Welcome to Adventures with AI, a column exploring what happens when artificial intelligence takes control of everyday tasks.
Eating at a restaurant is one of my great pleasures; the kitchen is not.
Unfortunately, since the start of the COVID-19 pandemic, I have done a lot of these and hardly any of the first.
Preparing meals has become particularly tedious during the last London lockout. Then as an unhappy couple in a sexless marriage, I tried to spice things up in my domestic life. Only instead of strapping on a gimp mask and a ball gag I have experimented with with AI.
[Read: How do you build a pet-friendly gadget? We asked experts and animal owners]
I first looked for the culinary inspiration of GPT-3, a text generator intended either for Conquer the world or burn in a blaze of bigotry and pseudophilosophy.
The model was trained on a mind-boggling amount of data, including the entire English Wikipedia, two extensive books of books, and a filtered version of Common Crawl. With so many recipes now online, GPT-3 had to learn its way into the kitchen. Law?
I put my stomach on the line to find out.
Access to GPT-3 remains prohibitive for most, but nutritional benefits Refluxgate Kindly provided me with a selection of his recipes.
For my starter, I whipped up a plate of honey and soy glazed vegetables.
The model suggested all the ingredients you would expect from such a dish – except the vegetables.
Thanks to the cooking gods, he recommended serving this monstrosity with rice:
The dish was gloopy and overpowering but somehow palatable. Definitely not something I would ever cook, but what if I was the last man on Earth and it was the last food on Earth, and the future of the human race depended on us? I would probably eat my own foot.
Maybe I am unfair. My regular diet of fast food and cigarettes may have rendered my taste buds unable to appreciate the subtle flavors of GPT-3. Did I judge the model too harshly?
I asked someone with a more sophisticated palate: Ellen Parr, the chef at London restaurant Lucky & Joy.
She was not impressed.
The recipe does not include any vegetables or instructions on how to cook the vegetables used. Each vegetable has a different cooking time, so these are wrong instructions. The AI also recommends storing the vegetables for five days and the sauce for three days. I would say the lifespan would be the other way around.
Still, it was edible enough to give me hope for my main course: a tomato sauce generated by GPT-3, served on a bed of tagliatelle generated by Tom.
The main course
I love my meals as I love my stories: with a beginning, a middle and an end. But this main course made me wish I could move on to the epilogue.
There is no polite way of saying it: it sounds like crap. One of that really disgusting shit that makes you question your lifestyle choices.
In all fairness, it tasted better than it looked – but not by much. GPT-3 recommended so much raw garlic it felt like I was eating a vampire’s nightmare. On the plus side, garlic suppressed the flavor of undercooked tomatoes.
Parr spotted other questionable tips in the recipe:
It has a lot of raw garlic added at the end of the dish, which wouldn’t be so nice and potentially overwhelming. It seems that they were inspired by tomatoes cooked with a raw onion, but it is not so comparable that the onion is cooked in the tomatoes for 45 minutes. And blitzing the tomato with oil will make it emulsify, which wouldn’t be an ideal effect.
There were other shortcomings in the other GPT-3 recipes. They were often difficult to follow, lacked attention to detail, and sometimes suggested unsafe practices, such as putting a pot on a hot stove.
Refluxgate lead researcher Mel Kasulis told TNW that the problems stemmed from the model’s training data.
He is not able to assess the level of effectiveness of his steps and because he trains through many recipes, it may be a matter of mistakenly synthesizing information that is not meant to go together. This did not happen for all recipes, however, this could simply be the limitation of not being able to train enough for an individual Recipe.
GPT-3 also showed a curious admiration for Gordon Ramsay. The model attributed each of her recipes to the sworn Scottish chef.
Ramsay’s vast empire of restaurants, cookbooks, and TV shows has clearly spread to large swathes of the internet.
I lived alone during the pandemic (* sniff *), which means there was no one to compliment my amazing concoctions.
The loneliness also drove me a little crazy. So last week I decided to go back to live with my mom.
To show my appreciation for his hospitality, I generously offered to prepare an AI-generated recipe for him.
Now I’m obviously a disappointing son, but not so disappointing that I give him veg without veg. Instead, I whipped him pancakes.
I used the ingredients recommended by software company Monolith AI, which generated them by training a machine learning model on 31 American-style fluffy pancake recipes.
It was hands down the best AI recipe I have tried.
Even my Belgian mother was impressed with American pancakes – to an extent.
The dough has more flavor than the one I am making. But I prefer the thinner pancakes.
After a good meal, there is nothing I enjoy more than a stiff drink.
I’m not a mixologist, but a man called Tim who writes code has a YouTube channel called Tim Writes Code where he recently shared a few GPT-3 generated cocktail recipes including one called Tim Writes Code.
However, I didn’t do that one. Not because I don’t like Tim and his code, but my declining liquor cabinet didn’t have the alcohol needed for this elaborate brew.
Instead, I tried the equally imaginative GPT-3, a blend of vodka, lime juice, and blueberry syrup, topped with a lime wheel.
The cocktail was extraordinarily sweet, which suited me perfectly. But would that impress the experts?
I asked Peter Kelly, bar manager at the aforementioned Lucky & Joy, if he was considering serving him:
There is no measurement for lime juice on the GPT-3, and there is no mention of ice cream in any of the recipes. Methodology also much too simple. A lot of things are seriously wrong. On the positive side, the flavor profiles work. The the proportions also seem correct.
He didn’t envision GPT-3 earning Michelin stars, but saw some potential for AI in the kitchen.
“It might be beneficial to grab some ingredients and generate something for you. It’s economical and potentially eliminates waste. “
Email Advantage from Refluxgate had higher hopes for the future of GPT-3 in recipe design:
It has already evolved by leaps and bounds from its second model, so I think the technology is only getting better. Some things I would like to see are better organization of thought, and maybe intersect its training with some FDA guidelines on food safety to make it more convenient and safe for general use.
As for me, I have mixed opinions on the culinary skills of AI. It certainly created some interesting recipes that no good chef would have designed. Unfortunately the starter and main course were absolutely terrible.
Yet they couldn’t have been so bad; I ate every piece and left the table full. But my stomach will never be forgive I.
Humanoid greetings! Did you know that we have an AI newsletter? You can subscribe to it here.
Published March 12, 2021 – 18:17 UTC