To me these show what is lacking in ChatGPT. In college (like 12 years ago) we all had to make a text markov chain generator. The result were sentences which were grammatically correct but meaningless and random (He went to the library and saw a golden goat). I sorta feel like like Chat GPT is doing that but with more ability to weight on higher level structure and weight on patterns in outside text. But these obviously are not real recipes and wouldnt work in real life. There's still not logic. It just seems like like better markov chain text generation.
A few years before that when I was in school we had the MIT fake paper generator (unfortunately no longer maintained). It used context free grammar to generate something similar, linking academic CS terms together to write nonsense conference papers.
You're right, modern language models are the same. They are more polished but still just as stupid. They don't understand anything, they just put a pattern together mechanically.
Personally I like systema like in the article posted here better, because they have a funny "mad-lib" quality instead of the low-quality blog content style of language models. It seems like where the language models can be more funny is in imitating a person's writing style
(I also wish we had an HN norm against copying chatgpt output into posts unless it's specifically an article about chatgpt.)
Maybe I am not a great cook (I thought I was decent though) but these recipes look totally plausible and normal, even if not the best recipe you’ve ever seen. They don’t look like gibberish markov chains to me. The ingredients are prepared appropriately
(I know a farfalle bun is a non sequitur but that was given as input- not provided by ChatGPT)
If you gave the name to the chef they’d wonder wtf somebody was doing mixing pasta and bone marrow and putting it in a bun.
Putting pasta into a sandwich makes slightly more sense than making a bun out of a specific pasta shape and I guess that’s missing from the model, but really it did a decent enough job here.
I don’t know when we’ll be at a place where if you ask a model for something and it comes back and tells you no and why that’s a dumb suggestion.
There’s no pasta in the ChatGPT recipe, it’s a “farfalle bun” (which we can assume is some type of pasta-inspired bun) with marrow in it. There is no pasta separate from the bun which gets put in the bun
They may be individually prepared appropriately but that is just because it copies other recipe texts. Perhaps it could be used to generate new ideas but so could a random recipe generator that picks three different cocktail ingredients and tells you to mix them. I'm not sure how this is any different than that, other than presentation.
Presumably actual recipes you see in a cookbook written by a person have been made in real life and the person could vouch for their quality or ease of making.
Those seem reasonable. The recipe name is absurd so by necessity the recipes are. If you asked a chef to make a recipe for those dish names and they didn’t just ignore the name, it’d probably be similar.
