Photo calorie app Cal AI was built by two teenagers
techcrunch.com
techcrunch.com
I'm all for making it easier for people to lose weight but this app may honestly have the reverse effect. If the app estimates calories too low (and therefore the individual eats more), many people will get frustrated with the lack of progress and give up. If the app estimates too high, the individual will lose weight, but diet fatigue and other negative side effects of being at a >500 calorie deficit may make the diet too difficult to maintain.
I pick this example because I've seen specifically this cause problems for people trying to lose weight. They think their eating a salad, not realizing they've thrown an extra 500 calories on top.
Another case: I sit down to breakfast, having made myself eggs and toast. One of if not the largest contributor to my calorie intake will be the amount of butter on my toast. If I use four pats that will probably exceed my calorie intake from eggs. If I use one, not as much. I sincerely doubt it's realistic to tell the difference with any sort of precision.
Wait, ARE people eating that? Have I been out of the game too long?
Any other interpretive situation based solely on a camera has so many inherent flaws as to render this almost useless.
Critiquing: You still need to figure out serving sizes - it's going to need to ask how many servings. Nutrition labels also aren't available for any number of things.
If you have a really dysregulated metabolism, your body can definitely work against you when consuming too little.
That of course feels like a "weird" edge case, but it illustrates the general problem that butter/oil/sugars can pack a lot of calories and have no or almost no visual signature.
A salad wouldn't even be the hardest case to detect, since raw vegetables don't soak up as much oil as other kinds of food.
There's a sweet spot for an app that is inaccurate with a market that wants it but doesn't understand how inaccurate it is.
Kind of like how I could vibe code an app, get it to "work", think it's great and be ignorant of the many ways it will break or isn't working that a knowledgeable developer could.
The fact that people still believe in reducing fat as it's own goal (instead of being an easy way to reduce calorie content) is a testament to how bad the public is at identifying fact from fiction.
Then you have shit like the influencer foods, "Feastables" and "Hydration beverage" Prime, which is just flat soda. It's pathetic.
Or think of all the dude bros who insist on dry scooping cup fulls of protein+caffeine powder, and going home to gorge themselves on two pounds of chicken breast, and yet doing absurdly normal amounts of weightlifting or exercise that requires no modifications to their already protein overloaded American diet.
Diet culture is what is fucking American health. People read fucking tabloids that bad-faith regurgitate poorly done "science", funded by the council for selling more food, and insist that since "Woman's Health" says that scientists say chocolate both kills and saves you, scientists are dumb and know nothing, even though THE ACTUAL SCIENCE NEVER CONCLUDED ANYTHING, because the scientific paper was just an observational study!
It doesn't help in how far some foods have been bred in the past century and a half in particular. Or how different people with different genetics may react to certain foods.
You could level the same criticism at Cronometer and MacroFactor when you try to log food you received at a restaurant. Yet those apps are still useful (and I think requisite) for knowing what you're eating. And you should probably 1.5x the calorie estimation when you eat out.
What's interesting is whether this app can accurately estimate food at all. If it can, then that's a huge win and you can add your own buffer zone for oils like you already have to do when you count calories. ...Or chill on the butter and restaurant food when you're supposedly trying to lose weight.
Not really. In practice you need to know the ingredients to estimate the caloric value. Either because you prepared the meal, or because someone who wrote the recipe of it calculated and wrote it on the packaging/menu.
> If it can, then that's a huge win
But that's the point of the example. That it can't. If it could, that would be good. But it can't do it, and not because the app is deficient in some way, but because the necessary information is not available in the image.
> Or chill on the butter and restaurant food when you're supposedly trying to lose weight.
Yes of course. And that is one of the things you learn when you do calorie counting. The practice drives home that message, and many others. But you are not going to learn that if the app hides the signal from you.
Spoiler: It can't. It is physically impossible to determine calories from pictures of food.
This was technically correct but missed out on a viral app and millions of revenue?
1. If the food being scanned has a nutrition label, I don't need this app.
2. If there's no nutrition label, the app can't possibly extrapolate the nutritional information. It can't estimate portion size or the ingredients in the recipe.
3. If the app can't extrapolate nutritional information, I don't need this app.
4. I don't need this app.
What you need is a LLM.
I am happy with a rough number on a dish, I do not need precision to know if I am not eating enough calories. Note there are errors on the result and reasoning, but still this dish should be around the ballpark of 500 cal. I figure running this through a larger model might be more accurate. I wondered if maybe I wanted to create an app for that, but I am happy not to.
Yes, simple descriptions could be parsed with NLP, and yes, calories could be inferred with a database.
>>> Calorie count of a dish that has: rice, broccoli, two pieces of turkey, and a fri
... ed egg. Assume this is a normal portion, for instance, there are 80g of rice.. Re
... turn the result in a json structure
<think>
Alright, so I need to figure out the calorie count for a dish that includes
rice, broccoli, two pieces of turkey, and a fried egg. The user mentioned it's
a normal portion, like 80g of rice. Hmm, okay, let me break this down step by
step.
First, I should find out how many calories are in each ingredient separately.
Let's start with the rice. Rice is a common staple, so I think 80g of cooked
rice has about 130 calories. I've seen that number before, but I should
double-check to make sure it's accurate.
Next up is broccoli. Broccoli is pretty low in calories, right? A portion of
around 100g should be about 25-30 calories. Wait, maybe a bit more if it's raw
or cooked differently, but I think for this purpose, 35 calories is reasonable.
Then there are two pieces of turkey. Turkey can vary based on the cut and
whether it's cooked. Let's assume these are regular chicken breast slices, each
around 50g. Each slice would have about 110 calories, so two pieces would be
around 220 calories. But I should make sure that I'm not confusing turkey with
another type of meat. Turkey does tend to be a bit higher in calories than
chicken sometimes, but 110 per slice seems right.
Lastly, the fried egg. A single fried egg is usually about 140-150 calories
depending on how it's made. Since there's just one egg, that's around 140
calories. I think that's correct because yolks add a good number of calories
due to fat content.
Now, adding all these up: rice (130) + broccoli (35) + turkey (220) + egg
(140). Let me sum them: 130 plus 35 is 165, plus 220 makes 385, and then plus
140 gives 525. So the total would be around 525 calories.
Wait a second, maybe I should consider if there are any other ingredients or
whether portions were different. For instance, sometimes turkey portions can be
larger or smaller depending on the meal. But since it's specified as two
pieces, each at about 50g, that seems okay.
Also, considering preparation methods: fried egg might have more calories if
it's done in a certain way, but I think one fried egg is standard. Broccoli's
calorie count could vary slightly, but for simplicity, the 35 seems fine.
Is there any chance I'm missing something? Maybe adding spices or sauces, but
unless specified, I shouldn't assume that. So the total should be around 525
calories in a normal portion as described.
</think>
```json
{
"calories": {
"rice": 130,
"broccoli": 35,
"turkey": 220,
"fried egg": 140,
"total": 525
}
}
```Deleted comment
On other tries of the same prompt it happily adds oil, but miscounts other parts
In packaged foods, there is a whole science of masking the sugar and fat content to make it more addictive without triggering your inbuilt satiety mechanisms [1]. This is what today's engagement optimisers did for money in the 50/60s.
You could argue that these "innovations" were precisely to subvert the intuition that visual appearance of food (and other natural sensors) can be relied on to assess their nutritional properties.
[1] https://en.m.wikipedia.org/wiki/Bliss_point_(food)#:~:text=T...
God I wish that was true. Butter is far too expensive to do that, so restaurants will use the cheapest alternative (usually soybean oil with butter flavoring) instead.
I can make two dishes that look identical and have +/- 50% caloric content, easily.
This doesn't pass even simple scrutiny. There are so many caloric ingredients that aren't visible in food. You can't tell just by looking whether a rice dish contains half a stick of butter.
They claim 90% accuracy, whatever that means, but I have my doubts regarding it's usefulness.
You'd have to be kind of stupid to expect it to actually be 100% accurate for all meals
Without knowing the amount of sugar, butter, oil, etc. is used in a dish, one cannot know if a dish is worth 250 kilocalories or 750 kilocalories.
If I need to manually fill in details of ingredients and amounts to get to the calories to be have an error margin of less than 100%, then the app is not useful and is at best misleading.
4. I don't need this app.
1 tbsp of olive oil has 135 calories.
How would the app know which fat the food was cooked in?
This is extremely false. Please verify your sources better (and apply a skosh of critical thinking).
> 1 tbsp of olive oil has 135 calories.
This is false too, but at least it's in the right ballpark.
Sorry! I was using Cal AI
As we understand more about brain development in kids, I think and under studied aspect are kids who got access to a lot more money than normal typical kids have, and the results such wealth brings.
As opposed to the kids born on planets without atmospheric oxygen, of course. Those kids don't stand a chance.
There's always some advantage others have. Money, good looks, distribution, connections, right place right time.
"Wealth inequality" isn't going to disappear overnight, and lamenting about it won't get us closer to success.
We're all dropped into the Darwinian gradient landscape. Some of us have better starting positions. While we sit and wait for policy decisions to make things easier, our job remains to find gradients that aren't too steep, aren't over-explored by the masses, and that give us some modicum of joy to spend our lives upon.
The minnows and gazelles have it much worse than us. Praying mantises, anglerfish, and the hosts for the parasitoid wasps are practically living in a daily Kafkaesque horror. Meanwhile we're in our own dopamine drip Disneyland with near-infinite degrees of freedom and plenty of years on the clock (for most of us).
Seneca said some good things about this.
I'm not lamenting wealth inequality, nor discussing other advantages (real or perceived).
I think that achieving success and wealth (e.g. not simply being born with it) has an impact on the development of a human brain, based in large part the behavior of rich/famous/successful young adults, AND what their personalities ultimately become once the have fully formed brains (~27 years old).
Specific to these teenagers, how will effectively becoming "rich" and successful at such a young age change the final formation of their brains, and ultimately shape their behavior?
Like Theranos. And Nikola. And Fyre Festival. Etc.
Age is irrelevant, unless we're trying to tell youngsters that fraud is an acceptable means to get ahead in life. But then again given our current political environment maybe that is what we're trying to tell the next generation.
EDIT: the app now also just reads nutrition labels as a backstop. Nutrition labels already include the calorie data, so the app isn't doing anything there.
There's no trying.
The US has played "Just grift your way to money" as a standard means of operating since at least the dotcom boom. A reason so many young and poorly educated people jump at obvious MLM type scams and other scams is because they feel that everything is a scam, so surely they can get in on it too, right?
And everything IS a scam. Coke doubled in price in the past 5 years. I promise you their costs did not double. Their costs are Labor (highly automated), water (they almost always have sweetheart deals for dirt cheap water, cheaper than you pay), and one of the most subsidized commodities available, HFCS, or alternatively, a sprinkle of dirt cheap chemicals for their diet sodas.
People feel that, even when they don't understand or even recognize it.
People recognize that the US has been a scammy free for all for decades now. Everyone for themselves, fuck you, got mine.
We are on like the third generation raised this way. The people who took "Greed is good" to heart had kids, and raised them with it as a core principle.
Jordan Belfort, the guy who Wolf Of Wall Street is based on and spent time in prison for scamming his clients in basically the same way modern crypto pump and dumps work, now sells out auditoriums as a motivational speaker for fucks sake.
The kids LOOK UP TO SCAMMERS
“You can have a full self driving car with just a few cameras”
In a way both things are very much similar and the real accuracy is more of a fiction than reality.
That's unlikely. Try driving in a snowstorm, where visual inputs become effectively useless, and you quickly realize how much the motion inputs are factored in as well.
But with the right sensors you can figure out the chemical composition of food too, so...
Lol, I can relate. I started working in an office when I was 16, now 24, and regretting wasting my youth grinding when I could have been having fun in a period of your life you only get to experience once.
Don't grow up too fast kids. Make stupid decisions and ride out your youth as long as you can afford to.
As you get older, people expect you to be more competent in life and work. If you leave work for two years to bike around the world, it'll be a fabulous adventure and will in the grand scheme of things have little consequence down the road. Try that when you have a kid!
I made a new friend at a hackathon recently. Since you're here on hacker news you could try attending events such as programming meetups, hackathons, tech workshops, conventions etc. It's easier to bond when there are mutual interests.
Otherwise try joining some sports related club, e.g. a local gymnastics or football club or even just a jogging group.
It is different but there are people who want to hang out, even if in the context of doing some other activity.
But a lot of people sit around and have zero social hobbies and claim it’s hard to meet people in adulthood.
Its just a marketing stunt. They complain about not getting into Harvard but he could literally rent a penthouse in Cambridge, hang out on the campus throw big parties and get 90% of the experience while supposedly running a 30M ARR company. Absolutely no benefit to actually enrolling except ego.
One note, as someone who also built a calorie tracking app with ai as well as lost a good amount of weight with it: accuracy for calorie tracking doesn’t matter. You can honestly just detect if it’s a meal and return 600 cals. For most people the simple fact that they become aware of what they eat and think about their food in an ongoing basis will lead them to loose Weight. Sticking to it is the hard part.
That's ridiculous. What signal would that provide to the user? Let's say someone who is eating double the portions they should be eating. How will this hypothetical app help them figure that out?
> For most people the simple fact that they become aware of what they eat and think about their food in an ongoing basis will lead them to loose Weight.
And you think people will " become aware of what they eat" by shooting a picture of their food with an app which always say "600 cal"? I don't think you thought this through.
Take a picture of everything you eat and correlate it with symptoms. Have AI figure out what may be a trigger.
(I have a super rare food disease that took years to figure out and made my life unbearable).
That being said, I'm not going to renew next year. But there is something to this product that is not the "AI" but the simplicity vs. MyFitnessPal which has a ton of features I don't really need.
The critiques are good, but for me, the simplicity of the app is the most attractive part for me.
When I learnt machine learning one of the things was continually training the model. Like your spam filter. You show it what is spam and eventually it learns. Is this stuff continually trained on the user's BMI? That's the only way to tell if a diet is working. Or is it just making absolute claims based on universal training data?
To me, his college list indicates that he was mostly prestige hunting. There's nothing inherently wrong with that, but colleges can tell when a student wants to attend based just on branding. It comes across as if he wants to use college as a resume booster rather than as a place to grow.
The essay reads as a list of accomplishments, with little self-reflection. (Side note: referencing Steve Jobs is way overdone.)
Long story short, college admissions is not a VC pitch. If I had been this guy's advisor, I would have recommended he write an essay about something not related to Cal AI. Colleges will already know about the app from his activities list (and resume and, presumably, recommendation letters). There's a huge missed opportunity for him to write about something else.
The essays that worked for my students this year were often about more mundane topics that gave insight into their character. One of my favorites was from a student who started giving free haircuts to classmates. The essay implicitly shows that he's thoughtful and well-liked—someone you'd definitely want in your college community.
I'm Polish, here the only thing that matters is your final test scores, and nothing else. And I think it's same in the most of Europe and Asia too, right?
My impression is that American unis care way more about social aspect and so on, which I don't understand (but I guess it's a fine way of looking at things, too.)
It's true that this model is more fair, and that's good, but it still feels wrong. There are way too many professions where you're de-facto locked out if you didn't get the right credentials at the right age, regardless of your practical skills. That results in us putting teenagers through these absurd trials for no real reason.
The only thing that distinguishes applicants is the soft social stuff.
Japan and South Korea kind of fixed this problem with cram schools and ridiculously overtuned college admission exams. But e.g. KAIST isn’t really comparable to MIT.
Or is it that way because of some other factors? I was thinking how much of this is because of historical factors; I assume in times before standarised exams it would be a very convinient way of finding new students. But then, I don't know how it was historically in Europe/Asia.
The tests need to be harder, but people would complain.
I didn't study for the ACT at all (literally went in without knowing anything about it) and got a 35. It's a trivial exam.
I got a 1040 on the SAT in the 5th grade. The average score is useless for gauging how hard these exams are.
Bragging how you got a 34 on the ACT or 1450+ on the SAT for an elite college is like bragging about clubbing a seal
Bar making exams harder, the only other way is subjective methods. I detest subjective methods, but making exams harder is very unpopular
When you are locked in and have the grindset there is nothing else.
I would agree, talking about actual human stuff related to an actually interesting topic was a wasted oportunity. Nobody actually cares what the numbers on the app are, least of all admissions officers.
Average is quite a bit above the floor though, so that just makes it sound like he should have been accepted.
The question you need to be asking is how the university system made an enemy out of someone who is clearly one of the most talented members of his age cohort in the nation. That's a failure no matter how hard you try to explain or justify the status quo. It's time for some real accountability and soul searching from the system, not excuses. Trying to nit pick the essay and pointing out how he should have done X or Y instead is completely missing the point.
I can't understand why the admissions officers would rather read an essay about a kid who volunteered at an animal shelter or something. Anyone can do that.
Joe, a regular guy: Makes $120k at his desk job
Joe, the businessman: Made $20k in 32 days, $228k ARR
Joe, who launched 5 months into development and did 60k in the first 2 weeks: $1.5M ARR
In all three of these examples, Joe's financial outcome is the same. This business does not have any longevity, and all of its revenue is from converting paid advertising of various kinds. It's still impressive, but is most likely a >10x exaggeration on even the lifetime revenue he makes from this. Which is of course circular, because the reason he's doing all this is to make a business out of monetising the audience of people who want to make money.
All of this is clever social climbing, but is clever social climbing the thing that should be rewarded by colleges?
The app is fake - at best its puffery, and the essay was littered with grammatical errors.
I’m also quite sure letters matter for undergraduate admissions. They certainly do at the graduate level.
Yeah. That's how it works. When you do community building and participate in activities in addition to "the grind", people like you more.
This doesn't just apply to academia.
Unlikely
My conclusion is that while AI is excellent for augmenting your tracking experience, it's not yet reliable enough to be the sole 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://apps.apple.com/us/app/fitbee-calorie-macro-counter/i...
Like a small vision model combined with the size/measurements data from the AR sensors modern phones come with and an open source caloric values database should achieve the 90% accuracy they are claiming.
Ronald Wright writes about "progress traps" in A Short History of Progress. It's been awhile since I read that but I think about it more and more these days with AI products on the rise.
You cannot differentiate a high calorie meal from a low calorie meal on sight alone.
The waste is selling a lie, enabled by AI bullshit artists and the public's seeming inability to understand that the US has no legal (or market most of the time) requirements to be truthful, upfront, or honest in marketing.
Like people just take this shit at face value and I don't understand how you can live in the US for more than a few years and not recognize that marketing is just lies, like not even smart or clever lies.
App Quality and Luck are not the only two factors that go into App Success.
Another thing I noticed is that I saw a random guy on instagram with a rather big following being sponsored by Cal AI. Maybe your friend was unsuccessful in getting his app out there? Although I agree that luck will always play a role, but if the public don't understand what your app immediately does and they believe AI to be pure magic, then sprinkling that everywhere will get something like Cal AI flying.
I still think it is shit from a technical perspective in terms of the validity of amount of calories from a single image and nothing else. But it seems like that's not what people want, inherently because they are lazy. Actually counting calories is much harder long term. If regular people now think that this is magically replacing this process by just snapping a picture of their processed meal, then I can see why it's successful. Although quite depressing...
If you want to derive any benefit from doing this you should really be trying to get your numbers correct from the start. I wouldn't leave that to a LLM.
If I could, I am not sure that I would. This app seems actively harmful. I don't think it can actually do what it claims to do, and that's going to cause real people problems.
It's unfortunate that that disqualifies me from making that kind of money. It's unfortunate that they are allowed to do so.
> The result is an app that the creators say is 90% accurate, which appears to be good enough for many dieters.
It absolutely cannot be "90% accurate". But I'm sure it seems "90% plausible" to its millions of users.
Incredible that a product like this can exist. Do people just will the fiction into reality?
Incredibly, these are the types of app ideas you'd hear from non-tech "entrepreneurs" in 2012 looking for a co-founder. The problem being, the engineers knew it was impossible. You could fake it I guess by asking Google to search for "similar photos" and getting a plausible calorie count half the time. But the users wouldn't believe it.
We're now at the phase where any impossible idea can be fully marketable by slapping "AI" in the name. ChatGPT feels so magical that we now believe unicorns really do exist.
But it's still a great idea.
Also, they should calculate both the calories and power requirements for each meal analyzed. What I mean is, it should says something like: this burger has 800 kcal and the analysis consumed 1kw of datacenter power. ;)
No, it isn't, because it can never work. It is, in general, impossible to determine the calorie content of food by looking at a picture of it.
Define "great".
It has $2m revenue, so it's clearly a great idea (at this stage?) financially and 'people love it' (30% retention)
Technically it's a garbage idea, and I'd say they could get class-actioned without good T&Cs. It's literally impossible to determine the sugar and fat content of a meal.
I'd never make it for the latter reason, however you clearly need to believe in the former to make it big haha.
The $2m figure is unverified but, even if it was, turnover is meaningless without knowing their expenses. If they're burning through £2.1m in LLM compute and marketing then they're losing money.
Plus, the guy seems like a dick so I'm taking his story with an enormous pinch of salt.
It seems we're at a point where this obligatory xkcd [1] is no longer true.
Are you eating a 10% calorie deficit or a 10% calorie surplus? Cal AI can't tell you.
Not possible to know accurately enough from a picture. Potentially ever.
But I tried one of these apps years ago and it went a step further than photos. It used the front facing camera on iPhones to build a 3d model of the food and measure its volume as well. Even that was off by more up to 50% not 10%.
The interesting thing I found, and it’s obvious when you read it but not when you’re trying to diet, is if you don’t layer food on top of itself or other food, you (and a camera based calorie counter) will have a much better understanding of how much you’re eating. Bowls / mounds of food will deceive you.
Standard for cutting is about 500 calories deficit, for 1lbs lost a week. Lets say 2500 calories daily standard. That's 20%. If food packaging was off by 30%, food nutrition planning would be worthless, but we know it isn't because we see fairly consistent results from weightlifters (assuming they're actually weighing their food and not eyeballing/using a PoS app like this)
Maybe there's something where they're off by 30%, but how many people even track how much vitamin D they get from food?
Wrt Bangladesh... I imagine the job might be a little harder during Ramadan.
Usually if a teenage hacker builds something, the HNers would respond with enthusiasm, but then, this is a guy who builds *something* and *actually* makes a good business out of it, at the same time maintains his high school life, and all we have is skepticism and discouragement.
https://www.cnn.com/2017/04/05/us/stanford-application-black...
lol
(2) I did some work on a pre-product startup that wanted to build something like this about 10 years ago, when visual recognition was much less developed.