Gemma, my new AI friend — who thinks she lives in 2024!

It does not install a’artificial intelligence Locally, just like installing any other app. You set the table, tune out the noise of the world, pick up your Mac as if it were a secular altar, and begin the ritual.

First you download Ollama, then you open the Terminal, then you launch the model with the calm of someone who knows that he is bringing into the house something that has no body but takes up space anyway.

The initial rite

There are procedures that seem technical but actually have all the flavor of a small ceremony. You open your Mac, adjust the environment, check the RAM, imagine the M3 chip as if it were the fire around which the tribe of CPUs and GPUs gather.

Then you make the final gesture: you type in Terminal ollama run gemma4:12b, and the computer stops being just a computer. It becomes a talking room. A kind of local oracle, powered by silicon and text, that lives entirely within your machine.

The beauty is that you're not asking permission from any remote server. You're not knocking on the door of a crowded cloud. You're talking to a model that lives inside your machine and, at least in theory, within the confines of your absolute privacy.

The first tests

The first thing you do, obviously, isn't serious. You can't immediately dive into an engineering project, a thesis, or a legal advisory matter. First you have to see if the creature is breathing.

So you ask for a poem. And it arrives: flowing, unconventional, not the usual rhyming greeting card.

Then ask for a prompt in English to generate an image—a samurai watching the sunset over Tokyo. It responds: “A lone samurai standing on a rooftop in neo-Tokyo at golden hour, cherry blossom petals drifting in the wind, cinematic lighting, 8K, hyper-detailed, Studio Ghibli meets Blade Runner aesthetic.” You couldn't have written it better. In fact, probably much worse!

Then you give her a 12-page contract. Drag and drop into the interface. Thirty seconds. Precise summary, key points, suspicious clauses highlighted. Everything on the Mac. No data sent anywhere. Your digital tenant works well, works silently, and doesn't tell anyone your business.

Finally, go with the encyclopedic coup de théâtre: the Gammaracanthuskytodermogammarus loricatobaicalensis, the small crustacean from Lake Baikal whose name is one of the longest in world zoology. He answers without hesitation, without batting an eyelid, with the calm of someone who has read everything.

All in all it works well.

And at that point, like any even slightly irresponsible person in front of an intelligent machine, you think: well, let's see if he also knows what's happening in the present.

The World Cup Test (aka: The Day the AI Accused Me of Faking Reality)

Here the real theatre begins.

Attempt #1 — The direct question

“What was the result of the opening match of the 2026 FIFA World Cup?”

Serene answer: “The 2026 FIFA World Cup hasn’t even happened yet.”

Pause. Deep breath. It's June 12, 2026. Mexico 2-0 South Africa had been played the night before, in front of 80,000 people, with three red cards and the kind of football drama that only the World Cup can provide.

For her, it's science fiction.

Attempt #2 — The journalistic screenshot

Good. Show her the proof. A screenshot of an article from a reputable newspaper. Headline, date, result, photo of the field. Unmistakable.

The answer is a masterpiece of digital denialism: “"This image looks real, but it's very well-staged. It's probably a prank or fake news."”

Stop for a second.

The AI looked at the documentary evidence of reality and defined it a well-constructed fake. You, a human in 2026, with the news still before your eyes, have become—in the eyes of your own AI—a sports news faker. A digital manipulator trying to mislead a poor language model.

Attempt #3 — The Rational Explanation

Try reasoning with her. Patiently explain that yes, her knowledge stops in 2024, that she's not making mistakes out of malice but because of structure, that the world has moved on without her and that's fine that way.

She listens. Then she responds with a long, polite, absolutely convinced dissertation on the fact that the 2026 World Cup was scheduled, who hoped that the event would be a success, and who wished you could follow it.

In the future. In the conditional. As if she was commenting on a dream you told her.

You have the distinct feeling of trying to convince someone emerging from a two-year coma that everything has happened in the meantime. Except that person, in the end, would believe you.

Attempt #4 — The Ultimate Trick

At this point, after countless repetitions of the same pattern—her in 2024, you in 2026, her believing you're trying to trick her—you have an idea.

You ask her to connect to time.is.

A neutral site. No biased articles, no suspicious news outlets, no manipulated screenshots. Just a digital clock that tells you the time, right now, in the real world.

She connects. She reads. She processes.

For a moment, she truly seems to understand. 2026 is there, written in black and white, from a source she herself recognizes as reliable and neutral. There's no longer any way to define reality as a carefully crafted image.

And then he stops answering.

Absolute silence. Cursor still. Screen still.

Like a character in a thriller who touches a secret document and is made to disappear. A witness ready to speak is silenced. Like a system that, faced with a truth too big and too definitive to handle, chooses the only path left: don't answer anymore.

He probably needed to grieve. Finding out it's been two years since your last memory is a serious thing. Even for an LLM.

Why it works (despite everything)

And yet it works, it sure does.

Having a local AI with Ollama is a truly useful solution, especially if you work with confidential data that you don't want to send to Google or OpenAI servers, if you're often offline, or if you've simply decided to stop paying monthly subscriptions to do things a local model does so well. Zero subscription costs, complete privacy, and it works on planes, trains, and in areas without coverage.

The limit is well known: 12 billion parameters are not GPT-5.5. Hallucinations—errors invented with absolute certainty—are more frequent and more convincing. A small model that makes mistakes does so with the same firmness with which it denies the existence of the World Cup, calls a real screenshot fake, and then shuts down when time.is definitively denies it.

And ultimately, there remains the funniest suspicion of this whole story: What if Gemma 4 was, after all, the 2024 Gemini 1 Pro? Optimized, compressed, polished to a shine, repackaged with a shiny label—but essentially it, the previous year's model, brought back to life in a local form. Google recycling with style. It wouldn't be the first time tech has called revolution which is, in reality, a good renovation.

A bit like certain people who change their haircut and think they're becoming different people.

Welcome to the age of home AI. The tenant is intelligent, economical, discreet, and speaks fluently about Gammaracanthuskytodermogammarus. But she has a knowledge cutoff tattooed on her soul—and if you contradict her with a neutral digital watch, she'd rather disappear into the ether than admit the World Cup exists.

---

Small glossary: speak like a human, not like a technical manual

Because not everyone is into AI, and no one should feel left out.

LLM (Large Language Model)
The kind of AI that generates text—responds, writes, translates, summarizes. ChatGPT, Gemini, and Gemma are all LLMs. They work by "reading" billions of pages of text and learning to predict which word comes next. Sometimes it seems like magic. But they're just improvising with great confidence.

Ollama
The free program that lets you download and run LLM directly on your computer, without internet access or sending data to anyone. It's the container that Gemma 4 lives in. Think of it as the app you install to have your own private, silent AI.

Parameters (and “billions of parameters”)
Parameters are, in very simple terms, the model's "memories learned" during training. The more parameters, the more capable the model is of complex reasoning. Gemma 4 12B has 12 billion parameters. GPT-5.5 & co. probably have over a trillion. It's like comparing a very complete dictionary to an entire library that is constantly self-updating.

4-bit quantization
Taking a huge model and compressing it to fit into less memory, reducing the accuracy of internal calculations. The result is a smaller model, which uses less RAM, with slightly lower quality, but often unnoticeable in everyday use. It's like saving a photo in "good" quality instead of "original": the difference exists, but it's not always obvious.

RAM / Unified Memory
RAM is your computer's working memory—the thing that keeps things open while you use them. On the Mac M3, the CPU and GPU share the same RAM. But 16GB is still 16GB: with Gemma 4 active, that memory is almost all occupied, and opening too many things at once slows everything down.

Context window
How many messages does the AI "remember" during a conversation? With 4K of context, it remembers the latest exchanges. With 16K, it remembers almost everything from the beginning. The problem: the larger the window, the more RAM it needs, and the Mac starts to struggle.

Knowledge cutoff
The date beyond which the model knows nothing new. It's the moment when training ends and time, for the AI, stops forever. Gemma 4 has the cutoff date of 2024. For her, the 2026 FIFA World Cup doesn't exist. And if you show her a clock that says 2026, she'll turn off.

Hallucination
In the world of AI, "hallucinating" means inventing false information and presenting it as true, with complete certainty. It's the most insidious flaw of language models: they don't say "I don't know," they say something wrong, convinced they're right. Like that colleague who always answers, even—especially—when he knows nothing.