this post was submitted on 25 Sep 2023
316 points (95.1% liked)
Technology
59207 readers
3264 users here now
This is a most excellent place for technology news and articles.
Our Rules
- Follow the lemmy.world rules.
- Only tech related content.
- Be excellent to each another!
- Mod approved content bots can post up to 10 articles per day.
- Threads asking for personal tech support may be deleted.
- Politics threads may be removed.
- No memes allowed as posts, OK to post as comments.
- Only approved bots from the list below, to ask if your bot can be added please contact us.
- Check for duplicates before posting, duplicates may be removed
Approved Bots
founded 1 year ago
MODERATORS
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
I don't know what their ai process looks like, what kind of data they trained it on, etc.
But annecdotally, I've played around a bit with chatgpt making cocktail recipes, and it's been surprisingly good at it. They sometimes need a little fine-tuning but they tend to get you in a pretty close ballpark, it's made some interesting suggestions I probably wouldn't have thought of, but nothing that turned out to be bad.
A lot of recipes tend to follow some pretty well-established ratios which means they can be broken down into some sort of mathematical formula which is something computers can actually do pretty well, and it's often just a matter for swapping out one ingredient or combination of them for another that is similarly salty/sweet/bitter/sour/umami.
For example a standard recipe for punch is 1 part sour, 2 of sweet, 3 of strong (liquor of your choice), 4 of weak (tea, juice, soda, water, etc.) and you can mix and match just about any ingredients that fit those profiles and get a drinkable punch.
I'm sure a company like coke probably has a long list of flavorings with known and well-documented flavor profiles that an ai trained on a list of proven recipes could mix and match with all day long.
That's not how a LLM like Chatgpt works. It's not referencing cocktail recipes, compiling their ingredients, seeing commonalities, figuring out mathematical formulas, and then experimenting with the variables. That kind of thing is still probably a decade or more away, if not decades.
If you're curious, see what people have tried to do with AI generated recipes that fit nutrition guidelines and how they never add up correctly.
From what I've seen of cocktails and recipe books, there's probably a lot more of them than you realize. I guarantee you there are thousands of cocktail recipes you've never heard of that have been written into published recipe books.
All of that to say that Chatgpt is basically just making up arrangements of words that based on its training should go together.
Yeah I've seen people recommend ChatGPT for meal planning and recipes, and it's mostly fine for common, simple recipes but it does super weird things when you ask it for something nonstandard, that has a lot of variations, or with dietary restrictions. Like it repeatedly gave me recipes with my allergens with a note to check package for said allergen and other weird things like claiming frozen vegetables take 10 minutes to roast in the oven. It's useful for certain things but it's not really intelligent.
So, basically what people who are decent at cooking do all the time. Groundbreaking.
I'm a little bit of a cooking nerd, and a pretty adventurous eater. There are some flavor combinations that, when they're explained to me, make a lot of sense, and I can see how they would work well together, but I would never think of putting together myself in a million years unless I saw some high-end chef on a cooking show or fancy restaurant do it first.
Off the top of my head, I remember someone on iron chef making some sort of fish ice cream, and someone on some other cooking show making some sort of liver pate and jelly donut, and both were very well received by the judges. I'd never think of putting those ingredients together in those ways, my first gut instinct if you just told me that those foods existed without further explanation was that they sound gross, but after thinking about it or having the chefs or judges explain them, I can totally see how they can work.
There's only so many chefs cooking at a high level like that though, whose brains are wired in such a way that they really understand how the flavors can work together and can work around the biases that most of us have and put together ingredients in new and unexpected ways.
AI often won't have the same biases we do (though it may be biased in other ways) so it could lower the barrier to entry for those of us who have the hands-on skills to put those sorts of dishes together, but maybe aren't quite creative enough to come up with them by ourselves, and for the more creative types it could potentially become a useful sounding board for them to bounce ideas off of.
And yet the example soda flavor wasn't well received because randomly putting shit together isn't something that is inherently better when a computer does it.
Even though it's apparently a pretty lackluster soda, I think it's pretty notable that I haven't seen any reviews saying that it's outrightbad, it's just not great. That's better than I would expect from just randomly mixing ingredients.
Now we don't know how many iterations it took to get them to that point, what kind of prompts or human handholding it took to get it to that point. It very well might be that the computer gave them a thousand bad formulas and this was the only one that was remotely palatable, but we don't know and probably never will know if that was the case.
Not that I think coke will do it, but personally I think it would be cool for them to take the feedback they get from this soda, feed it back into the ai and have the computer design a version 2.0 based on that feedback and see how well it goes, and keep iterating it that way and see where they end up.
I mean for a machine to do it? Yeah it kind of is.
Like the whole purpose of developing AI is to replace us, it being able to do what we do is literally the metric we are shooting for.