The Doordash discourse because of Citrini's post is a bit out of control. Especially because the thing is, we don't need to speculate, we can literally test this. I spun up an environment with Doordash and 100 other AI generated competitors, with varying prices, ratings, delivery times etc. Then asked an LLM to pick one based on my order.

And it shows LLMs behave like a blanaced scoring equation - i.e., they don't care about brand and won't pick Doordash if its metrics aren't better than others. In fact, the LLM is even stricter than pure math, i.e., it picked it less than random chance.

When you advantage DASH, the LLM picks it slightly more often, but still only based on the actual merits of cost, delivery time and other such KPIs. Whatever the objective function is (price only, price + reliability, quality) that's what gets maximised. Not brand, not unless explicitly asked to.

So the question becomes can anyone actually compete with Doordash enough to get better metrics. And the answer is obviously yes, even if only for specific foods in specific locales. No reason the 100 competitors in Palo Alto should be the same as SF let alone Williamsburg or North Austin.

People forget that we can literally do empirical analyses at the speed of thought these days and you don't need to speculate endlessly about what a future hypothetical AI might do. The future's already here.

PS: FWIW I started this thinking obviously DASH would win, but the experiments proved me wrong, so here we are. If any of you want to play with it, repo is here: github.com/strangeloopc…