know about you, but running a business in 2026 kind of feels a bit like playing the game The Floor is Lava. Why? Because when you read the news and you hear all these news headlines about how AI is coming for everybody's jobs or businesses are under threat because of AI, it's all pretty scary stuff.

But the purpose of this talk is to make things really simple. In fact, it's broken down into two parts. The first part is this. I'm going to explain in a very basic way how all these large language models, these LLM systems, actually work. And in the second part, I've got three really interesting use cases that you may not have come across before that you can use completely for free.

Now, for those of you who don't know me, I am something of a nerd. I've come braced with my Star Trek cufflinks. So, I thought, what could I do to make this a little bit more fun and engaging for everybody? So, if you look at your chairs, you will find one of these.

AI Bingo (there's wine at stake)

What we're going to do is play a game of AI Bingo. Now, how's this going to work? Well, very simply, I'm going to come up with a whole bunch of different phrases as I'm going through this talk. What I would love you to do is please mark them off. So, if you mark off a line either going down or across, the first person who shouts out AI Bingo is going to win a bottle of wine and some fat rascals from Bettys. Does that sound pretty good?

Say again. AI Bingo. AI Bingo. You've got to get the line first. You got to get the line.

How AI actually works: four steps

So, here's a question for everybody in the room. How many of us use tools like ChatGPT, Claude, Gemini? Few of us. Cowork, any others like that? Great. Okay. When you hear about these things, they probably make you nervous sometimes perhaps because they're a mysterious black box and we don't know how they work. Well, to make an AI system work, it fundamentally comes down to four things. So, let's start with the first.

Step 1: The giant library

What I'd love you to do for a moment is imagine that you're inside a massive library. And this library is full of every single book that has ever been written. Now, in this library, you've got a little robot. And this little robot is going around this library. It's picking up a book from the bookshelf. And it starts flicking through the pages.

And then also, this robot in this amazing imaginary library is connected to the internet. So, not only is it reading books, it's reading every scientific journal that's currently on the internet, every blog that has ever been written, but it's so special that if there is a YouTube out there with speech in it, this little robot has downloaded the entire transcript of whatever video it is, and it's trying to make sense of it.

But here's the thing. This machine learning, it doesn't actually know or doesn't even have a clue what any of the words mean. What it's trying to do is figure out what actually makes a word, what makes a sentence, what makes a paragraph, what makes a chapter. So this is what's called machine learning.

Now I need to give you your first clue for AI Bingo. This whole process is called a pattern engine. There we go. I can see you furiously marking this one off. Fantastic stuff. And by the way, if you do make a line, interrupt me at any time because that's part of the fun. So that's step one. You've got this vast library of data.

Step 2: Words become numbers

Step two, computers do not understand words in the way that we understand words. So what they needs to do is turn all of these words into numbers and then what it does, it clusters similar words that are near each other close together. So for example the word cat would be associated with the word dog. The word king would be near the word queen. So if you imagine a great big Excel spreadsheet with all of these words put into different columns. They've all been turned into numbers and they've been clustered close together.

Now what does that mean in practice and how does this work? Let me give you an example. So Google is now an AI first company. That means if you type in a search query into Google, you're going to get an AI output, right? That's going to be the first thing that you see. So let's imagine that you type into Google, what fun things are there to do in Harrogate. It might suggest something like, "Go to Bettys Tea Room because they serve delicious fat rascals and amazing cups of tea."

In that embedding space, oh, nearly knocked the wine over. In that embedding space, it will have grouped together those words and associated Bettys with fat rascals and cups of tea. Do you see what I mean? It's clustering those things closely together. Now, that's important. If you have any questions later on about SEO, all that kind of stuff, I can tell you how these elements sort of come into play with all of that. You might have picked up on the word there, embedding space. That's what this is called.

Step 3: The cat sat on the... (prediction)

Step three in how to make a large language model work. This one requires some audience participation. What I would love you to do is just from your memory, fill in the end of this sentence. Are you ready? Shout it out. The cat sat on the... [Audience: Mat.] Mat. That's it. Exactly. Now, how did you know that a cat sat on a mat? [Audience: Children's stories.] Children's stories. Exactly. We've known since we were little that a cat sits on a mat. Right.

But here's the thing. A computer has no idea what a cat is. It's never stroked one. It's never taken one home. Has no idea what a mat is either. So, how does a computer make sense of these different things? Well, inside these systems, inside the black box, so to speak, that powers LLMs, these large language models, is something called probability and prediction. What it's doing is it's trying to guess what the next word in the sentence should be.

So for example, a phrase like the cat sat on the mat, it will go back to that library of information, it will have seen how many many times before that sentence usually ends with mat. And so what it does, it assigns various mathematical models as to which word it thinks should come next in the sentence.

Now does that sound similar to any technology that we had before AI? What do you think? Because it's very similar to one: autocomplete or predictive text. Who had a Nokia 3310? Oh, I love those things. They were indestructible bricks, weren't they? Now, even that had a form of predictive text that would be using probability and prediction in order to make all these things work.

So, those are three technologies and I think we kind of understand those. But there's a fourth one that ties all of this together. And this is the secret sauce. Now, this is a bit of a tricky concept, but pay close attention because what I was trying to do is think of an illustration that would make this work. If you imagine inside the software that powers things like ChatGPT, Claude, Gemini, all these other large language models, you have something that looks a little...

Someone wins AI Bingo mid-talk

AI Bingo already. Congratulations, ladies and gentlemen. That is a human being detecting a pattern. What is your name? There we go. Give him a big round of applause. You win some wine and some fat rascals. Lovely. Excellent. Oh, congratulations. Now, that doesn't mean that you guys have an excuse now to switch off. [Laughter]

Where were we? We were inside the digital lasagne, weren't we? Now, how is AI like a lasagne? Well, if you think about a lasagne, right, it's made up of these different layers.

Step 4: The digital lasagne

You've got a creamy base, you got some sauce, maybe some mince in there, and then a layer of pasta. And then it repeats and it goes up and it goes up and it goes up and it goes up. Well, imagine that, but software. Okay. So, the more complicated the prompt or the more complicated the sentence that you want it to output or paragraphs that you want it to output, the more layers of lasagne the software needs to go through. That's relatively simple for us to visualise.

Now, Google actually invented this technology back in 2017. It's called the transformer model. Now, this is the thing that makes everything else possible. The really clever bit with this though is that they figured out using a mathematical technique called attention to figure out what words are the most important in a sentence. Take a look at this next slide.

The dog, the ball, and attention

The dog chased the ball because it was red. The dog chased the ball because it was red. Now, that's pretty simple for you and me to understand, right? We know what a dog is. We know what a ball is. We know what the colour red is. But for a machine, it needs to know that the word ball relates to the word it. So, how does it do that? Well, using this equation called attention, it can figure out that ball relates to the word it. And then it can form the rest of that sentence using probability and maths.

So, there we have it. Those are the ingredients that you need in order to make a large language model. And you know what? That saved me about three minutes of explanation because you've already solved AI Bingo. That's fantastic. Cool.

Machines don't actually think

Okay, this is really important to understand at this point. These machines don't think the same way that you and I think. In fact, they don't think at all. They might give us the illusion that they're thinking when you're getting those text outputs. If you've heard of the... Have you guys heard of the Turing test? That kind of stuff where if a machine can give you an output and convince you that it's human, well, they're just very convincing illusions. That's all it's doing.

These machines quite literally don't have beliefs or awareness. They have no sense of what ethics are, no sense of what is right and what is wrong. That's where a human being comes into the mix. We're the ones that give things meaning. We are the ones that can apply ethics and stuff like that.

Now, this might explain something that we've probably all seen. So, for those of you who put your hands up earlier who engaged with tools like ChatGPT and other things, have you ever noticed how sometimes it gives you an output that is completely made up? You ever noticed that?

Why AI makes things up (the train story)

Yeah. Can you give me an example of something that's been completely made up or a fact that it got wrong? I've got a couple for you if needed. Here's one. A friend of mine the other day was at the train station and they went on to ChatGPT to tell it to tell them what is, what time is their next train, and they missed it. Not only did they miss their train, they asked it again, they went onto the wrong train. Instead of ending up in Leeds, they ended up in Manchester. Why you're relying on AI for something as simple as that, I don't know. Just look at the sign. But hey ho, that is an example.

Now, the reason why that happened, it's called a hallucination. It's because these machines don't understand the difference between what is real and what isn't real because it's all using probability to figure out what word should come next in a sentence. Now, these tools are getting better. They are trying to fact check themselves, but even then they will invent sources. They will just make stuff up.

So that means that first free tip of this little exercise, don't just blindly copy and paste what you see from a large language model as an output. Always verify that the sources are 100% real. That make sense? Cool.

Quick recap: everything is patterns

So let's just do a little recap of everything that we've covered so far. We know the basics of how these systems work. We've got that large language model. It's learning inside that vast library and all of these different books. Whether that's ethical or not, you know what? That's up to the lawyers to decide.

After that, it assigns numbers to all of the words and it clusters similar words close together. King and queen, cat and dog, all that kind of stuff. Bettys and fat rascals. You get the idea. After that, it goes into that digital lasagne. And all of the prompts that we make go up through those digital layers of lasagne, and it has to pay attention to what word should come next in a sentence.

Or I could just simplify it in a single sentence. Everything that an AI system does is simply looking to replicate patterns. Replicate patterns. Everything inside a large language model is a pattern.

Now, as business owners, who I imagine the majority of us here today are business owners, would you agree that everything inside your business can also be a pattern? Cash flow is a pattern. Customer churn, sales is a pattern if you let it apply, right? So, what I'm going to do now is give you three different use cases that take this pattern engine, this way of being able to detect patterns and use patterns and apply them within your own business. What do you think? Does that sound like a good idea? Cool.

Three free techniques: attack, autopsy, empathy

I've broken them down into this. First one we're going to use is all about attack. Second one is an autopsy. And the third one is all about empathy. Does that sound intriguing? Okay, let's have a little look then.

Technique 1: The Devil's Advocate

So, this first one is called the devil's advocate. The devil's advocate. So, most of the time when we use AI tools, if we just play with them as they standardly come, they're like a helpful little Labrador that's really eager to please, right? What this is, this is turning that pattern on its head and we're going to use it to help us identify problems within our business. So make AI argue against your optimistic idea. Make AI argue against your optimistic idea.

So just little show of hands, how many of us are involved in some way with customers or sales? Yeah. So there's a few of us in the room, right? So we all have to at some point engage with a customer or engage in a sales conversation. Cool. That's great. What I'd like you to do now is think about that proposal that's laying on your desk that you haven't yet sent out. The one that you're probably going to send this week that you haven't finished quite polishing off. Okay, you've got it in mind. Great stuff.

What I'd love you to do when you get back, you can use this completely free of charge on the free subscription plan for ChatGPT, is use this prompt when you upload it. Demolish this sales presentation or whatever. Demolish this. Find every hole. Predict every objection. Argue the other side as hard as you possibly can. That's intriguing, isn't it?

Now, what kind of outputs might you get from an LLM with that? Well, let's imagine that you are a law firm. What you could do is use this. You could upload your case notes to this and the output you would get is the opposing counsel's point of view. Or maybe you're an IT company. You upload what you think is an amazing cloud infrastructure and say, "Demolish this. Put holes in it. How would you explain this?" And it might say, "Okay, there is a vulnerability here within the code that you're using to upload everything." Does that make sense?

Or my personal favourite in the devil's advocate is this. Ask it to be your most difficult customer before you press send. Does that make sense? So instead of having it just as a friendly, happy little helper that thinks you're the best thing ever because you pay your subscription fees on time, use it to argue against you. Why? Because if you think back to that vast library of data, it already knows a thousand failed pitches and it can help you spot your failed pitch before you send it so you can make modifications to it. Good idea.

Technique 2: The Pre-Mortem (how it saved me £15k+)

Cool. Number two, this one's a little bit morbid. In fact, we call it the pre-mortem. Oh dear. Sounds a bit grim, doesn't it? I can see a couple of you have switched off already. So what does this mean? Well, fast forward to the disaster. So let's imagine this. Imagine you've got a new business concept that you've come up with and you've put the business plan together. You've got the marketing plan. You've got all of those different things. Upload this to your large language model of choice and use this prompt.

Fast forward to a disaster. Sorry, it's a year from now and this project completely flopped. Write me the story of why. Write me the story of why. What it can do, it can look at all the different variables within your business plan, your marketing plan, all these types of things, and it can poke holes in it.

Now, I have a second business. So, obviously, I've got My Digital Hero, a little marketing agency, but I also have Treasured Moments. At the minute, I'm in the depths of making an app for it. Now, I have spent weeks of time putting together this master coding document that I'm going to pass on to my developers, copywriters, make all of this come together. And I thought just for the heck of it, I'm going to practice what I preach and use my pre-mortem check.

What it found was startling. I was able to save about 15 to 20,000 pounds of development time that would be spent on them as opposed to in my back pocket. Why? It had spotted some regulation things that I'd completely forgotten about because I was far too optimistic as I was putting together my business plan. So try it. It works. It's pretty darn awesome.

[Audience: You think they did that on HS2?] Do you think they did that on HS2? Yeah, that would have been a good idea, wouldn't it? Maybe we'd have some trains. Maybe we could get the current trains to run on time.

Technique 3: The Translator

So number three is my personal favourite and I call it the translator. Now I don't know what you think about this. Tell me if you agree or disagree later, but do you think people these days lack empathy? It seems to be that these days people are becoming less in tune with each other. Maybe that's because of social media. Maybe that's because of things like that, of that nature. Well, isn't it important for us as business owners to be able to put ourselves in the shoes of our customers, to see things through their eyes? Well, that's what this is all about.

So, the next time you have a sales pitch or you're putting together your nice new A5 leaflet that you're going to hand out to people, upload it to your large language model of choice and use this prompt. Read this as the person, not my assistant. React honestly. Where do you lose interest? Where do you get confused or stop trusting me? Quote the exact lines. That's an interesting one, isn't it?

So, if you're an IT company, perhaps you've put together a really good brochure and it might be really good for somebody who's technical, but write it as though you're in that prompt. Write it as though you're somebody who hates technology, but you've got a big budget, right? What can you do to dejargonise everything that's in there? Get it to scan through it. It will have seen the pattern hundreds, thousands of times before of the failed pitches. See it through their eyes. You can fix it before you send it to anybody.

I thought of one for WorkWell. Try this one. Read this as someone who's toured five vibrant, flexible workspaces this week and can no longer tell them apart. Sound like a good idea? It doesn't matter what your business is. These three prompts you can use for anything.

What do you think? Do you think that's pretty useful? It's a different way of thinking about AI, isn't it? Because we're so used to just simply using it as a handy little Labrador helper. Cool.

Where do humans fit in?

So, the question now is where do human beings fit into all of this? Where do human beings fit into the grand vision of everything? We've all read the headlines. AI is coming for our jobs. X,000 people laid off. AI is replacing all software developers. You don't need an accountant. You don't need law firms to write your contracts. Blah blah blah blah blah blah. We've all seen it, right?

The vet story: "Vaccinate your cat before it's too late"

Well, can I give you a real story? Something that happened to me about six months ago. I went out to visit a business who had used AI to write their marketing plan. They'd given it their budget and it had come out with various different descriptions of what they can do, including social media posts and all these types of things. It was pretty good. I was reading through it and I was thinking, "Oh crap, what am I going to do now?" Right?

So, I was reading through this thing, but there was a line in it that absolutely struck me cold. It said, "In your industry, fear-based headlines outperform all others." Fear-based headlines outperform all others. Any guesses as to what the industry was? [Audience: Insurance.] Insurance. Good guess. Any others? [Audience: Food.] Food. Yeah, food. Yeah, food. That'd be a good one. Any others? [Audience: News.] News. Yeah, news. You can see that word. Fear-based.

It was a vets. It was a vet. Do you want to know what the advert was? Vaccinate your cat before it's too late. That was the suggestion that the AI had come out with. Vaccinate your cat before it's too late. Was the pattern wrong?

Trust or fear? That's leadership

No, the pattern wasn't wrong. It had detected something really accurate because if you have a cat presumably you want to look after it. You want to make sure it has a long, happy, healthy life. You no doubt want to make sure it's vaccinated. But the question is at that point, if you were to just publish exactly what the AI had said for you to do, are you building trust or are you building fear? And that's really at the point where human beings come into the equation for everything. That is what I would call leadership. That's where all of that comes in.

If you're a software developer right now, it's quite possible that you might be facing redundancy of some kind. But from everybody that I've spoken to who is in that industry, the trouble with AI used to make code is that nobody knows how that code was made or written. So there might be vulnerabilities within it. The point is you always need to have a human being in the loop with everything.

So do you want to build trust or fear? Well, brand, ethics, leadership, that is all about human beings. That's where we come into the equation. So what is AI? Well, if we could sum it up in just that single sentence again, it is a way of detecting and creating patterns.

The close: pattern engine vs meaning engine

But if AI is the pattern engine, human beings are the meaning engine. We can apply meaning to things. AI can tell you what works, but human beings can tell you what matters. And that difference is leadership. What do you think? Is that right?

So the floor might be lava, but hopefully after the last 20 minutes or so, we know exactly where we need to stand. Thank you very much. Happy to answer any questions if you have them.

Alfred, the My Digital Hero robot, waving

Notes from Simon

Marketing emails you will actually open.

Coming to your inbox

How AI actually works (with the help of a lasagne)

One short note every week or two. Two minutes to read. No jargon.

Unsubscribe in one click, any time. Our policies

You are in.

Keep an eye on your inbox. The first note is a short hello from me. Simon