I do love the appeal to bread making. It's a wonderful example. If people haven't made french bread by hand, it's a humbling exercise.

Recipes of course have evolved too. Old roman recipes were merely a list of ingredients. Water, flour, salt, yeast.

Written steps came after, then photos, videos, gradually replacing hands on training / kneading.

There are now recipes as code running sour dough assembly lines. Certainly capturing much more detail in technique than even a well made video. But I bet there is still human QA at the end judging "is this bread what folks expect?"

I suspect that in order of complexity you'll get "can I attempt to follow each step", "can I follow the intention of each step and understand if I've failed to meet it" (mitigated by using more specific and detailed steps) "can I follow the intention of the recipe itself - can I add or modify steps that are missing to give the ideal form of sour dough" (maybe you show a machine what good bread looks like, moisture content, crunch?) Those mostly overlap with the 3 you've called out. But I'd add "why would anyone make bread?" Why the heck are we still mixing flour and water together. Why does this recipe exist? Great crusty sourdough requires them all.

In my opinion, the bread example doesn't really work that well because it bridges into the physical domain which most cognitive systems don't have access to. That said, for grounding context and therefore creating truth having a version of a world model is very important (See Yan LeCun's work). My experience is that given the right world to operate in, an agent can indeed find flaws in recipes and fix it even though the agent has not been prompted explicitly to do it. This world, as far as I understand it now, is a combination of sequential (at which step am I in a process), conversational (what was talked about alread/ what had I done already), and context memory (what is the frame or reference/plane of existence).

Self-correcting agents are already here: https://jdsemrau.substack.com/p/hyperagents-and-self-correct...

Reminds me of Moravec's paradox, that it is easy to get computers to ace complex math tests but difficult to teach them to walk. We are very excited that computers have mastered the "know the recipe" step and are underplaying the complexity of actual intelligence required to really replace people.

My fear in your above example would be that we offload more and more of the "know the recipe" intelligence to computers and humans are slotted in as replaceable manual labor and are left arguing with a computer about whether the starter needs to be fed or not (or whatever equivalent scenario).