Hacker News Reader

Comment by wnorris510 | Hacker News Reader

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.

View on HN

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)

> 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.

^^ 100%. Accuracy, speed, weight loss efficacy. Improving on each of this is a different solution.

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.

What is the added value of the app? Ive been using ChatGPT to track my calories for 6 months now and dropped 20kg because of it. This is mostly me just telling it everything that enters my mouth and adjust the output I deem wrong. This isn't super accurate but apparently enough to get an idea, and thus lose weight.

Our app uses LiDAR for portion and is a lot more accurate than ChatGPT, ChatGPT doesn't have access to LiDAR and is less accurate.

Our app is free to use the most accurate model, ChatGPT is paid for the most accurate model and has usage limits.

Our app plots your data on nice charts and has visuals specific to understanding your data over time visually (e.g. how has my sodium intake been changing day over day or what has my average weekly caloric intake been when deducting active calories from fitness trackers).

Our app integrates with Apple Health and soon will integrate with Android Fit so you can export your data to other apps, share with health professionals, etc.

Once Meta, Google and Apple put this in their current to future smart glasses adoption will sky rocket .. it will done automagically and then that knowledge might shock people how many calories they consume vs. the recommended daily intake. It could be made a game with you and your friends and or family or workout buddies.

Ozempic be damned for at least 1/4 on it and or considering it.

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.

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.

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.

There may not be a silver bullet but adding randomly generated numbers is not an improvement

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.

You are still selling a product that says it can count calories when fundamentally it cannot. The fact that people believe you and pay you money for it doesn't change that fact. You are lying to people.

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

This is from your website. It is pure fiction. You admit as much in your first post

We published in the top conference of computer vision a peer reviewed study of our accuracy. On average we are twice the accuracy of a professional nutritionist. The product is free to try. If you're not happy with the accuracy, don't pay! We have around half a million people each month coming to our app and a vast majority stick around and are happy with the results.

If there's something we missed on your food shoot me a DM, I'd love to dig in.

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.

+/- 20% is way better than being off by 3x like in the example in this article

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.

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.

What is "an average amount of butter"?

And why can you assume any of the other things?

Probably because most people likely to use this app aren't going to be using it for serious self-monitoring like someone with diabetes would. The reality is that most people will use an app like this for a while, fail at their attempt to lose weight or gain muscle, and lose interest. "Good enough" is the mantra of much of what is in the health/nutrition/fitness niche and it goes beyond anything AI-adjacent.

Good enough isn't close to cutting it for weight loss. A 250 calorie surplus per day will add half a pound per week. That being off by two tablespoons of oil.

"Good enough" being that it doesn't matter because most people who use it probably aren't going to seriously stick it out to see results as they either won't put the work in or won't seriously stick to a diet for any meaningful length of time.

I do agree though, it isn't fit for purpose, like much of what's available beyond AI.

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?

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

The paywall says I get extra stuff from premium, including "unlimited photo, voice, and barcode logging", "unlimited AI nutritionist chat", "unlimited Personalized tips from your AI nutritionist".

It sounds like if I pay more, I get more access and can use the app more? So the original article was completely correct in characterizing the premium plan as $79.99

It says

> The app was free to download, no trial period necessary. There's a $79.99 per year premium plan, but it's intended to be a donation. The app caps free tier users at three photos per day, while all non-photo methods of logging are unlimited and free for everyone.

Which part of that is not correct?

Apple requires us to say you get more as part of premium. They have silly rules that tie our hands on stuff. You have access to everything as part of the free tier. Hopefully the recent court case from Riot will help make it easier for us to clarify this.

I just tried in the app on Android and it would not allow me to take more photos unless I pay.

When you say "you have access to everything as part of the free tier", does that apply to Android as well?

Because what you are saying is not true on Android. After 3 photos I cannot take another, the message says "You used all of your photo logs! As a free tier member you get 3 free Photo logs per day."

You are claiming that this is different on iOS?

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.

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.

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

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.

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

Reply on news.ycombinator.com