about a year ago i started logging every meal i ate into chatgpt. just typing "two slices of pepperoni pizza, a beer, handful of olives" and asking for a calorie estimate. pasted the totals into a notes app at the end of each day. did it for 6 months. lost 10kg.
i'd tried myfitnesspal years before and bounced off the database/barcode thing within a week. food logging felt like data entry. chatgpt was the first version of it that didn't feel awful.
four things i learned:
accuracy doesn't matter as much as people think. chatgpt is probably 10-20% off on any individual meal. fine. what matters is that it's wrong in the same direction every time. if it overestimates pizza by 15% consistently, you still get a real signal about whether today was higher or lower than yesterday. the trend is correct even when the numbers aren't.
the back-and-forth is what made it click. if it gave me a weird answer i'd just say "no the portion was smaller, maybe half that" and it would recalculate. you can't argue with myfitnesspal. that loop is impossible in a database app.
photos work better than i expected. snap a plate, ask for a guess. it gets the items right almost always. portion sizes are the weak point but you can correct those in the same chat.
the workflow itself was miserable. no history of what i'd eaten before, no way to quickly repeat the same meal, no search. conversations would get long and messy so i'd have to reset and re-explain the counting rules from scratch. and the math didn't always add up, the running total it gave me would sometimes not match if i added the meals up myself. but still it was working so i kept going.
i lost about 10kg over those 6 months, partly because the friction itself helped. i was thinking about food more because logging required attention.
then my second kid was born. stressful period, sleep deprivation, eating whatever was around, eating because i was stressed. i stopped tracking entirely and gained half of the weight back over the months after. not surprised in hindsight, the tracking workflow had too much overhead to survive a life event like that. the second i was tired or rushed it was the first thing to go.
so now im doing it again. this time i got annoyed enough at the original setup that i built my own wrapper around the same idea, mostly so the friction is low enough to keep doing it on a bad day. same openai gpt model under the hood as what i was using in chatgpt before, just with an interface built around how i actually live. type a meal or snap a photo, get an estimate, correct it if it's wrong, history is searchable, repeat yesterday's lunch in one tap. all the stuff i kept wishing chatgpt had plus a cool cyberpunk/space/sci-fi kind of vibe.
not linking it here, but it's called Excaloricate if anyone's curious. named it that because i liked the "excommunicate calories" idea.
if you're handy with code by the way, this stuff is genuinely buildable in a weekend. the openai api does all the hard part, you just wrap it in whatever interface fits your life. one of the more fun side projects i've done.
[link] [comments]