"I found the best burger place in America. Then I killed it."
Small Brain
Ok. Let's imagine burger places could be reduced to one dimension: Goodness. Local McDonalds: 12/100. Stanich's: 98/100. Then this article would make sense, and the aftermath would be just tragic unintended consequences. Maybe there's a Secret Burger Guide out there that actually knows the next Best place and is hiding it and you can only know it if you're in the Secret Burger Cabal and when you go to this place that is 97/100 you'll say "yes, this is definitely better than the 90/100 place I went to last week."
Medium Brain
Obviously, burger places are not one-dimensional and this is nonsense. Let's imagine burger places are reducible to a small number of dimensions, like 3: taste, atmosphere, and value. Then there is no one "best"; you could average them, but that's not really what people want. You as a reader could then decide "I think atmosphere is the most important" and make your decision based on that. There might still be articles like this: "I found the best tasting burger in America." We'll still have the same problem - it'll just be diluted a bit, as the Best Taste, Best Atmosphere, and Best Value place all get deluged with 1/3 the crowds. Also, each of those people will probably be a little disappointed, as they didn't really want only the taste of the burger.
Large Brain
Obviously, burger places are not 3-dimensional and this is nonsense. Let's imagine burger places have 100 dimensions. Then you could maximize each dimension and spread the crowds out 100 ways - everyone picks their favorite dimension and maximizes that. Burger places start to specialize in only 1 dimension: e.g. Fries Crispiness. Crispy Fries enthusiasts make a pilgrimage to the #1 Fries Crispiness place. The Fries Crispiness place stops making burgers, because everyone's just coming for their fries. Specialization of labor! Everyone's a little disappointed. Also, the fries taste bad because the #1 Fries Crispiness place isn't also the #1 Fries Savoriness place.
Obviously, reducing things to 100 dimensions then maximizing one is dumb and this is nonsense. We can turn every burger place into a 100-dimensional feature vector; your preferences are also a 100-dimensional feature vector; we maximize the dot product of these for you. Now everyone gets super-personalized burger recommendations. Turns out the best one for me is 2000 miles away - eh, still cool, I can make a pilgrimage there someday! There's still some clustering near the top, as those that are pretty good on a lot of dimensions get recommended to more people. But, it's not bad. Still, on Tuesday I might be looking for a quick tasty burger, and on Friday for the best burger in the city. Preferences change over time. Plus, each burger place changes over time: on Friday they're packed, so I'm not gonna go there then.
Exploding Brain
Ok, so burger places aren't 100 dimensions; they're 1000 dimensions plus 1000 temporal dimensions, and so am I. Now we can no longer get enough data to fill this absurdly high dimensional space. We can do pretty well, and we do. Still, I have to take a long survey every time I want to go out, just to see what my temporal preferences are like now, and that's annoying. Plus, sometimes it's just wrong. To test it out, I spent a month picking burger places by throwing darts at a map, and I did just about as well, and it was a lot easier.
Galaxy Brain
Burger places aren't 2000 dimensions and this is nonsense. Burger places are in roughly 4 categories: bad, bad and I have a story about it, good, good and I have a story about it. Here, "story" can mean:
- this place is better than the average "good" place; I would go out of my way to eat here
- this place is fine in most dimensions and great on one
- this place is not good usually, but if it's 2am and everything's closed and you were out partying it's great late-night food
- it's so bad you never want to go there; I got sick after eating there
- I have fond memories of this place from college
These stories are in too many dimensions to even model. You can never get enough data to accurately predict if someone will have an "interesting story" about a place. And importantly, you should stop trying. Every second you spend on it is spinning your gears; every story about "The Best burger place" might accidentally ruin it.
Conclusions That Are Not As Profound As I Thought When Starting This Post
Burger places have the following properties: temporally varying, high supply, low stakes, result measured in feelings and stories. As a result, if you wanted to pick a place to live and you loved burgers, you should optimize for "the greatest number of diverse burger places" rather than "the best." See also: coffee, wine, music, movies.
Edit: Even Bigger Galaxy Brain
One review doesn't ruin a place - spousal abuse and ignoring court orders do. Welp :P
Small Brain
Ok. Let's imagine burger places could be reduced to one dimension: Goodness. Local McDonalds: 12/100. Stanich's: 98/100. Then this article would make sense, and the aftermath would be just tragic unintended consequences. Maybe there's a Secret Burger Guide out there that actually knows the next Best place and is hiding it and you can only know it if you're in the Secret Burger Cabal and when you go to this place that is 97/100 you'll say "yes, this is definitely better than the 90/100 place I went to last week."
Medium Brain
Obviously, burger places are not one-dimensional and this is nonsense. Let's imagine burger places are reducible to a small number of dimensions, like 3: taste, atmosphere, and value. Then there is no one "best"; you could average them, but that's not really what people want. You as a reader could then decide "I think atmosphere is the most important" and make your decision based on that. There might still be articles like this: "I found the best tasting burger in America." We'll still have the same problem - it'll just be diluted a bit, as the Best Taste, Best Atmosphere, and Best Value place all get deluged with 1/3 the crowds. Also, each of those people will probably be a little disappointed, as they didn't really want only the taste of the burger.
Large Brain
Obviously, burger places are not 3-dimensional and this is nonsense. Let's imagine burger places have 100 dimensions. Then you could maximize each dimension and spread the crowds out 100 ways - everyone picks their favorite dimension and maximizes that. Burger places start to specialize in only 1 dimension: e.g. Fries Crispiness. Crispy Fries enthusiasts make a pilgrimage to the #1 Fries Crispiness place. The Fries Crispiness place stops making burgers, because everyone's just coming for their fries. Specialization of labor! Everyone's a little disappointed. Also, the fries taste bad because the #1 Fries Crispiness place isn't also the #1 Fries Savoriness place.
Obviously, reducing things to 100 dimensions then maximizing one is dumb and this is nonsense. We can turn every burger place into a 100-dimensional feature vector; your preferences are also a 100-dimensional feature vector; we maximize the dot product of these for you. Now everyone gets super-personalized burger recommendations. Turns out the best one for me is 2000 miles away - eh, still cool, I can make a pilgrimage there someday! There's still some clustering near the top, as those that are pretty good on a lot of dimensions get recommended to more people. But, it's not bad. Still, on Tuesday I might be looking for a quick tasty burger, and on Friday for the best burger in the city. Preferences change over time. Plus, each burger place changes over time: on Friday they're packed, so I'm not gonna go there then.
Exploding Brain
Ok, so burger places aren't 100 dimensions; they're 1000 dimensions plus 1000 temporal dimensions, and so am I. Now we can no longer get enough data to fill this absurdly high dimensional space. We can do pretty well, and we do. Still, I have to take a long survey every time I want to go out, just to see what my temporal preferences are like now, and that's annoying. Plus, sometimes it's just wrong. To test it out, I spent a month picking burger places by throwing darts at a map, and I did just about as well, and it was a lot easier.
Galaxy Brain
Burger places aren't 2000 dimensions and this is nonsense. Burger places are in roughly 4 categories: bad, bad and I have a story about it, good, good and I have a story about it. Here, "story" can mean:
- this place is better than the average "good" place; I would go out of my way to eat here
- this place is fine in most dimensions and great on one
- this place is not good usually, but if it's 2am and everything's closed and you were out partying it's great late-night food
- it's so bad you never want to go there; I got sick after eating there
- I have fond memories of this place from college
These stories are in too many dimensions to even model. You can never get enough data to accurately predict if someone will have an "interesting story" about a place. And importantly, you should stop trying. Every second you spend on it is spinning your gears; every story about "The Best burger place" might accidentally ruin it.
Conclusions That Are Not As Profound As I Thought When Starting This Post
Burger places have the following properties: temporally varying, high supply, low stakes, result measured in feelings and stories. As a result, if you wanted to pick a place to live and you loved burgers, you should optimize for "the greatest number of diverse burger places" rather than "the best." See also: coffee, wine, music, movies.
Edit: Even Bigger Galaxy Brain
One review doesn't ruin a place - spousal abuse and ignoring court orders do. Welp :P
No comments:
Post a Comment