The Moment the Story Changes

Indianapolis, Hawaii, and the conveyor belt that finally explained everything

I’d been asking AI questions for months — the same way I used to ask Google. A question, an answer, a closed tab. Useful, but nothing more than that.

Then one day Copilot put a suggestion in front of me: Try making flash cards for the 50 US states. Not a question to ask. A thing to build. I felt curious. So in March 2026, sitting at my kitchen table, I decided to actually try it — make something, not just ask about something.

Two hours later I was still there.

Fifty states. Fifty capitals. Copilot kept submitting cards to me one at a time, asking me to imagine — to actually picture — an image for the capital and the state, over and over, the same request dressed up slightly differently each time. It wasn’t fast and it wasn’t simple, but it was working, in the slow, grinding way these things sometimes work, and I kept going because I was nearly there.

Then it asked me to describe a flash card for Indianapolis, Hawaii.

Indianapolis is not in Hawaii. Indianapolis has never been in Hawaii. Somewhere in two hours of talking, Copilot had lost the plot entirely, and it asked me, in complete sincerity, to help it build a card for a state capital that does not exist.

That’s when I cried. Not because a chatbot made an odd mistake — that happens, that’s forgivable on its own. It was the two hours underneath it. I’d wanted this one small thing to work, I’d put in the time, and it had unravelled into nonsense right near the end, and I didn’t understand why.

A short pause, because the why actually matters.

What a context window is, and why it ate my flash cards

Here’s the thing nobody explained to me at the time, so let me explain it to you. Every AI conversation has something called a context window — think of it as a conveyor belt. Everything you’ve said, and everything it’s said back, gets placed on that belt at the start of the conversation. The belt only holds so much. As the conversation keeps going, new things get added to the front, and the oldest things eventually fall off the back end and are gone.

Copilot is extremely verbose. It doesn’t just answer — it explains itself, expands, elaborates, fills the belt with its own commentary as much as with the actual task. Fifty rounds of “imagine an image for this capital and this state” adds up fast. Which means the belt fills up fast, and the early instructions — the actual point of the conversation — fall off the back long before the conversation feels finished. By the time it asked about Indianapolis, Hawaii, the part of the belt that remembered we were talking about real state capitals had already gone.

That’s not a quirk. That’s the mechanism. And once I understood it, the two hours — and the tears at the end of them — made a different kind of sense.

What happened after

I haven’t opened that Copilot conversation since. It’s sitting there, untouched, since March — and if I opened it today, its own memory of how it started would be long gone, fallen off the back of its own conveyor belt months ago.

But that afternoon did something. Still at the same laptop, I opened Claude.ai in the browser for the first time. I’d talked to Claude on my phone before, in passing, the way you talk to anything new and not yet trusted. This was different.

I described what I wanted. We talked it through. And in under an hour, I had it — the same idea Copilot had spent two hours grinding me through and still got wrong at the finish line, built properly, with no wasted rounds and no Indianapolis in Hawaii.

That’s the part I actually want you to sit with. Not the tears. The hours. I didn’t write code. I didn’t follow a tutorial. I had a conversation — just a conversation — and a working thing came out the other end. Two hours of one tool fighting itself versus one hour of just talking to another. That contrast is the whole post.

An invitation, not a warning

I’ve told you about the tears because they’re true, and because I think the cleaned-up version of this story — the one where I sound capable from the first paragraph — wouldn’t actually help anyone. But the tears aren’t the headline. The hour is the headline.

If you’ve been circling AI the way I circled it for months — curious, a bit wary, not quite sure what you’d even use it for — I’d say: go and tinker. Not with a plan. Not with a course you’ve enrolled in. Just open something and try to build a small, slightly silly thing. A flash card set. A list. A tool for the one tiny, specific problem that’s been bothering you for years and never seemed worth fixing properly.

You might hit your own Indianapolis, Hawaii. If you do, now you know why, and you know it passes.

But here’s what I actually want you to take from this: that Tuesday afternoon, the one where I opened Claude in a browser for the first time, a conversation — nothing more than that — turned into a thing I’d actually built, in less time than I’d just lost going nowhere. That was it. That was the whole moment.

It doesn’t announce itself. It just leaves you holding something that works, made out of nothing but talking, and you realise, a beat too late, that you’re not the same person who opened the laptop an hour ago.

Have the courage to tinker. With any luck, no two-hour detours required. But even if there are — they’re not the end of the story. They’re just the part before the good bit.

Sandi is a Melbourne-based problem-solver, crisis-averter, and translator of the technical into the human. She spent decades being the person everyone called when something was broken, confusing, or just needed explaining properly — earning a reputation that preceded her wherever she went. Now she’s channelling that same instinct into AI: making it accessible, practical, and genuinely useful for people who think it isn’t for them.