I don’t think an LLM is required for the UX you want. Online grocery delivery already exists. They could easily offer a list-based UI where you upload a list and it populates a cart for you.
I order groceries online a lot. My personal opinion about why this UX wouldn’t work well is that people want to feel like they made an optimal choice. Like when you go to the store to buy an apple, do you just pick the first one? Or do you pick the organic variety that looks freshest? Or maybe you go in for an apple buy decide to get grapes because they look better. There are a few staples I always buy, but generally my grocery purchases are more deliberate. From what I see at the store, many people exhibit such behavior.
My father-in-law visited last week to help with the baby and my wife's broken leg. He asked if I wanted anything from the grocery store. I said, "an instant ramen cup, but not the cheapest brand. Also, watch out because they sometimes have beef broth and I'm allergic." He came back with a choice that I would never have picked myself, but it fit all my criteria and perfectly satisfied my sodium craving.
When I'm outsourcing a chore, I'm doing it to exchange money for time and effort. I would prefer to be able to abstract away the details, and trust that my request will be satisfied. In the current UX, I have to specifically request a tub of "Ben and Jerry's Extra Choco Choo Choo Churn", then be disappointed when the desperate minimum-wage worker somehow buys a version with almond milk because it was the closest one in stock. All I wanted was "a $5 pint of chocolate ice cream" to cheer up my depressed wife, and she's allergic to almonds...
Perhaps an LLM could provide that level of abstraction to customers, while reliably compiling it down to a sufficiently concrete list of options for a worker to retrieve. The abstraction would likely improve the UX, as the workers would have more freedom to choose compatible products rather than arbitrarily specific ones. It would be closer to a personal shopper experience virtually built on top of commodity labor, which seems like the whole goal of these grocery services.
> I said, "an instant ramen cup, but not the cheapest brand. Also, watch out because they sometimes have beef broth and I'm allergic."
This strikes at the heart of the issue. The criteria you’ve placed on the ramen is actually pretty complicated, and yet it is simpler than the criteria many use to select produce, dairy, and meat. Needing to meticulously articulate criteria for the 100+ items purchased every month would be a pain. Items also have interdependencies, e.g. I can only afford the fancy yogurt if I got a deal on something else, and context dependent, e.g. I want to spend more on chicken for the fancy dinner, and less for a weeknight meal.
Of course, these criteria are theoretically learnable by an LLM, but extracting the prerequisite information is hard. So much context exists in your mind. For instance, just looking at your ramen purchase history, do buy chicken ramen over beef ramen because you like chicken better, or is it because you actively dislike beef? It’s hard to say without your explicit input. If every purchasing decision requires explicit input, it seems better to cut out the LLM.
The best UX would be to have a personal shopper. That's the point of using human language. I can be explicit about the properties I care about and vague about the properties I don't care about. I'm not going to pay someone $100 to do my grocery shopping or have my father-in-law do it every week, though. An algorithm that can robustly interpret human language can also optimize a shopping cart significantly faster than a human can, based on massively indexed databases.
Not to be rude but to highlight my point, why are you trying to infer my ramen flavor preferences from my purchase history? I already explicitly said that I am allergic to beef. That's the whole point. It takes practically no effort for a literate human to speak or write down the criteria that are important to them. It is significantly harder for a computer to correctly interpret it, but that's why it seems like an LLM could help.
I order groceries online a lot. My personal opinion about why this UX wouldn’t work well is that people want to feel like they made an optimal choice. Like when you go to the store to buy an apple, do you just pick the first one? Or do you pick the organic variety that looks freshest? Or maybe you go in for an apple buy decide to get grapes because they look better. There are a few staples I always buy, but generally my grocery purchases are more deliberate. From what I see at the store, many people exhibit such behavior.