Sure, I’d be happy to help train your model 🔗
My continuous glucose monitor works with an app. The app is mostly cool: Good historical data, great for developing day-to-day awareness, and it’s hooked into other stuff so I can make connections between sleep, exercise, and blood sugar.
It also has a meal diary. When I first started using the app, you could choose between a photo entry or a detailed entry for each meal. If you chose a photo entry, it’d just let you take a picture of your meal and timestamp it. If you felt like going back and filling in what you ate, it would provide a link to a database on kinds of food and their potential metabolic impacts. If you chose a detailed entry, the picture became optional and you could fill in more detail on the spot.
I preferred photo entries. I know what’s in a given meal so it was fine to look at a week of historical data and check in on photos of the meals that seemed to be connected to spikes or dips.
Recently the app succumbed to the AI fad, so the photo log UI changed: Now the app tries to guess what you’re eating by analyzing the photo. I don’t think, in over a month of meals, it has had a better than 10 percent success rate. Some of its guesses are simply wild. It isn’t learning from me specifically, because I have a very consistent breakfast choice and no amount of correcting the same basic picture taken almost daily has swayed the app.
There’s no way to turn the AI assistant off, though. They wedged it into the workflow and your choices are “correct it” or “ignore it.”
I tried to correct it for about a week then snapped out of it: There is no way the damn thing made it out of beta without everyone realizing it is infuriatingly bad at what it does. I’m sure they intentionally released this half-baked thing counting on people to train it by correcting it for the sake of accurate records.
So I just accept whatever it proposes. You think that bowl of lentil soup is a chocolate sundae? Sure. Yes, now that you mention, I thought I was eating a tasty omelette, but I can see now that it was a ham.
I’m happy to send them an invoice if they’d prefer different behavior on my part.