[vector_spaces]

This is a deceptively difficult problem. Food is incredibly messy -- grocery delivery is very far from being a solved problem even with a human in the loop. You have to deal with stockouts and sensible replacements, and driver/picker error, and quality variance. Quality variance is a huge issue in perishable categories and a major reason why foodtech is tremendously difficult (and fun, IMO).

Simple quality variance examples: banana ripeness. Or size of items that can only be ordered by each instead of by lb. Or one of the two onions you needed looking mostly fine on the outside but rotten on the inside.

As an experiment, try ordering all the ingredients to make a specific recipe several days in a row. You'll tend to hit an failure rate between 15-30%. That failure rate is usually fine if you're just restocking for home -- you can always pick up milk/sugar/whatever tomorrow -- but it's pretty awful if it means that something like 1 in 5 of your dinner plans are ruined or you have to leave your guests to rush to the store to pick up some missing ingredient

Also: the LLM will need to be aware of your home inventory, unless you're fine with it ordering lots ingredients you already have

So there's lots of hidden complexity here. If they turn this on, it will be a fun party trick that will work once in a while, but getting burned with ruined plans causes people to churn out fast.

[anthonypasq]

all you've done is outline a series of mildly tricky but completely solvable problems to a use case that most humans would find incredibly useful. its very strange that you cant extrapolate 5 years down the line and see that this is completely reasonable.

[vector_spaces]

Underestimation of the problem space is why foodtech is a tarpit for tech companies. Many have tried and failed to solve these very problems over decades. I don't blame you, to a green outsider the food industry seems like it would be simple, but the devil is in the details. I'd love for you to prove me wrong though

None of this is to say that LLMs have nothing to offer here. There would still be value in being able to tell an agent "Here's my list, get this ordered for me". But being able to say "find me a recipe for dinner that my guests will like and have the ingredients delivered in time for me to make it" without getting burned every other time is actually a much harder problem.

[kyleee]

I tend to agree with you overall. Think about a small case study with avocados; what’s the confidence level you could order 3 avocados that you need to use for a recipe upon delivery and that you’d be satisfied with the quality and ripeness level? I’d put it at probably 20%, which means it’s a non-starter.

[vector_spaces]

Literally tonight I ordered heavy whipping cream, strawberries, and biscuits from Doordash's dashmart store to make strawberry shortcakes, having forgotten about this thread. Until I saw the whipping cream was puffy and expired nearly a month ago. Haha