Since my taste profile went into a project, the answers have been better. They fit, they name amounts that are realistic, they don’t send me to an importer who doesn’t deliver here. Then I looked back at what I had actually bought this quarter, and nearly all of it resembled each other.
That is not a fault in the model. I asked it to match my taste, and it did.
The profile is a filter, not a portrait
What sits in a profile is always a snapshot of your preference, written in the words you happened to have that day. Mine said, among other things, that I am lukewarm about oak and that I want acidity in a white. Both true, then.
The difference with a person who knows your taste is that someone in a shop occasionally puts something else in your hand. Not because it matches your profile, but because it just came in and they are curious what you will make of it. That is a social move, and it is not in a project’s instructions.
A model does exactly what is written down. Ask for wine that suits you and you get wine that suits you, in a circle that tightens slightly with every round. You buy it, you enjoy it, you confirm that, and the next recommendation lands a notch more precisely on the same spot.
What it costs
With wine the damage is small. You spend a year drinking decent bottles from two regions and miss what is happening in the third. Annoying, not serious.
The pattern is the serious part, because it is the same pattern running through your news, your music and your work. Wine is a conveniently small example: you can trace exactly what you bought, you have the receipts, and the circle becomes visible within a quarter. With information it takes years and you never notice.
There is a second layer. My profile was written by me, so my circle is my own choice. In most recommendation systems the profile is derived from your behaviour, which means it no longer records what you want but what you did. Those two diverge quite a bit. This year I bought more bottles I would recommend to myself than bottles I actually wanted to try.
Three rules that open the circle
I did not throw the profile away. I added three rules to it, and they do the work.
With every recommendation, include one bottle that does not fit my profile.
Say which part of my profile it contradicts, and why you think it is
worth trying anyway.That is the important one. Not “surprise me”, which produces randomness. The model has to point at the rule it is breaking, which turns it into a reasoned exception rather than a guess.
Once a month: look back at what I bought and tell me where my taste has
become narrower than my profile describes.This one showed me that “acidity in white” had shrunk in practice to two regions, while the profile was written much wider.
When I reject a bottle, ask whether I am rejecting it on taste or on habit.Irritating in daily use, and useful for exactly that reason. Half the time it was habit.
What it doesn’t fix
These three rules make the model more critical of my profile, and it is still the model that takes my profile as its starting point. It stays inside what it knows, and it has no idea what is on the shelf around the corner from me this week.
The real breaks this year did not come from a prompt. They came from someone pouring without saying what it was, and from a bottle I bought because the label was ugly. That remains the most reliable method: tasting blind, and occasionally buying something for the wrong reason.
If you take one thing from this, make it that first rule. Ask for the bottle that contradicts your profile, and have it name the rule it breaks. You don’t have to buy it. Seeing how narrow the circle has become is enough.