This is neat. I've spent a lot of time thinking about implementing something similar for my company Eat This Much, but end up pushing it off in favor of focusing on our core meal planning features.
When something doesn't have a reference listed, and just says "sourced from a publicly available first-party datasource", what does that mean? Crawled from other sources and you'd prefer not to say? The wording does feel a little sketchy when contrasted with entries that do list sources.
When something does list references that don't seem super close to the actual food, what is the process like there for interpreting those values? Example, this Chicken Salad inheriting from Chicken Spread: https://www.opennutrition.app/search/chicken-salad-37mAX17YX...
The quality of the data might feel rough now, but I can see this being valuable for our users even if it's just an opt-in "show estimated micronutrients" or something. Would require labeling values as not being directly from a source of truth.
One thing that a lot of people are missing is that there is already a lot of inaccurate nutrition data out there. Even on information directly from the manufacturer, sometimes there are errors, or just old versions of the product that never get scrubbed from the internet (I imagine the latter case would be tricky for an LLM to deal with too). Just logging your dietary intake in any form will get you 80% of the benefit of tracking via some self awareness of your intake. Of course, it's an easy argument to point out that if you had the choice between verified data and fuzzy LLM data, you should go for the human verified data (for now).
> When something doesn't have a reference listed, and just says "sourced from a publicly available first-party datasource", what does that mean?
It depends, and the degree to which it depends is why the citation is ambiguous (although it is true, if imprecise). My goal is to individually cite the individual nutrients but it was simply too costly and time-consuming at the stage of the project at which I did this work.
> what is the process like there for interpreting those values?
Because the degree to which something in the database might be related to those values is so varied, it depends. The reasoning agent had access to those database entires, which is helpful because they tend to contain micronutrient data. It also had access to web data, as well as its own world knowledge, and considers sources in that order. Ultimately it was left up to the agent to decide what the most reasonable fit for each food was, thinking through what an average user likely meant by that entry (e.g. a typical user probably assumes a 'Tomato' is raw), and then to choose the best sources from there. For the chicken salad, it used approximate micronutrient values from the listed references to inform its answer, but adapted the end values for how the dish is described in the description.
> if you had the choice between verified data and fuzzy LLM data, you should go for the human verified data (for now)
Human verification isn't free, and that means it is not available to a lot of people who can't or don't want to pay for something. But if that's something that someone values, I would certainly not diss the human effort!
I was under the impression that Tesla has put a lot of effort into building out support infrastructure? My perception is a bit clouded by the fact that I live a few miles from a service center.
Try to get spare parts for teslas. Or fix them at a non-Tesla owned shop. Both are 100x more complex and expensive than any other mass produced car made in the last decade 20 years. And it is like that by design: it was never a priority for them.
I believe you'll meet the "Added or Free sugar" recommendation with 2 medjool dates, but AFAIK, the guideline isn't as strict on naturally occurring sugars, if there is any guideline at all aside from general carb consumption. Dates also have a surprisingly low glycemic index for how sweet they are.
You need to activate the plan in the US and return to the US every 90 days (or lose your mobile data). So it’s not really an option for people outside the US.
I think it's a bit more subtle than that. The code of this tool runs in your browser and makes it download the model from huggingface. So it does not host the model or provide it to you, it just does the download on your behalf directly from where the owner of the model put it. The author of this tool is not providing the model to you, just automating the download for you.
Not saying it's not a copyright violation, and IANAL, but it's not a obvious one.
My company tried using this a few years ago to build a cross platform app (and hopefully share a lot of code with our svelte web interface). We made a lot of progress in 3-4 months of trying, but we ended up running into too many odd bugs and edge cases that were difficult to debug. We eventually switched to Flutter for the mobile codebase.
No doubt this and nativescript have improved since then, but from my early impressions, I wouldn't have recommended nativescript for anything other than the simplest of interface needs.
Also the old Cordova equivalent and Kivy. Take mobile development and try to shoehorn a browser-based solution (they don't want) into it. Now you've got two problems and not enough expertise available to help.
Did they have JS _AND_ React experience beforehand? Or did they just pick RN because it was the cool thing on the block?
I foundered on my first attempt at an React app (side project) coming from a C,C++,Java and non-React JS background, but the second one(professionally) after I've had exposure to some React got released and an fork of it is still in use today. And the company I work with has 2-5 other apps out for customers using RN/Expo (thanks to the positive experience I had).
We're also strict about using TypeScript though so we're not chasing basic type bugs (since debugging is a tad weaker with RN/Expo even if it's usable if shit the fan only occasionally) and we have plenty of React experience outside of it, biggest "quirks" are often RN/Expo upgrades if we haven't maintained things in a while (the oldest app used a relatively early version of Expo though) or platform specific notification crap (that I've heard is painful regardless of Expo/RN).
For me personally, it’s appealing that it just provides a mirror to my existing MacBook. It gives me more flexibility on when to use glasses vs laptop screen, like if I need to take notes in a meeting/lecture hall.
When something doesn't have a reference listed, and just says "sourced from a publicly available first-party datasource", what does that mean? Crawled from other sources and you'd prefer not to say? The wording does feel a little sketchy when contrasted with entries that do list sources.
When something does list references that don't seem super close to the actual food, what is the process like there for interpreting those values? Example, this Chicken Salad inheriting from Chicken Spread: https://www.opennutrition.app/search/chicken-salad-37mAX17YX...
The quality of the data might feel rough now, but I can see this being valuable for our users even if it's just an opt-in "show estimated micronutrients" or something. Would require labeling values as not being directly from a source of truth.
One thing that a lot of people are missing is that there is already a lot of inaccurate nutrition data out there. Even on information directly from the manufacturer, sometimes there are errors, or just old versions of the product that never get scrubbed from the internet (I imagine the latter case would be tricky for an LLM to deal with too). Just logging your dietary intake in any form will get you 80% of the benefit of tracking via some self awareness of your intake. Of course, it's an easy argument to point out that if you had the choice between verified data and fuzzy LLM data, you should go for the human verified data (for now).
reply