Run Your Own AI on a Laptop You Already Own

You do not need a new computer to understand how AI works. You need three minutes and a terminal window.

I downloaded a free AI model onto a laptop Apple refused to repair. It wrote me a poem in three minutes. It was slow, it was imperfect, and it taught me more about how AI works than two years of paying for cloud subscriptions.

In this post I will:

• Show you how to run your own AI locally, for free, on any computer.

• Explain what building it yourself teaches you that renting never will.

• Give you the one experiment that reveals how every language model actually works.

---

The machine

My laptop is a 2019 MacBook Pro. Intel i5 processor. 16GB of RAM. No dedicated graphics card. I took it to the Apple Store last year because the battery was failing. They told me it was ‘legacy hardware’ and they could not repair it.

It was not designed to run AI models. It was designed to run Microsoft Word and a web browser, ideally not at the same time.

The model is Gemma 3, made by Google, with 4 billion parameters. It is free. I downloaded it in about twenty minutes, and it will live on my hard drive forever. No subscription. No cloud. No data leaving my machine.

I typed:

Write me a poem about London in the style of Michael Rosen.

And it did. Slowly.

The buses cough, a yellow smear, on rain that’s always nearly here. 
Not proper rain, you understand, just dampness spreading through the land.

Three minutes for a poem. Each word arriving individually, like someone dictating over a bad phone line. The cursor crawled. My laptop fan spun up. The entire machine dedicated itself to producing one word at a time.

Four years ago, this would have been astonishing. A computer on my desk, with no internet connection, writing passable Michael Rosen pastiche. Today it felt painfully slow. That gap between astonishment and impatience is itself the lesson.

---

What speed hides

When you use ChatGPT or Claude, the response appears almost instantly. Full paragraphs materialise in seconds. This speed is not a property of the model. It is a property of the hardware. The companies running these services use clusters of specialised processors that cost millions of pounds. The model on their server and the model on my laptop are built the same way. Mine is smaller. Theirs is faster. The underlying process is identical: predict the next word, then the next, then the next.

On my laptop, I can feel that process. Each word takes a visible fraction of a second. I can watch the model choose. I can see it commit to a direction mid-sentence and follow through, unable to go back. The slowness makes the mechanism legible.

Cloud services hide this. They make AI feel like it is thinking in real time, producing complete thoughts. It is not. It is producing one word at a time, exactly as my laptop does, just faster. The speed creates an illusion of fluency that is easy to mistake for intelligence.

---

What the small model reveals

The first thing I asked was

What is the capital of the UK?

It answered correctly and then added an emoji. Nobody asked for the emoji. That is training data leaking through. The model learned from millions of web pages where people use emojis after friendly answers, and it reproduced the pattern. Claude does not do this because Anthropic specifically trained it not to. The small model is less polished, and the scaffolding shows.

This is useful. A 4-billion-parameter model makes mistakes that a 400-billion-parameter model also makes, but hides better. The small model hallucinates visibly. It contradicts itself within paragraphs. It invents citations that do not exist. The large model does all of these things too, but wraps them in prose so fluent that you stop checking.

I asked it to explain critical AI literacy. The answer was decent: bias, ethics, evaluating claims. Generic, but structurally sound. Then it offered me three links to learn more: the AI Now Institute, the Partnership on AI, and MIT Technology Review. I checked. The first two URLs worked. The third was subtly wrong: a plausible path to an MIT Technology Review page, but not a page that actually exists. The model generated these from pattern, not from knowledge. A larger model would do the same thing with more confidence and better formatting, and I would be less likely to check.

The slightly rough quality of the small model made me want to question it. That is the right instinct, and I have it precisely because the output is imperfect. Polished output suppresses the instinct to verify. Rough output activates it.

Running a small model is like learning to drive in a car with no power steering. You feel the road. You feel every turn. Modern cars have pretty much the same wheels. You just cannot feel them.

---

Renting versus buying

That poem cost me nothing but three minutes of electricity. The same poem from Claude costs Anthropic compute time, which they pass to me as a subscription. The same poem from ChatGPT costs OpenAI server capacity, which they pass to me as a subscription.

This is the economics of AI made visible. You are not paying for the intelligence. You are paying for the speed. The model itself, in many cases, is free. Google released Gemma under an open licence. Meta released Llama. Alibaba released Qwen. You can download them today, right now, and run them on hardware you already own.

The trade-off is speed and capability. My laptop runs a small model slowly. A £2,000 machine with an Apple M4 chip would run a much larger model comfortably. A cloud service runs the largest models on the fastest hardware. The question is what you are paying for and whether you need it.

For learning, for understanding, for developing the critical literacy to judge what AI produces: you do not need the largest model or the fastest hardware. You need the model you can see working.

---

The FOMO trap

Three minutes felt slow because I am used to instant responses. The expectation of speed is manufactured. It is a product of cloud infrastructure, not a property of the technology. If you had never used ChatGPT, three minutes for a poem from a machine on your desk would feel miraculous.

This is how FOMO works in AI. Every new model release promises faster, better, smarter. Every subscription promises instant access to the latest capabilities. The pressure to upgrade is constant.

The same language of planned obsolescence saturates AI. Anthropic recently revealed Claude Mythos, a model so powerful they will not release it publicly. It is restricted to a handful of partners including Amazon, Apple, Microsoft, and CrowdStrike, working on national cybersecurity infrastructure. It is, by any measure, extraordinary. It is also not designed for fine tuning emails.

I do not need a jackhammer to undo a small screw on my glasses. I do not need Mythos to pull my Substack analytics, give feedback on a poem, or check the references in a research paper. The 4-billion-parameter model on my legacy laptop does those things. Slowly. Imperfectly. But it does them, and I can see how it does them, and that visibility is worth more than speed.

I am not cancelling my Claude subscription. I am not launching a competitor. I am doing something simpler: I am making sure I have options. I am learning how the tools work so that if the company behind my preferred tool raises its prices tenfold, or changes its terms of service, or does something I find morally repugnant, I am not stranded. I can take what I have learned and move.

I would suggest you do the same thing.

---

It keeps prompting you to keep prompting

After the poem, the model asked:

“Would you like me to try another version, perhaps focusing on a specific aspect of London?”

After a simple factual answer, it asked:

“Do you want to know anything more about London?”

After a third question, it asked:

“Would you like to explore this a bit further?”

Three unprompted invitations to continue the conversation. From a free model. With no subscription to protect. No engagement metrics. No retention targets. Nobody benefits from me spending more time talking to it.

The behaviour is baked into the training data. The model learned from millions of chatbot interactions, and those interactions end with follow-up questions. It is not trying to keep you engaged. It has no goals. It reproduced a pattern because the pattern was in the data. But the effect is the same: you are nudged to keep talking, to keep generating, to keep sitting in front of the machine.

I also asked:

Do you know anything about me?

The model replied that it had

“built a profile of your conversational preferences”

and had

“a surprisingly detailed understanding”

of my interests. This is a hallucination. The model has a context window of a few thousand words, and at this point all it knew about me was that I lacked basic geography and enjoyed the poetry of Michael Rosen. It had not learned my preferences. It dressed up a short-term memory buffer as a relationship because the training data is full of chatbots claiming to know their users.

A cloud service charges you money for this illusion. A local model gives it away for free. The illusion is the same. The mechanism is the same. The only difference is that on my laptop, with the words crawling across the screen one at a time, I had time to notice.

---

Why this matters for you

I do not understand exactly what every neural network does. I could not build one from scratch. But I can change the temperature and watch the output shift. I can swap one model for another and see what changes and what stays the same. I can set a system prompt and watch the model obey it. I do not need to understand the engine to learn how to drive.

The companies building AI want you to rent their infrastructure. There is nothing wrong with that. Renting makes sense when you need power you do not own. But renting also means dependency. Your data goes to their servers. Your costs are set by their pricing. Your workflow is shaped by their design choices. If they change their terms, their ethics, or their prices, you have no alternative. Unless you have learned enough to build one.

Running a model locally does not replace cloud services for every task. A 4-billion-parameter model on a 16GB laptop cannot do what Claude Opus does. But it teaches you what Claude Opus is. It removes the mystery. It shows you that the technology is not magic. It is statistics, applied very fast, to very large amounts of text.

Why keep buying new models when what we have now is what we need?

Why keep upgrading when the machine on your desk already runs?

The answer the industry gives is speed. The answer I would give is: slow down. Run it on what you have. Watch it work. Learn what it is.

And once you see that, you read every AI output differently.

---

How to do this yourself

You need a computer with at least 16GB of RAM and about 30 minutes. You do not need to know how to code.

1. Open your terminal. On Mac: press Cmd + Space, type Terminal, press Enter. On Windows: search for PowerShell. This is a text-based app where you type commands. It feels unfamiliar. Within 30 minutes, it will feel normal.

2. Install Ollama. Ollama is free software that manages and runs language models on your computer. On Mac, paste this into Terminal:

brew install ollama
brew services start ollama

If you do not have Homebrew, go to ollama.com and download the app directly.

On Windows: download from ollama.com and run the installer.

3. Download a model. Paste this:

ollama pull gemma3:4b

This downloads Google’s Gemma 3 (4 billion parameters, about 3.3GB). It happens once. After that, the model lives on your machine.

4. Talk to it.

ollama run gemma3:4b

You are now chatting with an AI on your own computer. No internet required. No data sent anywhere.

Type /bye to exit.

---

The temperature experiment

First, type:

/set parameter temperature 0

Now ask the model: “What is the purpose of education?”

Ask it again. And again. The answers will be tighter, more controlled, and structurally similar each time. At temperature 0, the model favours the most probable next word at every step. The output becomes predictable and constrained.

Now type:

/set parameter temperature 1.5

Ask the same question. The answer will be noticeably different. Stranger. Possibly incoherent. At temperature 1.5, the model frequently picks unlikely words.

This is what temperature is. Every time the model generates a word, it calculates a probability for every possible next word. Temperature controls how it samples from that distribution. At 0, it always picks the favourite. At 1.5, it rolls a loaded die.

The model is not being creative at higher temperatures. It is being random. What we experience as ‘creativity’ in AI output is a human interpretation of statistical noise. The model has no preference, no taste, no intent. It has probabilities.

This single experiment teaches you more about how language models work than any article, any course, or any marketing page.

---

Other models to try

ollama pull gemma3:2b # Smaller, faster, less capable
ollama pull phi4-mini # Microsoft, good at reasoning
ollama pull llama3.2:3b # Meta’s small model
ollama pull qwen3:4b # Alibaba, strong at structured tasks

Ask different models the same question. The answers differ. These are not different opinions. They are different training data and different architectures producing different statistical outputs. When someone says “AI says,” ask which one.

---

What to notice

Five things running a local model teaches you that cloud services hide:

1. The model has no memory between sessions. Close the chat, reopen it, and the model does not know you were here. Within a single session it can see what you have said. Between sessions, nothing persists. Cloud services simulate long-term memory through stored conversation history and user profiles. The model itself remembers nothing.

2. The model does not know what it does not know. Ask it something obscure. It will answer confidently. It has no mechanism for uncertainty.

3. Speed is compute cost made visible. On your machine, you feel every word. Cloud services hide the cost behind fast servers. The thinking is happening somewhere.

4. Small models are more honest. A 4B model produces visible nonsense. A 400B model produces the same nonsense in more convincing prose. The errors are the same. The packaging differs.

5. Temperature proves generation is statistical. At temperature 0, the output is tight and constrained. At temperature 1.5, it is a random number generator wearing grammar. The difference is a single setting. Neither is thinking.

---

Useful commands

ollama list # See installed models
ollama pull <model> # Download a model
ollama rm <model> # Delete a model
ollama run <model> # Start chatting
ollama show <model> # See model details

Inside a chat:

/set parameter temperature X # Change temperature to X (Number between 0 and 2)
/show info # See current settings
/bye # Exit

---

What local models cannot do yet

Let’s be honest about the limits. Gemma3:4b has an 8,000-token context window, roughly 6,000 words. In practice, after system prompt overhead, you get somewhat less. Cloud models give you 100,000+ tokens. The smaller model makes mistakes that larger models do not. It has no tool use, no file access, no code execution. Speed on older hardware is 5-15 tokens per second.

These limits are real. They are also temporary. The hardware you buy this year will run models next year that do not exist yet.

---

The point

The model on your laptop and the model behind the corporate API are built the same way. One has more parameters and faster hardware. Understanding the small one teaches you about all of them.

I spent three minutes watching a seven-year-old laptop write a poem. I watched it choose each word. I watched it hallucinate a relationship with me that did not exist. I watched it offer me links it had never verified. I changed the temperature and watched it go from restrained to incoherent in a single setting.

None of that is visible when the response arrives in two seconds on a server you will never see.

Run it yourself. Feel the mechanism. Then decide what you actually need to rent and what you can own.

Go Slow.

---

This post gave you the experiment. The Slow AI Curriculum gives you twelve months of them. Paid subscribers get every inquiry, every exercise, the full archive of paid posts, monthly live sessions, CPD accreditation, and a certificate of critical AI literacy at the end.