I used AI-powered calorie counting apps, and they were even worse than expected

https://lifehacker.com/health/ai-powered-calorie-counting-apps-worse-than-expected

by gnabgib • 1 year ago

245 292 1 year ago

muratsu2• 1 year ago

These apps look gimmicky but they are growing like crazy. CalAI, for example, is making 35M+/yr. I’m hoping that some people really find value in these apps and that they are not just a pure marketing driven play.

tekla• 1 year ago

They provide value in that they lie to you and you feel better about yourself.

klabb3• 1 year ago

That’s exactly what I was thinking when I read the article. ”Why are you thinking this is meant to work?”. But of course, normal trusting people will believe it if it’s on the ”trusted app stores”.

These apps with graphs are just selling you a sense of control, of compulsive metric driven decision making. Its partly the consumers’ own fault – many people won’t care even if they knew. But it’s also predatory by the companies. It’s consumer tech in a nutshell: put a pretty but shitty app on the App Store, spend 80% on marketing, show ads in the free tier, sell overpriced premium offerings to those who are careless with money, rinse and repeat. I follow some subreddits and groups for ”entrepreneurs” and it’s just like this.

muratsu2• 1 year ago

I don’t mind if the reality of their business is similar to gyms in the sense that they make money off of peoples desires but have low utilization.

I used calorie counter apps before and they helped me lose weight (by mostly educating me what to eat/avoid). With these apps, I feel like the core premise - counting calories - isn’t even working and they’re selling people hopes and dreams. That’s danger territory.

xnx• 1 year ago

CalAI, for example, is making 35M+/yr.

Source?

debuggerson• 1 year ago

I happened to scroll through the founder’s interview on YouTube. You may take a look. https://youtu.be/t0U3UREdjmo?si=14oeWqtL9oVszab5

asdev• 1 year ago

I wouldn't trust the revenue numbers, plus the churn is probably insanely high. Flashes in the pan

wahnfrieden6• 1 year ago

There is good money in the App Store once you crack social marketing. The founder shows several videos of reloading the App Store Connect page showing revenue numbers. It's not unbelievable.

(Not to defend the app itself obviously)

AstroBen• 1 year ago

What's going on with the google reviews of that app.. basically all 5 star with one word over and over again - 'good'

They're solicited, right?

m_ke• 1 year ago

I've heard from someone who knows that they're scamming people like crazy. Supposedly they also setup a bunch of LLCs to hire influencers then never paid them.

TuringTourist• 1 year ago

I suspect this might be because diet marketing before AI was one of the most fraught with misinformation subjects you could run across. This is because the "sale" of the idea has an effect on every salesman, so all of the salesmen are trying to sell every thing at once,since that's also selling their thing, and ground truth gets stampeded. Now when you combine diet marketing and AI, you get a multiplier. Both in the sellers and the buyers, since the desire to believe is stacked.

jimjimjim• 1 year ago

This makes me sad. Yes a fool and their money are soon parted but this drives user expectations for apps even lower, pushing competent apps out of the market because of the price difference.

dyauspitr4• 1 year ago

Gimmicky my ass. I speak a sentence every night on a thread to ChatGPT about what I had for breakfast, lunch and dinner along with quantities and it spits out my macros and nutritional breakdowns effectively. It’s the easiest and most no nonsense way I have found to record this information. Since it’s all on the same thread it always outputs the data in the same format and I can ask it for information over custom time ranges on the fly. I’ve also uploaded the nutritional labels for all my protein shakes and supplements to the same thread so I can just say “I had my nighttime shake” and it knows what I am talking about immediately.

rahidz• 1 year ago

I speak a sentence every night on a thread to ChatGPT about what I had for breakfast, lunch and dinner along with quantities and it spits out my macros and nutritional breakdowns effectively.

Have you verified that these are mostly accurate?

ineedasername• 1 year ago

I do this too and it's been spot on. I also have a very good idea of what I'm eating though: "2 oz of steel cut oats, 1 banana, probably about 4oz by weight, without the peal, about 12 ounces of coffee with 1 tsp of sugar and about 4oz of light no-sugar almond milk"

dyauspitr4• 1 year ago

Yes, a lot initially and now I’ve looked at them so much I can roughly tell without looking it up.

jvanderbot• 1 year ago

Not OP, but I have verified, and it is accurate. At least accurate enough from images or descriptions when assuming average proportions. Where it's really accurate is when you have measurements, even rough. I took to having a 1/2 cup scoop around the house so I could get volume measurements.

Initially I verified against labels by eating the serving size precisely. Then, I switched to weighing and doing the math on random days or meals.

For some items I would tell it the calories and macros and it would remember (a pb spoon is a common treat around my house).

It's good.

But like all food journalling the benefits are not so much having macro balances or tallies, its in making you aware and making you think about what you're eating. That alone takes you out of autopilot and lets you eat better.

It's just bad at math. So I would ask it to output a json struct for each day, that I could inspect by sight, then I'd save that and had a python script to actually tally.

I had predictive analytics and the whole 9 yards. Maybe I should have started a business! https://jodavaho.io/tags/diet.html

DontchaKnowit• 1 year ago

This is different than the style of ai calorue app that extrapolates caloric content from pictures of your food. That kind is gimmicky BS

wahnfrieden6• 1 year ago

The app is also using a low quality and outdated model (Gemini 1.5) to reduce costs, whereas this poster is likely using GPT 4o

bigmadshoe• 1 year ago

Tasks like summing a bunch of numbers from different parts of the input over a specific time period are still pretty error prone for LLMs. I would exercise caution if these results are important to you. A database or spreadsheet is going to be far more reliable – maybe you could have ChatGPT output a structured format and you could do the summation in a google sheet?

dyauspitr4• 1 year ago

Yeah I know, it’s not something I need do a lot.

tcoff91• 1 year ago

Read the article. What you are doing has nothing to do with the apps the article is talking about.

You are providing accurate info to an LLM and it is assisting you with information processing.

These apps are trying to just take a picture of arbitrary food and just trusting a neural network to be magical and tell you the macros in the food from a photo.

dyauspitr4• 1 year ago

I actually think it’s quicker to just speak to the LLM about your diet than even taking a picture. But yes, I understand what you’re saying.

jaronheard• 1 year ago

I like how in the article the author includes the instructions for Cal AI: - Include a reference object (like a coin or your hand) for scale And then the screenshots show just a pictures of bowls of food with no reference objects at all.

Honestly curious if that would have improved the author's experience.

PlunderBunny• 1 year ago

I noticed that too - I wonder if the coin/hand was cropped out?

horsawlarway• 1 year ago

I'm guessing yes. I think you can even see half a coin in the cropped photo of the tofu salad (lower right edge, under the metal bowl).

TeaVMFan• 1 year ago

I prefer efficient manual entry to vibe-coding calories. I built my own Java-based SPA: https://frequal.com/cf/

It's local-first so you control this sensitive information. It features easy reuse and exercise tracking.

Terr_• 1 year ago

Also, a certain level of human involvement is what gets you to really care and change behavior.

I exploit my "frugality" neuroses by treating calories as an expense.

tunesmith• 1 year ago

Weird article, it mentions three apps at the beginning, doesn't even test two of them in favor of two others it didn't mention, and then declares all the apps failures.

duskwuff• 1 year ago

Funny, I was just discussing this on HN a few days ago:

https://news.ycombinator.com/item?id=44185892

There's a lot of questions about the nutritional content of food which you can't answer just by looking at a photo, like "is that a glass of whole milk or non-fat", or "are those vegetables glossy because they're damp or because they're covered in butter". No amount of AI magic is going to solve that.

dwetterau• 1 year ago

It will get to the point where it could know to ask those questions, like a human would. Clearly these tools are not there yet, but I think the models are probably good enough for this already.

bitmasher9• 1 year ago

It’s probably the only thing that would get me to start taking photos of my food, but it would need to be about as accurate as I am in chronometer.

hammock2• 1 year ago

Those are easy manual adjustments to make. Far easier than estimating an entire meal

fullshark• 1 year ago

Maybe in aggregate they are fine if the noise is unbiased and you do some things like correct it when it says an apple is tikka masala, but yeah, someone using this and lying to themselves about the calorie content of junk/comfort food seems pretty likely.

nvahalik3• 1 year ago

I work for a company that offers nutrition tracking on an app in the App Store.

We are not shipping camera functionality yet. But our concept is to not necessarily guarantee the accuracy of portions but to make lookup easier.

We also spent the time to get the AI integrated with a verified database. This made our results far more accurate.

We tended to find that without the lookup the calories and macros would be generally correct. The math was usually within a margin of error of 5%. This was acceptable except that… there was no micronutrient values and you couldn’t really adjust the portions at all. The system just dumps the macros and while you can halve something… the user experience isn’t great.

Ultimately, if you want precision: manual entry is the only way to go. I feel like out approach will end up being very great once we work out the kinks. Our search isn’t spectacular and as a team we are learning a lot of about prompt engineering and how to make best use of the AI.

irrational• 1 year ago

Does that work for homemade food as well? The vast majority of the food we eat is homemade with recipes that don’t have any sort of nutritional information. I’ve always wished there was a simple way to figure out the calories. Taking a picture would be ideal.

adrian_b4• 1 year ago

For homemade food it should be easier to make reliable estimates of the calorie content, because you know with certainty all the food ingredients and their amounts.

The food ingredients with the highest calorie content, like various kinds of seeds or nuts or flour or meal or oil or fat or sugar or dried fruits or dairy products, come usually with calorie estimates from their vendors.

For other ingredients, like various kinds of meat or of fresh vegetables or fruits, there are online databases with typical nutritional information, like the USDA database. Some of that information can even be found in the corresponding Wikipedia pages.

jurgenkesker2• 1 year ago

Yes, I would think that would work better indeed. As a augmentation or help tool. I would love to be able to say to MyFitnessPal that 'I ate this and that food, same as usual, and oh yeah drank this.' Just as a easier input interface. I wouldn't trust a pure AI solution without some proper database behind it.

nvahalik3• 1 year ago

Yeah. The big problem is that "augmentation" is hard because we (humans) have an internal process for how we think about things that is hard to define and building a flowchart for how we understand foods doesn't even necessarily capture things. Very well. You can take something like "chocolate chip pancakes" where the context can be "<brand> <item>" or "<modifier> <food item>". And then you can search.

But even though we've integrated it with a good food database, the process of searching isn't great because sometimes things like brand names don't get recognized and/or modifiers may get confused because... is it a brand? Is it a way of preparing something?

Ultimately we are working on improving how our search works by not just searching by the name, but by getting information about brand, the product, and possible serving options as well. These would better inform the search and allow us to, say, fallback without a brand if we can't find the brand.

The other problem has to do with variant detection. I can say "kirkland sous vide egg bites" but there are 3-4 variants of them. And right now most databases are just "here is the item you requested" without looking at possible variants, which is a problem that we are going to end up solving ourselves.

It's been interesting because we've learned a lot about how people "think" it should work vs. how it actually works.

cowanon2222• 1 year ago

I've had pretty good results using the AI features in Macrofactor. It's certainly not perfect, but it does a pretty good job with mixed text and photos and allows you to easily fine-tune the results.

Macrofactor is also the only app I've seen that actually estimates your underlying metabolic rate and adjusts accordingly. It predates the recent AI surge, and seems to have a team that's studied nutrition science behind it.

thearrow• 1 year ago

Same. I was very skeptical when Macrofactor introduced this feature, but have since been incredibly impressed. The ability to give it text alongside a photo and then adjust the results (broken out by ingredients) are critical. I’ve also been taking pictures of food sitting on scales and it will take the measurement into account.

Seems like the Macrofactor team took their time developing this feature, as it felt like they were one of the last to roll it out, but the extra polish definitely shows and was worth the wait.

sunnyam• 1 year ago

I think the key difference is that they perform a search of the foods in Macrofactor's database which means that you're more likely to get a good estimate.

From someone who weighed and scanned a lot of foods, it has really improved the workflow

Kiyo-Lynn• 1 year ago

I’ve tried a few AI-powered calorie tracking tools too. At first it felt super convenient, just take a photo and get the numbers. But over time I kept seeing mistakes. Sometimes it would even tag a salad as soup. Eventually I went back to the old way. Fewer photos, more mindful eating. Honestly I feel better that way.

spacecadet• 1 year ago

Honestly this is it for me. Im actively trying to reduce my reliance on tech, even more so when it comes to gimmick tech. A nightly walk with no phone will make you drop pounds too...

solumunus• 1 year ago

I genuinely can’t understand why anyone would even TRY using these apps. Of course they’re going to be terrible, but how could anyone with a modicum of intelligence suspect otherwise?

surfmike3• 1 year ago

The founder of parrotpal (another AI calorie tracking app, supporting both text and photos) points out that using photos is one of the least accurate ways for tracking food: https://www.instagram.com/parrot.pal/reel/DAB9NtfM250/

Users also seem to like using photos less than text for food tracking: https://www.instagram.com/p/DBoSeQzM3bQ/

The author of this article seems to be against the concept of calorie counting as a whole too, but calorie counting does work well for many people. They also bring up intuitive eating as an alternative, but intuitive eating is not intended for weight loss while that's what calorie tracking is usually used for (though it can also be used for maintenance and for weight gain).

Personally, after using MyFitnessPal for a couple years, switching to ParrotPal made calorie counting way less time-I just need to give it a quick text (or voice) description and it does a surprisingly good job of estimating. There are a few times when I need to adjust it, I mostly try to overestimate. It's not perfectly accurate but it gives me enough accuracy to have successfully lost and kept my weight off.

tayo423• 1 year ago

seems to be against the concept of calorie counting as a whole too

What are peoples criticisms of calorie counting? Its the only thing that keeps my weight from creeping up. There are way to many calorie dense foods that can easily sneak in if im thoughtless. (I want my 20 year old life style back :( )

schmichael• 1 year ago

My unprofessional nutritional advice is: do what works for you! So if that’s calorie counting: great! No need to give it a second thought.

But if you are interested in critiques, a fair summary might be: calorie counting is at best extremely imprecise, both on how we measure the calories in food and how we estimate energy expenditure by activities. A little googling should lead to numerous discussions. I really enjoy the podcast maintenance phase, and even if you don’t want to listen to the episode they helpfully include tons of links: https://maintenancephase.buzzsprout.com/1411126/episodes/106...

darkwater• 1 year ago

What are peoples criticisms of calorie counting?

Probably the same criticism that applies to all "methods to do a thing". That people often miss the forest for the trees and obsess on the wrong metric (counting calories while still ingesting preprocessed industrial food and beverages) instead of the right one (losing weight while being healthy at the same time).

shlant• 1 year ago

most people do not arrive at the idea that CICO (Calories In Calories Out) is useless or incorrect on their own - they usually buy into that belief because they follow specific diet communities or health influencers that are incentivized to tell them that.

rsynnott• 1 year ago

Sometimes people get a bit obsessive with it, would be the main obvious problem.

edanm2• 1 year ago

Personally, after using MyFitnessPal for a couple years, switching to ParrotPal made calorie counting way less time-I just need to give it a quick text (or voice) description and it does a surprisingly good job of estimating.

I haven't used ParrtPal, what makes it easier than using MyFitnessPal?

surfmike3• 1 year ago

- More freeform -- just type what you ate (and add as much detail as you want) and it will log it, nothing more to select or search from after - You can log several dishes or meals in the same sentence - You can use voice

With MFP, there was always searching, then selecting the best entry, then fiddling with portion size. It usually ends up taking many times longer for me.

Sleepthinker• 1 year ago

As a gym enthusiast, if this app can’t give me accurate calorie counts and nutrient data, it’s basically useless to me. But for the average person, this is a big deal because we tend to seek convenience and want the easiest path to our goals. If this app wants to truly deliver results, it still has a long way to go.

Fun fact: The founder of Cal AI is just 18 years old.

tjpnz• 1 year ago

If you want calorie counting to be of any value to you you're going to want to be accurate. Thankfully we're mostly creatures of habit. Source accurate numbers for everything you eat in a week and you're 90% covered for every other week. I recorded 30 items over two weeks and then added 15 more over 2.5 months. Don't sabotage your efforts with tools that don't work.

tayo423• 1 year ago

This is my experience too. Not even 2 weeks, it was like a really tedious 3 or 4 days then I pretty much knew what to take out.

smartgamer112• 1 year ago

Yeah cal AI is the worst of them all

senko• 1 year ago

CalAI is just objectively a bad app even discounting its AI calorie estimator as an irrelevant gimmick:

* the UX is bad

* the stats are laughable

* charts worse than nothing (really, misleading to look at)

* lacking basic functionality such as “duplicate for today”

* can’t see or export the meal photo

* their “streak” attempt at gamification is so bad it literally cannot count

And yes, the AI there is atrociously bad. Not only can it not reliably detect calorie content (I don’t think anyone could on a basis of a photo in general case), but it so often mislabels food (like the example in article) it’s just a comedic relief. And there’s a “Fix” button which does nothing.

It’s like they never really tried hard. Combined with the general lack of quality, looks like a typical quick money grab riding the AI hype wave. I’m pretty confident the app could be vibe-coded in a week (and that’s generous). Maybe the backend is more complex? Don’t think so as there are no social features, it just stores data.

Paid for a year (Android app) to see how well they applied AI to the field, now continue to use it like a meal photo log with some (manually entered) metadata, otherwise Excel would be a nicer solution.

Traubenfuchs• 1 year ago

We live in a post quality age.

Consumers just don't care.

AI dialed this to the max: The early vibe coding startup catches the worm. Working with designers, doing A/B tests, QA, product management, nowadays all of this will make you lose to the competition.

You just need to get something out that looks as if it works, that makes users feel as if it works, whether it really does or not, whether the quality is atrocious or good.

fakedang• 1 year ago

Exactly. At what seems to be 30 USD a year, that's laughably low for most users. I'm actually impressed on the kids for finding a niche snake oil market and selling it to them.

username223• 1 year ago

It's not that consumers don't care -- it's that "they" (i.e. "we") have been beaten down.

Connect an image labeling API to the USDA's database of nutrients, make a slick-looking app, and you have a product. Pay for some marketing, offer a three-day trial after which you auto-bill for a year, make refunds as difficult as possible, and you have a business. At $30/yr, 10k subscriptions (deliberate or accidental) gets you $210k after app store commission. If you're lucky, you get bought out for even more. If not, you still made good money for not much work.

jonplackett11• 1 year ago

I always just assumed these were terrible.

There just isn’t enough info in a photo, even for the world’s leading calorie counting expert, to give a sensible estimate of the calories content. Ans even more so for macros.

Add to that the fact vision is still pretty crap and you have a complete joke that I can’t believe anyone is paying for.

automatic6131• 1 year ago

I bet @onionweigher could - if all you ate were onions

teddyh• 1 year ago

Human children can’t live on onions.

<https://scarygoround.com/badmachinery/index.html?pg=374#show...>

jclardy2• 1 year ago

I've used the AI features in Lose It and was pretty impressed, not for the caloric estimate, but when given a picture of a breakfast burrito it accurately split it out into the component parts, from there I could manually adjust amounts easily for food I make without having to manually search out every ingredient manually. As an additional tool it is great.

The issue of course is these new apps built with a sole focus of AI images for tracking. With a photo of restaurant food you can't see a sauce and the 5-10g of additional sugar content, can't get accurate guesses of what is in breads/pastas and figuring out general volumes of foods is not possible unless you have some kind of standard size reference in the frame.

Iulioh6• 1 year ago

Honestly the best use case i have seen is the ability to read a labels and add it to the database. (Similar to an OCR but better)

Correcting errors in MFP is the bane of my existence.

I have tried some other apps but i hated every single one of them for a reason or another and came back to MFP but that was one of the best features i would really love to have

mapt4• 1 year ago

However bad these algorithms are (and we're talking very wide intervals in multiple dimensions of uncertainty), you could immediately improve the vision/volume calculation part with a modest change - like requiring a size reference. "Place a unit of currency next to the plate, then take at least two photos of the plate from different angles" can get you pretty close to an accurate volume.

If they aren't doing even that much, then they're not very interested in accuracy at all.

jonplackett11• 1 year ago

The problem is, if you are serious about losing weight, you need to ideally aim for something like a 300 calorie deficit. Much more and you’ll feel shit, a lot less and it’ll be quite slow or go the other way. It’s quite a fine margin and even tracking using bar codes you’re not going to be perfect.

But even more important - you really gotta track your macros if you want it to work well - and again also not feel crap.

There is no way in hell an AI camera is tracking macros well

fwip• 1 year ago

The article says that at least the first app trialed does recommend putting a size reference in.

gavinray• 1 year ago

I have faith that ML can approximate calories from images under the following conditions:

- The food is inside containers of a standard size, which the model can use as a volumetric reference

- Optimally, the containers are clear plastic, for depth approximation

- The food is homogenous, or mostly-homogenous. Dishes containing a lot of mixed ingredients seem like an intractable problem to me

Now, whether or not abilities like this are useful to you would seem heavily dependent on your diet.

jonplackett11• 1 year ago

So here’s why that ain’t true.

Yoghurt.

You can buy zero fat Greek yogurt that has very low sugar. It’s perfect for losing weight. It’s about 50kcals per 100g. And also super high in protein.

Or you can buy yoghurt that is full of sugar and is 100kcals per 100g and has lots of fat and hardly any protein.

Even a human expert could not tell the difference without tasting them.

This is just one example of many.

davidcbc7• 1 year ago

And how is it going to tell how much oil or butter was used in a dish?

james_marks2• 1 year ago

What’s interesting to me is a tool can be terrible at its stated problem and still add value.

I suspect many of the results dieters get is “mere measurement”, e.g. it helps them be more thoughtful, seek education, etc.

So the value of the tool is in its ability to keep them engaged and hopeful (the unstated need), and the actual mechanic doesn’t have to be that good at the stated need.

davidcbc7• 1 year ago

It can also be a negative value.

If I diligently track calories using one of these things and don't lose weight I'll say "screw it" and stop bothering because obviously counting calories doesn't work. I think this is much more likely to be the case because even proper calorie counting is difficult and frustrating

jonplackett11• 1 year ago

It’s the ‘What gets measured gets managed’ truism - but does it hold true is your measurement is BS?

flippyhead• 1 year ago

There is a study somewhere that said people who wanted to lose weight and simply tracked their calories (without any attempt at reducing them) LOST weight vs those who did nothing. Awareness alone is valuable.

Maybe this one: https://today.duke.edu/2019/02/tracking-food-leads-losing-po...

Too• 1 year ago

Yeah. It doesn’t take much imagination to figure out that a Coke Zero looks identical to regular Coca Cola, yet they have vastly different calorie content, to put it mildly.

Sauces loaded with butter, sugar or other goodies will of course be the same story.

jonplackett11• 1 year ago

I used MyFitnessPal to lose 15kg a few years ago and I have a theory that the act of calorie counting (ie the effort of working it out and labour of putting it into the app) is a key part of why it works. It becomes just slightly more effort to eat something and it makes you stop and think. Making it easier I would expect to make it also easier to eat more.

kamma44342• 1 year ago

Attention and focus make it work. You must be ‘in it’ - otherwise it wont work.

myhf• 1 year ago

Of course it will work. Just tell ChatGPT that it is working and it will agree with you.

jclardy2• 1 year ago

I find a similar thing with budgeting apps. Manually tracking requires effort and for you to think twice before every purchase, whereas automatic transaction syncing with your bank means you can just not think about it.

In the case of dieting/calorie counting - I don't think you can get away with not thinking about it, especially with inaccurate estimates.

yoz-y• 1 year ago

Yep. Iirc any kind of diet works as long as it is restrictive. Even if the restriction is “do not eat anything that has blue on the package”.

devmor5• 1 year ago

As someone who lost nearly 100lbs (about 45kg) calorie counting and working out, this was a big part of it.

It also just made me cognizant of my bad eating habits. I’d absent-mindedly snack on something, go to log it, then think “wait, why am I eating this right now?” and stop.

It’s essential educating yourself on nutrition through habit.

jonplackett11• 1 year ago

Great achievement well done! Out of curiosity, did you track macros too and what calorie deficit did you aim for, and how long did it take?

jxjnskkzxxhx• 1 year ago

Yes!! I used an app called Waistline 2 years ago to keep myself on a brutal diet. I did < 1500 calories per day for 2 months. Lost 9kg. Came out with the same insight: the hassle of meticulously keeping track helps with the restraint.

ginko• 1 year ago

The app did recognize it as an apple this time, but estimated it at 80 calories when the actual count should have been closer to 120.

Where does the author take the 120 number from? The very link they gave says a large pink lady apple is 78 calories.

hyperbovine• 1 year ago

Presumably every apple has a different calorie count depending on ripeness, size, and variety...?

llm_nerd• 1 year ago

They do claim their apple is 222 grams, so presumably they took the 100g nutrition and multiplied it by 2.2.

wahnfrieden6• 1 year ago

Others have found that CalAI likely uses Gemini 1.5 (not sure which variant) which is used because it's very cheap, not for its accuracy.

I understand the fundamental issues with calorie counting via photo analysis/LLM. However these articles would be more interesting if they included comparison with results from actual cutting-edge LLMs, rather than the budget models. I acknowledge that apps don't use these models though to protect margins and lower cost.

Also:

Cameras capture two-dimensional images of three-dimensional objects.

iPhone camera captures depth, too. I don't think it's being used in CalAI though. This article is more about current commercialized solutions rather than analysis of state of the art capability.

tough2• 1 year ago

what's worse, this could be 100% done in local llms since its such a small batchable job

wahnfrieden6• 1 year ago

What model can run on a phone with accuracy closer to 4o than Gemini 1.5?

woadwarrior012• 1 year ago

I don't think it's being used in CalAI though.

They don't, but that doesn't stop them from lying about it on their homepage.

Snap a photo with Cal AI, and your phone's depth sensor calculates food volume. Our AI then analyzes and breaks down your meal to determine calories, protein, carbs, and fat.

The grift must go on.

wahnfrieden6• 1 year ago

Oh interesting. I wouldn't put it past them that they include that data in the input though. Why are you certain they don't? (I guess that they don't, too, because someone reported getting identical calorie readings from giving photos to Gemini 1.5 directly)

p1dda• 1 year ago

I welcome every article that exposes just how bad "AI" is.

blibble• 1 year ago

it still also can't count the legs on a dog

https://vlmsarebiased.github.io/

another decade and we'll be able to get the Not Hotdog app from Silicon Valley

perching_aix2• 1 year ago

How many times a day and how many unique food items does one need to eat to continually remain so hopelessly unfamiliar with their caloric intake that they need an app to track it?

What role can AI possibly have in this, when all you need to do is either turn around the packaging or do a web search for the dish you cook to get a number?

cactusplant7374• 1 year ago

Restaurants and dishes at home made by someone else that have a combination of different foods.

matkoniecz2• 1 year ago

they need an app to track it?

"track" is important here

How packaging or a web search will help with exactly remembering how much was eaten in past?

perching_aix2• 1 year ago

By enabling you to think in terms of what you ate, rather than numbers. If you eat a varied but regular diet, this enables you to stop having to track. You'll just know your approximate caloric budget and intake for the day, and what that means in terms of food items.

When you have to take ownership of your diet (and thus your weight), that means a permanent lifestyle change. Tracking is simply not sustainable indefinitely, so you must establish better consumption habits while you do so, that will stick past the tracking. Tracking and getting to know what you eat, rather than tracking your calories, is what enables you to form such a habit and knowledge base.

Eating random stuff and thus having to use AI due to a lack of easily accessible and personally learnable nutrition information makes a total mockery of this process, and runs pretty much antithetical to it.

ulfw• 1 year ago

This goes for most "AI" things nowadays. What was once "blockchain" has become "AI" for companies now.

(PS: doesn't mean there aren't real use cases. Sadly the term AI has been stretched to cover all kinds of charlatanery)

bandrami• 1 year ago

AI aside we don't really know how many calories a given person will get from eating a given piece of food. We measure food's calories by burning it, but we don't get nearly all that energy from dissolving it (just consider the fact that solid human waste burns, for instance). At best we've established some (fairly broad) minima and maxima.

bitdivision• 1 year ago

Presumably the recommended daily calorie intake figures (and the calculators for these) do to some extent take this into account? i.e. I would imagine they're based on experimentally tested values rather than theoretical energy expenditure.

matkoniecz2• 1 year ago

"We measure food's calories by burning it" - are you sure that how it is done?

Humans (unlike say cows) do not digest parts of what is available in plants, which makes some vegetables filling and ultra-low caloric.

"we don't really know how many calories a given person will get from eating a given piece of food"

exact numbers of unknown, but difference is not so great, except truly extreme cases (and vast majority of fat people is not fat due to them)

neither_color• 1 year ago

add to that food calorie labels arent required to be 100% accurate, there's an acceptable range like +/- 20% iirc. That's enough to throw off a precise deficit. Calorie counting is more about directionally helping you lower your food consumption but it isnt as accurate as people think.

zeta01342• 1 year ago

Honestly the main thing was learning that perfectly satisfying food I cook at home tends to land around 600-800 cal for a meal, where eating out almost anywhere is often twice that, even if my beverage is water. Just being aware how how heavy those otherwise delightful meals are gives me pause. Granted these figures are all a side effect of my personal eating habits and preferences, but that's rather the point of the exercise isn't it? There's a lot of individual variation, and the individual is probably best served by a personalized approach.

ap99• 1 year ago

We might know the inner workings of the "machine" but we can control inputs (food) and measure outputs (weight).

Let's say I eat 3 hamburgers per day and gain ~1kg per month.

If I reduce that to 2 hamburgers per day I maintain my weight.

And when I reduce that to 1 hamburger per day I lose ~1kg per month.

That is simple. As for how I feel while eating 1 hamburger that's another question.

I can try to improve the machinery of my body by e.g exercising, eating healthier food, adjusting my macros, taking supplements, etc.

So now I can eat 1.5 hamburgers per day and I feel better and still lose 1kg per month.

mapt4• 1 year ago

We don't get it from raw bomb calorimetry ("Burn it and measure how hot it gets"). There are additional absorption coefficients applied before the nutrition label is printed. But for proteins, fats, and simple carbohydrates the coefficients are quite high (~95%), while progressively more complex carbs have a declining curve that ends in indigestible fiber. The difficulty is that the same person will digest more or less of the protein in their hamburger depending on how many fries they ate, how greasy those fries were, whether they ate lunch, and how much beer they drank that night. It's variable.

reedf12• 1 year ago

While this is strictly true - measurement of calories gives the upper bound of possible energy absorption. This makes it sufficient to calculate if TotalCaloriesIn >= TotalCaloriesOut. Obviously there are some biological complexities - but within the boundaries of the system the outcomes are constrained by the laws of thermodynamics.

atleastoptimal• 1 year ago

it’s crazy how much money you can make with a garbage app. The fact that i’m not means i must be stupid

snakeboy• 1 year ago

This feels like the same type of critique people have about code-assistant LLMs. We need to be a lot more realistic about their role as a 'next-gen autocomplete' who can write some boilerplate code, not a replacement for a mid-level software developer.

These apps seem useful to reduce the friction of entering your meal's calorie contents from 30 seconds of ingredient estimating, to 20 seconds (auto-filling main ingredients and a baseline portion estimate). As everyone points out, it's impossible for these to get any better than that (GIGO). But reducing the overhead for calorie-counting seems like a reasonable goal and a valuable service. Of course the marketing departments for these companies is going to say whatever bullshit they want; that's their prerogative...

mkfs• 1 year ago

Fat is the most calorie-dense macro at 9 per gram (protein and carb are both 4), so your accuracy is going to be heavily skewed by how accurately you can identify the fat content of the meal. The problem is that it's often in forms like oils that don't seem well-suited to AI image recognition.

boydYVR3• 1 year ago

Yeah. Relying on images of food to track calories is like ... dancing about architecture.

dack• 1 year ago

Calorie tracking is also about educating yourself about how many calories certain things are, so you can make better decisions.

Like, oil is insanely caloric and can accidentally add hundreds of calories, but it's nearly impossible to eat too many greens.

Once you learn this, then the tracking is just to keep you honest - your brain knows what to do but it lies to you when it wants to bend the rules and those little cheats add up enough to throw off the whole diet.

cainxinth• 1 year ago

You develop a sixth sense for how calorific foods are when you track calories consistently. I kept a meticulous food journal for about six months before I realized I didn’t really need to anymore because I had gotten so good at just estimating my totals.

Protip: most people underestimate the calories in alcohol

yreg2• 1 year ago

most people underestimate the calories in alcohol

Funny, I feel the other way around. I kept hearing about how much calories there are in alcohol, and then when I started calorie counting I didn't find it so high.

Like 6 shots of gin are ~550kcal and enough to get anyone pretty drunk. Unless one is a regular heavy drinker it's not that hard to once in a while budget calories during the day to be able to get a few shots when going out in the evening.

Obviously staying sober is the healthier option.

taeric• 1 year ago

Oddly, I think how you track can matter here. If you do it as a lookup of "I am eating this, and it says it was this many calories" then you are not likely to remember and develop a sense of how many calories things have. However, if you do it as "I am eating this, I think it has X calories, but looking it up I found it was Y" that can help remember.

I can try to find the studies, but basically, "guess and check" is ridiculously powerful in learning. Is part of what makes flash cards so strong. Just "ask and answer", not so much.

hammock2• 1 year ago

Extra light beer or shot 100

Light beer or wine 150

Regular beer 200

mixed drink 300

Trashy mixed drink 400

coltoneakins• 1 year ago

100% agree it is about re-educating yourself about calories in food. I wasn't happy with my weight but it was never 'obese': I am 5' 8" and weighed 187 lbs.

Back in November I started tracking calories in the app Cronometer. I lost 35+ lbs down to 151 lbs as of this morning.

Even as a 'relatively healthy' dude, I realized just how bad my perception of calories and macros in food was. So, I totally agree with this.

XorNot4• 1 year ago

This was my story too - calorie tracking made me reevaluate my whole diet. Suddenly when I couldn't "just have more" I was getting much more interested in the value of what I was having and the quality improved a lot.

jonplackett11• 1 year ago

Yeah there's like 2 or 3 foods you realise you're eating that totally mess you up. For me it was peanut butter and sausages (not together...).

Just so many calories and it's not like I even cared about either that much.

skohan• 1 year ago

I had the same experience. For me peanut butter and apples were my "healthy snack" and were accounting for a huge amount of calories per week.

const_cast2• 1 year ago

Yes, there's a lot of low-hanging fruit that is so bad that it's best to just not do it at all. Soda was the big one for me. There's really no responsible way to drink full-sugar soda, it's just too many calories. Oh, and ice cream.

testing22321• 1 year ago

This was the huge benefit of weight watchers. I saw hundreds of people utterly shocked they could eat a literal mountain of salad for the same number of “points” (calories/50) as a tiny piece of something sweet or fatty.

They genuinely had no idea, and it changed their lives.

libraryatnight• 1 year ago

I did Weight Watchers once when I was young. It didn't stick at the time, but it did give me a great deal of perspective that lead to healthy eating and exercise habits that did stick. I worked at Barnes & Noble then, and coming in and realizing that a grande frappuccino wasn't just 'unhealthy' it was all the points. I started reading labels.

Ultimately what worked is what you say, "educating...better decisions" Every diet I tried I concluded the secret sauce was just doing what I already knew - stop eating so much crap and move around more. It was hard yet liberating to get to a point where I wasn't on a diet I was just consciously choosing salads over fries, passing on desert, and not buying junk food so I just don't have it when the compulsion hits. (ADHD meds helped with compulsions, so it wasn't all iron will)

It was like getting sober in a way. I knew, but I also had to want to stop.

rurp2• 1 year ago

Yes exactly, all that's needed in practice is to learn a handful of simple heuristics and you'll be fine if you follow them. Unfortunately simple does not equate to easy, losing weight is quite hard for most people but the hard part is not knowing what to do, it's actually doing it.

I think that nearly all of the consumer weight loss industry, of which these silly AI photo apps are a small part, is an attempt to turn a simple but hard process into one that's complicated but easy. In practice such shortcuts generally don't work which is why products like these have miserably low success rates.

trod1234• 1 year ago

In many respects calorie counting is an exercise in misinformation induced stupidity, from bad science.

What people fail to realize is the calories on the label aren't the calories they are actually getting. Those on the label are the calories for the substance found by burning it in a lab. Calories absorbed are quite different, and depend on the method of preparation and personal factors. Different methods release different percentage amounts to be taken up by your body in digestion.

Additionally, the standard saying "Calories in Calories Out" only applies when you are healthy, and have no weight or medical issues. The moment you have any kind of metabolism-mediated or norepinphrine-mediated reactions (i.e. allergies, chemical exposures [pfas], etc), that paradigm fails.

If you are getting ravenously hungry or nauseous , you are starving yourself, and this can damage your metabolism, and it won't result in long-term weight loss. You lose metabolically active tissue over fat which adapts to lower BMRs.

When you are eating healthy at the proper times, you aren't getting ravenously hungry.

Anyone who has done Atkin's knows that after they transition into ketosis, they don't get hungry. They eat very high caloric foods, but they eat much less, and the carbohydrates, or excessive protein, and usually so low that the extra fat they eat doesn't get stored as fat. Not everyone can do that though because of issues with kidney, liver, or gallbladder, but the ones that can lose amazing amounts of weight effortlessly. The diet itself is also anti-inflammatory, and cholesterol isn't an issue.

This is going to shock people.

I've weighed my food out to get that daily calorie count. Here's a mindblower. I can eat 2700 calories of fats, and protein in a 60 30 ratio keeping carbs at 20g for the day, and I will on average lose 1-2 pounds a day, and this is done after the transition where water weight drop off has already plateaued (the weight loss isn't water weight, and its mostly not muscle mass either).

Each pound is 3500 calories. My BMR is supposed to be 2200 for my weight and height/body composition. How am I in what amounts to an effective 9700 calorie deficit with no exercise, and no hunger?

I've done this alongside college friends too. They see almost the exact same results. They found it a little annoying because they had to get checked out by a doctor first for those issues, and had to drink a lot more water, eat more fiber, and ensure they got the essential vitamins. Aside from that, the fat just falls off.

Calories in vs. Calories out is a lie.

Kirby644• 1 year ago

What people fail to realize is the calories on the label aren't the calories they are actually getting. Those on the label are the calories for the substance found by burning it in a lab. Calories absorbed are quite different, and depend on the method of preparation and personal factors. Different methods release different percentage amounts to be taken up by your body in digestion.

Bomb calorimetry is not how calories on a package are determined today. Individual constituent macros are used to determine calorie content, accounting for fiber and thermic effects of digestion. Calorie metrics on foods are largely accurate, when used for caloric intake reasons.

If you are getting ravenously hungry or nauseous , you are starving yourself, and this can damage your metabolism, and it won't result in long-term weight loss. You lose metabolically active tissue over fat which adapts to lower BMRs.

"Starvation mode" is a myth. If you are in too great of a deficit, sure, you will lose muscle in addition to fat, which can lower your metabolism... but it's not going to somehow damage you forever. If you weigh less, you burn fewer calories... just a fact of life.

I've weighed my food out to get that daily calorie count. Here's a mindblower. I can eat 2700 calories of fats, and protein in a 60 30 ratio keeping carbs at 20g for the day, and I will on average lose 1-2 pounds a day, and this is done after the transition where water weight drop off has already plateaued (the weight loss isn't water weight, and its mostly not muscle mass either).

Nobody is losing 1-2 pounds of weight a DAY in any sustainable way. You might be able to pull that off for a little bit, but unless you're morbidly obese that just isn't happening for any reasonable period of time.

Each pound is 3500 calories. My BMR is supposed to be 2200 for my weight and height/body composition. How am I in what amounts to an effective 9700 calorie deficit with no exercise, and no hunger?

BMR is not the same as TDEE. Most people's BMR is substantially lower than their actual caloric expenditure. It would not surprise me at all that you could lose weight at 2700cal/day if you have a BMR of 2200. You are likely in a caloric deficit.

rurban• 1 year ago

A classic tale of a founder winning over engineering. "This stock model from huggingface is good enough, no need to train it."

tomp• 1 year ago

The best summary of the situation I found on Twitter:

The most important part of marketing and PR is lying to your customers.

tantalor• 1 year ago

Not hotdog

blitzar2• 1 year ago

Hotdog

ksajh• 1 year ago

Sandwich

vjulian7• 1 year ago

I use ChatGPT (4o model) to track my daily macros, specifically calories, protein, fat, and carbs. I have a prompt containing my macro targets and a list of the foods I typically eat or have at home. Throughout the day, I report my meals, for example five large eggs, two slices of sprouted wheat toast, and a cup of spinach with a pat of butter. It calculates the macros for each meal and keeps a running macros total so I can see how I’m tracking. It’s so easy—I don’t even type, I use (the overall excellent) ChatGPT’s TTS function in the iOS app.

If I eat something unexpected, I have it recalculate what I need for the rest of the day. If I were planning on 8 ounces of chicken for dinner but had an unplanned snack, it might tell me that 6 ounces will suffice now, and I now have room for more fat, so I could add some cheese. I find it quite accurate in this.

Where it loses accuracy is over time. Even when focused on just a single day, the cumulative totals can be wrong when the context window gets too large. I catch this by mentally adding up the numbers it spits out. If the numbers seem off, I open a new chat. Also, it’s not reliable for looking back over multiple days or trying to track long-term patterns.

When I take a photo of a meal (like when eating out), it performs very well. I ask it to list the contents of the plate, then I correct any assumptions. Once corrected, the resulting macros are usually within 5% accuracy. I also tell it the quality of the ingredients and restaurant overall, which can help a bit.

cj2• 1 year ago

Do you find this easier than scanning barcodes with traditional calorie tracking apps?

jmacd• 1 year ago

How much of your daily intake comes from food that has a barcode on it?

netdevphoenix• 1 year ago

I assume a lot of fruit, fish and meat people buy has no barcode unless you only buy stuff at a supermarket

jurgenkesker2• 1 year ago

Did you verify it sometimes how accurate it is? I would be wary of it to be honest.

lysecret• 1 year ago

I'm doing the exact same thing. I have built a small app for it too but I like using the plain ai too.

echelon2• 1 year ago

I really don't get the bearish on AI folks.

One company has a tool that can replace what we professionally do.

This app would have taken person months to build, yet here it is, almost as an afterthought.

The next ten years are going to be wild. Hope everyone has a plan for leveling up and riding this tsunami.

Even if LLMs presently can't do intricate things, that doesn't matter. I know in my gut that my generous total comp job isn't going to be available in the future in its current form. In aggregate, the work we do isn't going to be as hard or as in demand as LLMs devour more of the market share and growth of the human economic field and attention economy.

The LLMs are likely going to increase demand for software, but in doing so they'll eat more of the market share.

ZIRP, Section 174, industry-wide layoffs, offshoring, and now this. Plan accordingly. I'm sure there are gradients with tremendous success to be had, but for many there will be a decrease in fitness.

hooverd2• 1 year ago

It's that extra 20%. Everybody right now can do the hardware store and rewire their house for cheaper than an electrician, yet they'll still hire one to do it. That or maybe everyone will just except things being slightly more convenient but worse.

kelstar183• 1 year ago

Is it ironic that my AI detector is ringing alarm bells at this comment (both its overall cadence, as well as it’s particularly unnecessary use of an em dash)?

MangoToupe3• 1 year ago

Is the em-dash ever "necessary"? It was beloved by nerds on the internet long before there was any whisper of LLMs on the internet. What do you think chatbots are imitating if not humans who also use it?

Iulioh6• 1 year ago

Some non-English speakers sometimes throw whatever they wrote in English into a chatbot to correct the text, just in case.

hn8726• 1 year ago

why are people insisting that an em-dash is exclusive to llms somehow, when on a mac it's super easy to type dash, en-dash and em-dash — just type dash, option + dash or shift + option + dash. I got used to defaulting to em-dash (even if incorrectly) and now I'll be categorized as an llm for that?

vjulian7• 1 year ago

Not AI-written. Personally, I love em dashes and semicolons.

beepbooptheory• 1 year ago

With stuff like this, I am always curious: ok, you send it a picture of a plate, how does it know the food was cooked in butter instead olive oil? Or either way, with 2 tbsp of fat instead of 1?

Iulioh6• 1 year ago

Oh no, it is TERRIBLY wrong

Using VTT is better but still not perfet

The only certain way to do it is measure everything with a scale

pacerwpg2• 1 year ago

Cool, it can replace myfitnesspal features from 10 years ago, but do them worse.

Revolutionary.

vjulian7• 1 year ago

No, I find that app cumbersome.

To respond to others’ comments, I tend to eat whole foods. I don’t eat TV dinner or shop at Costco or such places.

vjulian7• 1 year ago

Also, do you mind dropping the sarcasm and explaining why you believe MyFitnessPal might be superior to ChatGPT in the usage I’m suggesting?

maxehmookau• 1 year ago

I can't see any way in which AI could ever solve this from a photo.

A tablespoon of oil is 100 calories, but probably not visible in an image. Sugar in cooked meals is impossible to see.

Is this aimed at a world where every food product is individually wrapped? Because most of the world doesn't live that way.

rTX5CMRXIfFG3• 1 year ago

Calorie counting is a tedious, impractical activity for anybody who isn’t even entering a bodybuilding competition. Just go on a total deficit with high-quality foods, whatever the composition.

fastball2• 1 year ago

How do you know you are on a total deficit if you don't count calories?

rTX5CMRXIfFG3• 1 year ago

You lose weight after a week.

dsr_• 1 year ago

Surprisingly, at no point does the author write:

"This is clearly fraudulent as it does not and cannot work, so I reported it to Apple/Google."

danesparza• 1 year ago

I think it's very telling that the images were not oriented correctly. I'm wondering if this is a simpler bug than the the author is letting on.

hippari2• 1 year ago

Would be fun if someone try this on something from East Asian cuisine vs Fish and Chips. See how well it does in complex dishes.

weird_trousers• 1 year ago

It would be “eventually good” if, and only if, Vision problems were solved at 100% F1-Score. Which is absolutely not! So, to me, I expected bad and I received… bad results.

Paying for this crap is just throwing money out the window. Those startups should solve problems and them let people pay for it, for 99% accuracy, not at minimum of 66%.

Qem• 1 year ago

I wonder what are the errors bars for calorie counting, even assuming one measures food intake by weight accurately. Are the reference values in food package labelling any good, or is the variability from one batch to another and variability in absorption from one person to another so large that such calculations become an exercise of garbage-in, garbage-out?

XorNot4• 1 year ago

It doesn't matter provided it's not systemic. For anything related to diet, average over time is the much more relevant factor.

For example you could track calories by just tallying your grocery bills if you lived alone.

xutopia• 1 year ago

I just ask ChatGPT based on the weight of the food I eat and it gives me pretty accurate results.

wnorris51010• 1 year ago

Founder of https://snapcalorie.com and ex-Google AI researcher here. The article actually gets our pricing wrong, it's completely free to use SnapCalorie, the $79.99 price is a donation to support our research and help us continue providing the best tools to everyone for free.

We've published in top peer reviewed academic conferences and our algorithm (can't say the same for other apps) IS actually more accurate than people trying to visually estimate portion size, but the truth is both are quite inaccurate.

Most people won't spend the time to track, so photo logging is a faster and easier approximation. If you want more accurate logging you should use the voice logging feature and a kitchen scale (also completely free in our app).

As many mention the goal is to learn, not to tediously track every little thing. Do what's sustainable for you and helps achieve your health and fitness goals.

davidcbc7• 1 year ago

What about when people give up because you massively underestimate their calories? Or when people develop eating disorders because you massively overestimate their calories?

If the goal is to learn then accurate information is important, although I suppose it's harder to get a VC to fund.

ffsm8• 1 year ago

What about the people that would have given up on the diet without even trying because they just cba to do it "properly"?

Ultimately, there is no silver bullet in nutrition.

wnorris51010• 1 year ago

We do a lot of research on preventing eating disorders. As the article mentions we did not suggest she lose weight down to being underweight like other apps did.

Our app has the fruit and veg counters ABOVE the calorie meter on the dashboard. The reason for this is that maximization mindset (e.g. maximizing fruits and veg) is way healthier than a minimization mindset (e.g. minimizing calories or carbs).

We actually even tried to fully remove calories from the app at some point but we had a vast majority of users churn and decided it would be healthier for people who want calorie tracking to stick with our app by having it present, but requiring them to scroll past the features that promote a healthier mindset to get there. Feedback from users has been amazing that they've slowly started focusing more on fruits and veg.

gadders4• 1 year ago

The calorie counts on food packets themselves are +/- 20%, so it's inaccurate all the way down.

Some of the value also just comes from writing down what you eat, and noting the snacks, dressings, glasses of wine etc that you forget about but all add up.

jablongo3• 1 year ago

Do you use the LiDAR Scanner on the iPhone for your predictions?

wnorris51010• 1 year ago

We do! You can read about our original approach back at Google AI in our CVPR publication here:

https://openaccess.thecvf.com/content/CVPR2021/papers/Thames...

We've since come up with a much more accurate approach. That said we also try to do some advanced technique to figure out what people were trying to log and not what the amount of food in the photo was. So for example if take a bite out of a bagel, we'll try to say 1 bagel since it's unlikely you wanted to log a bagel less a bite.

haliskerbas• 1 year ago

How do you deal with sauces / toppings that can be widely variable in macronutrients?

JohnMakin• 1 year ago

This is the classic refrain product owners counter back with any time people present reasonable criticism of their AI app.

User: "This (AI product) doesn't work!"

Product Owner: "Well, humans are also bad at that."

That's not the promise of these apps in general! The whole selling point of AI is that they're vastly better - if my eyeball estimate is "pretty inaccurate" and by your own admission the app is "pretty inaccurate" then why the hell would I use your app??

From the very top of the page you linked:

SnapCalorie is the first app where you can take a picture of any meal and get an accurate calorie count and nutrition in seconds

(emphasis mine)

kevlened• 1 year ago

The whole selling point of AI is that they're vastly better

Not to nitpick, but "better" isn't on a single axis. Taking a photo is a better experience than searching and keying in every component, even if the accuracy is identical.

AI tools don't even have to be "vastly" better. They could be, and often are, even worse on several axes, because they often trade quality for ease-of-use.

You may assign a lower weight to the ease-of-use axis (as engineers, we tend to), but then you're likely not the ideal customer for many of today's AI products.

wnorris51010• 1 year ago

I got into this space because I saw many family members struggling with weight and nutrition. One of the leading causes of death in the United States is Diabetes, largely caused by poor nutrition and lack of exercise.

I've spent the last 10 years of my life researching why this is and how I can help. What I've heard from countless users is that it takes to long to track what they're eating and learn how they can improve.

I've never tried to claim that the selling point of AI is that it's "vastly better", I've just tried to build tools to help people. The voice note feature in our app accompanied by a kitchen scale is the most accurate thing you can do to track nutrition and it actually is much faster than what existed on the market before we launched.

The photo logging feature is more accurate than what most users did before we launched and is by far the fastest way you can track, but yes, it has it's limitations, and unlike our competitors I won't pretend it's perfect. If you're eating out at dinner, and the alternative would be you didn't log the meal, it's a great option.

At the end of the day accuracy actually is not the most important thing if you really care about helping people. Education is what matters. People need to learn which foods and ingredients are the problematic ones and why. Our app accomplishes that better than any solution that existed on the market before we launched.

ericmcer• 1 year ago

Thanks for mentioning voice, I prefer that to photos. Photos seems like the most complex technically.

AI seems to be going heavy into text/chat and images but voice is really the perfect UI to me. I would be totally fine if I could trigger a "call" from the app and talk to an AI on the phone for 30s and it could log my meal, give me some advice, etc.

hammyhavoc3• 1 year ago

Do you talk to your voice assistant on your devices as-is? Do you do it in the company of others?

With the exception of a few legally blind folks I know, I don't know anyone who openly talks to a machine, nor anyone who does it in private.

malfist4• 1 year ago

How can your AI hope to be accurate. Visually, Can you tell how much butter or oil someone used in the recipe? Visually, Can you tell if it's 73/36 ground beef or 97/3? Can you tell if that taco meet is beef or turkey? Can you tell if that glass of milk is skim or actually just heavy cream?

You can't tell if I cooked my scrambled eggs in the rendered fat from the sausage I cooked just before it, or if I used a teaspoon of olive oil

There's too many sources of calories you just can't see at all. I made jambalaya last night. If I'd sauteed the veggies in a tablespoon of oil or a cup of oil you wouldn't be able to tell.

mminer237• 1 year ago

You can just assume an average amount of butter for such a recipe, 80/20 beef, beef, and 2% milk and be reasonably accurate, especially for people who otherwise wouldn't have any idea of their calorie intakes.

MacNCheese23• 1 year ago

This is without a doubt the worst App I installed in the last years. It is predatory and has the worst user experience I have ever seen. I'm serious.

It begins with that i haven't even tried out anything, before logging in - that the google store popup comes up that I should rate the App - on the same screen you even tell users please rate us - HOW?? HOW should i rate something now that i havent even used yet.

Then, like in the 20th century i have to supply an e-mail/password and wait for a OTP instead of just using the Google EcoSystem and let me login with my existing Google Account.

After that, i went through at least 10 pages of bullshit were I have to give my information, weight/height/metabolic rate etc - even adjusting the calorie deficit didnt work outright if it is 0 and you want to put a minus before it it just goes away. You first have to enter a postive > 0 value to be able to ad a Minus.

Mind you, i just want to see what that App is about what can it do - i don't want to enter and click next next for a gazillion informations first after being annoyed by "please rate us".

Then, after completing the "setup" i get a fullscreen overlay to annoy me about PREMIUM membership - AGAIN I HAVE NOT USED ONE SINGLE FUNCTION OF YOUR APP YET.

But not enough, I want to try it out with a photo from my meal today and on the camera lense screen it annoys me again with PREMIUM!!! You can only do 3 Photos a day. Dude your app is so predatory its just trash.

Then i select the photo from the gallery and have to CROP it??? Its a fullscreen picture of just the meal - nothing else and you annoy me with cropping it?

Then, the actual function - while it recognized my meal it was off by the calorie count and weight by a factor of 2. The same picture, for free, google gemini flash recognized correctly too and gave an accurate display of average calories (it approx. 450g). I added the exact weight to gemini afterwards to have the correct estimation and it gave me the real value.

I tried to do the same in your app, instead of 300g i edited the entry and choose 600g. And clicked save. I was puzzled that nothing changed. Then i saw the portion size went from 1 to 0.5 - just by me changing from 300g to 600g. What a stupid user experience. So i had to change the weight AND return back the portion size to 1.

Again, everything your App does apart from massively annoying users, gemini and chatgpt can do better for free.

mgraczyk3• 1 year ago

What do you mean by this?

it's completely free to use SnapCalorie, the $79.99 price is a donation to support our research

I just downloaded the app and I don't see any way to use the premium plan without paying 79.99 or 6.67/month. How do I choose not to donate?

Or do you mean it's a required donation? And if that's true, in what sense is the article wrong, given that it says exactly that?

wnorris51010• 1 year ago

You have access to all of the same features without upgrading to premium. Just click the "x" on the paywall.

yqiang• 1 year ago

I'm working on an app in this area [1], and I've spent a lot of time exploring how to responsibly use AI for food tracking.

My conclusion is that while AI is excellent for augmenting your tracking experience, it's not yet reliable enough to be the primary tracking method. Consistency is key to successful food tracking, and AI can certainly help users avoid the common issue of missing a meal and losing momentum. However, inaccuracies, like consistently being off by 100-200 calories per day, can significantly impact results, especially for those on lower-calorie diets (like 1,200-1,500 calories/day, which is common for many women due to their physical size).

With FitBee I landed on communicating to the user that these are estimates and you probably shouldn't use it as your primary method of tracking calories.

[1] https://fitbee.app

jablongo3• 1 year ago

The article claims that none of these apps use "depth analysis", but newer iPhones have this capability. I would guess at least some of these apps are using some kind of volumetric analysis for the food when available.

jraines• 1 year ago

The benefit of AI as a feature in calorie counting apps is great. But you have to track stuff manually for a while to realize the benefit. Specifically: using AI for restaurant meals is a convenient starting point, from which you use your hard-won feel for weight & caloric density of things to tweak the first pass to something closer to the mark.

As of now I would not at all trust a fully AI calorie counting app

marcellus23• 1 year ago

This is exactly correct. I find AI in my food tracking app (FoodNoms) to be really useful, but no, you cannot rely on it on its own. You always have to tweak the results to match what you feel in your gut (no pun intended) the counts should be. Interestingly, very similar situation to using AI for code: you shouldn't just blindly trust what it puts out, but if you know what you're doing, it can save you a lot of time.

wnorris51010• 1 year ago

100%. You should not trust photo based calorie tracking and should only use it when there is no other realistic choice (e.g. eating out).

You especially should not trust an app that does not do any research on the accuracy of it's algorithm or claims to be "consistently accurate" as many of them do currently. The truth is these algorithms have the potential to be accurate when values are averaged out over the long term, but dish to dish they will have occasionally wildly inaccurate results.

davidcbc7• 1 year ago

You especially should not trust an app that does not do any research on the accuracy of it's algorithm or claims to be "consistently accurate" as many of them do currently.

Your website says:

SnapCalorie is the first app where you can take a picture of any meal and get an accurate calorie count and nutrition in seconds.

So which is it? Can you get an accurate calorie count of a meal from a picture or not?

cochne• 1 year ago

I started using an AI calorie tracking app. For non-branded foods, the calorie counting is pretty inaccurate and sometimes very off. However, having an app that can track pictures and create short descriptions that someone else (my trainer) can easily review has been pretty helpful.

jablongo3• 1 year ago

Establishing ground truth for this is not easy. Often the labeled calories on foods are quite inaccurate themselves, based on n=1 bomb calorimetry tests. There are also incentives that may lead to lower than actual reported calories on the label.

homefree• 1 year ago

I suspect none of this stuff will matter in the near term. People will just take ozempic or whatever the best iteration is and be done with it.

I’ve lost substantial weight (75lbs) the hard way multiple times, it’s doable but it sucks. If the drugs make it easy with seemingly no downside then why waste time with this other stuff.

Use the drugs, lose the weight, then when thin add in an exercise regimen if you don’t have one.

sudobash1• 1 year ago

I expect that plenty of these apps are or will be mining these photos for data about their users. I know the cat is already out of the bag as far as online tracking goes, but until now, tracking every meal & snack someone eats has been impossible (except perhaps for the most consummate doordash user).

And the companies most interested in the data would be the large food corporations who bear a lot of responsibility for getting us into this obesity crisis in the first place.

foxyv• 1 year ago

Calorie counting is so unreliable it hardly even matters. The calorie measurements are way off. There are so many sources of macronutrients and they are all digested differently and lead to so many different amounts of "Fuel" for the body. Just the different types of carbohydrates are bad enough. An actually accurate food label would look like the spec sheet for a 5th generation fighter jet.

This doesn't even take into account the variations in the human body. Some people digest different types of foods better than others. Then there are the gut microbiomes of each individual. By the time you get done, if you have counted 2000 calories you could be off by 25% or more.

The big advantage that makes calorie counting work is that it makes people actually think about what they are eating.

boydYVR3• 1 year ago

Yes, all bodies process food differently, and all measurement tools are inaccurate. One of the bigger benefits of tracking is the ability to see trends. If you're chronically low or high on a nutrient, a good tool will make this apparent to the user.

alis0nlaura• 1 year ago

SnapCalorie is a very useful tool, and not just for calorie counting but for macros and solving for nutrient deficiency.

boydYVR3• 1 year ago

I've felt the same way for years. I used one of the big names for a month about 10 years ago, and it was a great way to learn about my habits and diet, and to experiment with IF (intermittent fasting). I tried it again, now that I'm a serial entrepreneur and single parent... and I just did NOT have the patience for the data entry. I feel same as this article that photos are not a solution. Vodka and water look the same to a camera.

Since April I've been using a tool I developed for myself. It's very much a MVP with an emphasis on M, but it's cheap to run, so I'll share the url if anyone wants to try it out and share feedback.

https://words.my

YeGoblynQueenne• 1 year ago

> Maybe the old-fashioned approach of listening to our bodies works better than any algorithm.

My body is perfectly capable of eating an entire jar of peanut butter with an entire jar of honey on the side in one sitting, so, no.

I used AI-powered calorie counting apps, and they were even worse than expected

Related Stories