Is There Anything You Only Learn by Showing Up?

I have a Duolingo streak somewhere past 1,000 days, which sounds impressive right up until somebody asks me to order lunch. I know the Spanish word for turtle (tortuga, and I will find a way to work it into a conversation if you give me long enough). I’ve lived in San Juan since early 2023. Nobody here has ever needed me to say turtle.

That’s a fair summary of what 80-plus countries actually did for me, and since the travel is the thing people ask me about more than anything else on my resume, let me answer the version of the question that’s genuinely interesting instead of the version that goes over well at dinner. Is there anything you only learn by showing up?

Short answer: yes, though it’s a much smaller yes than the one people want, and I’ve been trying to describe it accurately for about 10 years without a lot of luck.

Start with the humiliating half, because nobody in my position ever leads with it. Sit me down against Claude (or Gemini, or GPT, I’m not precious about which) for a quiz on the history, politics, economy, cuisine or geography of any country I’ve been to, and I lose, badly, every time. It has read more about Laos than I will ever know about Laos, including the specific parts of Laos I walked through on my own feet. Every time somebody tells me that travel gave them a deep understanding of a place, I want to very gently ask them to name the current head of state. (I’d fail that on something like 60 of my 80. Possibly more. I haven’t audited it, I’m not going to audit it, and if you’re about to helpfully offer to audit it for me, we can still be friends but it will be strained.)

So “travel broadens the mind” is mostly a thing we say at parties. Most of what I picked up is trivia, and I paid a genuinely irresponsible amount of money for trivia that Wikipedia hands out for free.

Then there’s the other part, which is smaller and real.

In 2025 I hiked Son Doong in Vietnam, the largest cave in the world (I’d like to tell you I went for the geology, but I went because it was the biggest one). I knew it was the largest cave in the world before I got on the plane. I’d seen the photographs, and the photographs are extraordinary and completely useless! A cave that size has nothing in it a camera can use for scale. The comparison people reach for is a 40-story building. I’d read that comparison maybe 10 times and it did precisely nothing to me, and then I stood in the thing and it did quite a lot to me, and I still can’t hand you the difference in a sentence.

Same story on the Everest Base Camp trek with Dawn this April. Everything I read beforehand was about oxygen and altitude, which is fair, since altitude is the part that can kill you. Nothing I read prepared me for how boring it is. 2 weeks of deciding, again, to walk for another hour, at 18,571 feet, with no view left that you haven’t already photographed twice (I photographed everything twice. Dawn has seen the folder. Dawn has opinions about the folder). Not one of the 30-odd articles I read beforehand mentioned that, and I doubt the model has read it either, because nobody writes it down. It doesn’t make a good article.

That’s the category. People call it intuition or wisdom or the human touch when they want to sound profound about something they haven’t actually defined, and what it is, is the stuff nobody wrote down.

The map is not the territory, which is a very old idea and definitely not mine. What I’d add is that a model doesn’t have the map either. It has every map anybody bothered to publish, all at once, and that’s enormously more than I have. It also comes with a bias baked in, because it contains what got written down, weighted by how often it got written down. The things everyone knows and nobody records are just missing, and I don’t think there’s any reliable way for the model to tell missing from not-true. Douglas Adams got here first, obviously. The Hitchhiker’s Guide outsold the Encyclopedia Galactica because it was cheaper and had DON’T PANIC on the cover, and its entry on Earth was “mostly harmless.” Enormous coverage, correct as far as it goes, 2 words where you needed the paragraph.

Now take that to your finance department, since that’s where I actually make a living and I’ve been circling it for 600 words.

Your model knows the written-down version of your company. General ledger, policy docs, chart of accounts, the last 3 years of variance commentary (assuming somebody kept it, and be honest with yourself about that). It knows all of it better than anybody on your team, faster, at 3AM, without complaining once about the formatting.

What it doesn’t know is why that cost center exists. The number in row 42 has been wrong since the 2019 acquisition and 5 people adjust for it every month without ever writing down that they do. One customer always pays 40 days late and that’s fine, while a different one paying 4 days late is your leading indicator that something has gone badly wrong over there. And nobody has ever told it which of your 4 revenue numbers your CFO actually trusts, because nobody has written that down and nobody ever will, and the person who knows retires in March (this is not hypothetical, by the way. Somebody is always retiring in March).

That material is most of what a finance organization actually runs on, which took me an embarrassing number of years to believe, because I kept assuming somebody somewhere had it written down in a document and I just hadn’t been handed the link.

So the highest-return AI project in a lot of companies isn’t a deployment at all. It’s spending 6 weeks getting the unrecorded material out of 5 people’s heads and into text a model can read: why the adjustment exists, which number is authoritative, what the exceptions are and who granted them, and about 40 other things somebody will tell you in a hallway and never once in a document. I’ve spent 30 years inside Hyperion, then Essbase, and now Oracle EPM and the MCP servers we build at Caprus to reach it, and every organization I’ve walked into has this material sitting in people rather than in files (rare few exceptions, and every one of those had a documentation person everybody privately thought was a bit odd). It lives in tenure, which works fine right up until somebody takes a better offer, and in my experience the person who knows the most is also the person most likely to get one.

Do that and the model gets better immediately. The part I’d pay for is the second-order benefit, which is that you still have it after the person leaves, and you were carrying that risk the whole time whether or not the AI project ever happened, so you may as well get an AI budget to pay for fixing it.

Keep the rest in the room with a human. Somebody has to notice when an answer is fluent and specific and confident and wrong, and that’s a judgment call made against years of watching things go sideways, and I have yet to find the article about it. It’s the first panel on the safety card I put up on the site this week, and I put it first because it’s the one I’ve watched go wrong most.

Anyway. I’d still lose the quiz, and the genuinely non-trivia portion of 80 countries fits in about a paragraph. (Edward, you spent 30 years and an amount of money Dawn has receipts for, and the yield was one paragraph and a Spanish word for turtle. Nobody made you do any of this. You could have read a book on the couch like a normal person, and the couch was right there the whole time.)

If your company is somewhere in this and you’d like to argue with me about which part, I’m easy to find. Write me at Edward@Roske.AI, or grab 13 minutes if you’d rather just say it out loud.

Asking good questions, Edward