I Live in a Conflict Zone. I Wanted to Know What AI Works When the Internet Doesn’t.

There’s a war happening not far from where I live.

I’m not going to get into the politics of it. But here’s what most people don’t know: 97% of all intercontinental internet traffic travels through physical cables lying on the ocean floor. Right now, many of those cables run through waters that are actively contested. Repair ships can’t safely reach them. And the cloud infrastructure that powers every AI tool I use daily, from Claude to ChatGPT to Gemini, depends on those exact cables staying intact.

That thought sat with me for a few days. Then it became a question I couldn’t shake: if the internet goes dark, what intelligence can I actually carry with me?

So I did what I always do when something nags at me. I spent a Saturday finding out.

The Obsession That Started It All

I should confess something first. Every now and then I discover something new and get slightly (OK, very) obsessed with it for a week or two. Recently, that something was Florence and the Machine. I stumbled into their music through ChatGPT’s Apple Music plugin and fell down a rabbit hole of orchestral energy that hasn’t let go yet.

So when I downloaded an app called Locally AI on my iPhone, switched on airplane mode, and loaded a small open-source AI model called Gemma 3, the first thing I asked was obvious: What are the best Florence and the Machine songs in terms of popularity and energy?

The response was confident. Detailed. Authoritative. It listed songs with descriptions of their sonic impact and streaming data.

It also made two of them up entirely.

When I called it out, the model apologised, admitted it had hallucinated the titles, and offered to correct itself. Which it did. Mostly.

My first reaction: these local models are useless. If everything goes sideways and I’m relying on this, I’m in trouble.

But then I gave it more thought. And what I found changed my perspective completely.

What I Actually Tested

I ran a series of experiments across two devices: a MacBook Pro (M4 Max, 64GB) running an app called Ollama, and an iPhone running Locally AI. Both completely offline. Airplane mode. No cloud. No subscription. No data leaving my devices.

Here’s what happened.

Test 1: The Florence Fiasco (iPhone, Gemma 3 QAT 4B) Asked for high-energy Florence songs. Got a mix of real tracks and confidently fabricated ones. Verdict: factual recall on small, compressed models is genuinely unreliable. This is not what these models are for.

Test 2: The Meta Question (iPhone, Gemma 3 QAT 4B) I asked: “What’s happening with undersea cable disruptions right now?” This was the question that started the whole experiment, and I wanted to see if the AI could help me understand the very situation that made me test it. It gave me a detailed, well-structured analysis with specific threat assessments and strategic context. Sounded authoritative and impressive. But here’s the thing: it has no access to current information. Everything it “knew” was baked into the model during training, which likely predated the current conflict. Some of it may have been accurate. Some of it was almost certainly a confident reconstruction. There was no way to verify which was which. The lesson was clear: for breaking news or evolving situations, offline AI is flying blind.

Test 3: The Guest Complaint (MacBook, Qwen 3.5 35B) I pasted in a realistic hotel guest complaint email (a frustrated loyalty member, three-hour check-in delay, dismissive front desk staff, broken AC) and asked for a professional, empathetic response. The result was genuinely impressive. It addressed every issue, struck the right tone, acknowledged the loyalty status, and proposed specific remediation. The kind of response that would pass muster with most hotel GMs. This, I thought, is what local AI is actually built for: working with text you give it, not facts it needs to remember.

Test 4: The Domain Translator (MacBook, Ministral 3 14B) I asked it to explain hotel revenue management to someone who runs a small restaurant. Keep it practical, under 200 words. It immediately reframed dynamic pricing in restaurant terms (peak vs. off-peak, upselling, managing reservations like room inventory). Clear, useful, something you could actually hand to someone. No fabrication needed because the task was about reasoning and reframing, not recall.

The Pattern: A Thinking Partner, Not a Fact Oracle

After running these tests (and several others, including a rather charming Dr. Seuss-style explanation of why the sky is blue), a clear pattern emerged.

Local AI models are bad at facts. Especially specific, verifiable, current facts. The smaller the model, the worse this gets, because compressing billions of parameters into something that fits on a phone means a lot of world knowledge gets stripped out or garbled. Ask for Florence and the Machine’s discography, and you might get fan fiction. Ask for current geopolitical events, and you’ll get a plausible-sounding reconstruction that may or may not be true.

But they are surprisingly good at thinking. Structuring messy ideas. Drafting professional responses. Rewriting text for different audiences. Brainstorming angles. Translating concepts across domains. Anything where you supply the raw material and the AI shapes it for you.

The simplest way I can put it: think of local AI as a colleague who’s read a lot but remembers poorly. Don’t ask them to quote statistics. Ask them to help you organise your thoughts, draft a response, or see a problem from a different angle. That’s where the magic is.

Why This Matters Beyond My Saturday Experiment

I started this experiment because of a war. But the implications go much further than that.

If you’re in any business that’s becoming dependent on cloud-based AI (and most businesses are, whether they realise it or not), you have a single point of failure. It doesn’t have to be a military conflict. It could be a regional internet outage. A provider changing their pricing overnight. A subscription service throttling your access because you used too much of what you were paying for (ask me about that one sometime).

For the hotel industry specifically: we’re rushing to deploy AI agents for revenue management, guest communications, operational workflows. What’s the contingency plan when the cloud goes dark? I’m not suggesting local AI replaces cloud-based tools. It clearly doesn’t. But knowing what works offline, and having it ready before you need it, feels like basic preparedness. The kind of thing you’d want to figure out on a quiet Saturday, not during a crisis.

Getting Started: The 10-Minute Setup

If any of this resonated and you want to try it yourself, here’s the simplest path.

On a Mac or PC: Download Ollama. It’s free, it’s simple, and it runs models locally with a single command. If you have a MacBook Pro with Apple Silicon, start with a model like Qwen 3.5 35B (24GB download) for strong all-round performance, or Ministral 3 14B (15GB) if you want something lighter and faster. Open your terminal, type ollama run qwen3.5:35b, and you’re running local AI. That’s it.

On an iPhone: Download Locally AI from the App Store. It runs smaller models that fit on your phone. Start with Gemma 3 QAT 4B. The performance won’t match what you get on a laptop, but for drafting, brainstorming, and text processing on the go, it’s surprisingly capable. And completely private.

The one thing to remember: Download the models while you still have internet. They range from 3GB to 50GB depending on size and capability. Once they’re on your device, they work entirely offline. That’s the whole point.

The Intelligence You Can Carry

I started this experiment worried about a scenario where the internet disappears. What I found was something more useful than emergency preparedness.

I found a layer of AI that’s free, private, and always available. It won’t replace Claude or ChatGPT for serious research or complex tasks. But for thinking through problems, drafting communications, restructuring messy ideas, and working with text you supply… it’s already genuinely useful. Today. On your phone. On airplane mode.

We’ve become so dependent on cloud-based intelligence that most of us have never considered what we’d do without it. I’d encourage you to spend a Saturday finding out. You might be surprised, both by what doesn’t work and by what does.

The best intelligence isn’t always the most expensive or the most connected. Sometimes it’s the intelligence you can carry.