Then you have me pegged wrong. I’ve been running local LLMs since 2022.

But this is what ChatGPT likely does, under the hood:

• When you ask it for recipes, it goes to Google and searches for them.

• It dumps the top results into its context… the same SEO spam you are using your chat to avoid, I’d presume.

• It then averages that with its own amalgamated knowledge to “improvise” a recipe. It’s not thinking through the logic and physics of cooking or referencing a cookbook, it’s just taking a guess at what sounds like a plausible recipe written out.

• On top of that, line by line, every measurement, every ingredient, every step is also pseudo randomized; there’s a chance it may pick 8 oz for some ingredient, or 16 oz of another. It’s a roll of the dice. That’s how LLM sampling works.

• Outside the thinking block, it also has no way to go back and correct itself if one of those randomized measurements it’s obviously wrong; it just goes with it, like an improv actor.

So…That’s just a computer randomized recipe.

It may work, it’s not bad necessarily, but you have no idea if it’ll work or not until you assess it and try it yourself.

So what’s the point?

What’s the reason for doing it?

There’s no advantage over a real cookbook, where at least the actual recipe has been assembled, eaten, and enjoyed enough to be deemed worthy of inclusion by the author. ChatGPT does no such thing unless it finds real cookbooks to reference.

Now, if you feed it real cookbooks into the context (or use some built in feature that does so), and ask it “pick me a recipe I would like. Modify it.”

That’s fine!

That’s right up an LLM’s alley.

But you can’t feed it SEO spam and expect something good out. Garbage in, garbage out is the term data scientists use.

This is what I try to emphasize every time I talk about LLMs. They have to be machines you understand the inputs and outputs of, not a black box you can trust to give you reliable info when you don’t know how it gets it.

On a separate note, I really, really hate OpenAI. I think they are a pyramid scheme.

I always suggest using literally any LLM but ChatGPT. Literally any one. Ideally an open weights one, but please, just not ChatGPT for so many reasons…