That’s not how it works.

LLMs are text models. They’re like weather models; they live in the context of the input, it’s literally their entire “world view.” The inputs are the most important part of the quality of the answer.

You need to be able to audit the inputs. If ChatGPT doesn’t even show them, that is a tremendous issue.

And I don’t mean to criticize recipes it’s given that have turned out well, but that’s, literally, mathematically, objectively, a roll of the dice if it’s not grounded in real recipes.

As an alternative, I would highly recommend:

• Asking the LLM service to find the most accredited cookbooks in the niche you want.

• Get them locally, copy and paste the categories you want into chat.

• Ask it to pick or (or modify/synthesize) a good recipe from that.

This is what text models were designed to do before OpenAI commercialized them.

thinking

See, this is where I think you're wrong. There's no thinking going on. It's a probability analysis based on word adjacency in a text dataset, with a little bit of $RND thrown in. If you ask the same question multiple times, you'll get different answers because of the presence of the random seed. If you ask it why chocolate coated shrimp is a bad idea, it's not going to answer based on culinary training saying that those two flavour profiles don't belong together, it's going to say no because some other crazy person actually tried it and that data made it into the dataset. (And yes, someone did, I checked...)