June 25, 2026
Taste issue at mile 26
You can't unit test for taste
App maker wanted AI to find scenic running spots, but the comments turned it into a taste war
TLDR: A developer used public map data and AI to add interesting landmarks to a virtual running app, then discovered that choosing what’s “good” is the hard part. Readers turned that into a lively fight over whether taste can be taught to software at all — with bonus marathon nitpicking and Iceland slander.
A developer set out to make a wholesome running app more inspiring by adding famous sights and hidden gems along huge virtual routes, using public location data and a little help from the AI assistant Claude. But the real action wasn’t in the code — it was in the comment-section philosophy cage match that broke out over one spicy claim: can you “unit test” for taste, or is good judgment forever unprogrammable?
That question sent readers into full debate mode. One camp basically yelled, “Of course you can — just tell the machine what you like and score the results,” with one commenter even taking a detour to roast Iceland as a terrible example because it’s supposedly just Reykjavik and a ring road. Ouch. The other camp went full poet mode, arguing that taste is the part of the plan you forgot to write down — or the part you couldn’t write down even if your life depended on it. Another reader delivered the line of the thread: testing can make sure something is not wrong, but it can’t magically make it delightful.
And then came the marathon jokes. The article’s “life is a marathon, not a sprint” vibe got lovingly dragged by runners pointing out that, actually, both are brutal and end with your tank completely empty. So yes, the app is about scenic long-distance motivation — but the internet made it a referendum on art, bias, AI ego, and runner pain.
Key Points
- •The article describes adding points of interest to *In the Long Run*, a virtual running app that maps users’ accumulated Strava mileage onto long-distance routes.
- •GeoNames was selected as the source dataset because it provides extensive geographic data, categories, links, downloadable files, and a Creative Commons license.
- •The processing pipeline used Python, Apache Parquet, and DuckDB, with Claude assisting in planning and implementation.
- •The author organized the work into milestone-based plans and used shorter context windows in separate AI agent sessions to improve response quality.
- •Initial data processing involved joining GeoNames files and filtering out administrative divisions while keeping selected feature types such as parks, historic sites, castles, monuments, and mountains, plus additional population and elevation filters.