5 ms·

Like.. the very fact I typed "javascript" or "css" or "html" means you probably don't need to apply your "did you mean" logic to my searches. If I'm searching

Like.. the very fact I typed "javascript" or "css" or "html" means you probably don't need to apply your "did you mean" logic to my searches. If I'm searching for something highly technical, I absolutely DO NOT NEED your help, Google.

p1esk 2y ago
You’re not the type of users they care about.
- akira2501 2y ago
  Well, I write software for the web, and at this point I consider Google to be an entirely negative company. So now, to the maximum extent possible, I work to undermine their interests and to not participate with their products or addons or browsers. Great strategy on their part.
- godelski 2y ago
  You've over simplified and created a major error. It's a common one though, so allow me to explain. There is no single entity Google cares about. There is no "average user" (nor median). Trying to make the product best for an "average" user makes it bad for everyone. You have made the assumption that this solution space is smooth and relatively uniformly distributed. Consider this: on a uniformly distribution, if you interpolate between two points, those interpolations are representative of the distribution. But if you do so on a gaussian distribution, this is not true. This is because the density of these distributions are not evenly distributed (this is what "uniform" in uniform distribution means). For more complicated distributions you need different interpolations. Many real world distributions are the agglomeration of varying distributions (you probably have clusters and pockets). The distribution of searches (or customers) is neither normal nor uniform. To understand some further complexity, it is important to remember that the burden of success is the requirement for further nuance. Pareto. It may take few resources to get to "80%"/"good enough" but significant to increase that another 10%. Think about a Taylor Series (Fourier Approximation, or pick your preferred example), the first order approximation will only take you so far. It may be good enough to start, but the need for higher order (and this more complex calculations) approximations are increasingly necessary for precision (non-linearly). So simplify where you can, but don't forget what you've done. If you do, you'll be stuck. Or worse, you will compound your error (also very common).
Suppafly 2y ago
>probably don't need to apply your "did you mean" logic to my searches I wish it had a yes/no prompt that you could click and when you click no it would refresh with the results you wanted, and possibly use those answers to update whatever algorithm does the suggestions.
- godelski 2y ago
  With how much they track us you think that returning to the search page and selecting a different result would result in their prior being updated. You'd think that variations of the same search in a small time frame would be recognized as a signal of bad results. I understand the problem is difficult, but it appears to me that the signals they have that can help are being ignored.
  - Suppafly 2y ago
    They may actually do that sort of improvement, and just be somewhat bad at it. At least if they let you check yes/no, you'd feel better about it, even if they ignored the answers.
godelski 2y ago
I wish I could get it to respect advanced filters. I can understand how a problem I type in might be very similar to a simple one, but there needs to be a way to signal differentiation. I know they track me, and you're telling me that clicking on the first 5 links to only revisit the search in under a minute and try a new link is not a signal that I'm not getting what I want? Is visiting page 2 not further signaling? I don't need a different variation of the same result, I need a different result. While the former can be certainly useful, it is absurd to think this is all there is. If you ever search anything with any bit of nuance, you'll have experienced this. I'll give an example: try searching "out of distribution" and try to find a result that is not about ML based OOD detection. They do exist and you'll find some better queries, but that result still dominates (and frankly, they are poor results). For our example query `statistics out of distribution -"machine learning"` has a top result of mathworks on neural networks and all results on the front page work. "statistics" is ignored before variations of "Machine learning." Top few results for `statistics out of distribution -"machine learning" -"deep learning" -"artificial intelligence" -"neural network"` returns to me an ICLR paper who's title that ends in "Deep Neural Networks" and many results are still ML based... How do I tell Google (or any engine) that I want to exclude results!? And before someone says GPT/Claude or some other LLM, the problem persists. Yes, I've paid. Yes, for the most advanced models. Derailing these tools from the wrong track is difficult. Even harder the longer the conversation goes (i.e. if you don't recognize you're headed in the wrong direction). The nuance matters.

p1esk 2y ago

You’re not the type of users they care about.

akira2501 2y ago
Well, I write software for the web, and at this point I consider Google to be an entirely negative company. So now, to the maximum extent possible, I work to undermine their interests and to not participate with their products or addons or browsers. Great strategy on their part.
godelski 2y ago
You've over simplified and created a major error. It's a common one though, so allow me to explain. There is no single entity Google cares about. There is no "average user" (nor median). Trying to make the product best for an "average" user makes it bad for everyone. You have made the assumption that this solution space is smooth and relatively uniformly distributed. Consider this: on a uniformly distribution, if you interpolate between two points, those interpolations are representative of the distribution. But if you do so on a gaussian distribution, this is not true. This is because the density of these distributions are not evenly distributed (this is what "uniform" in uniform distribution means). For more complicated distributions you need different interpolations. Many real world distributions are the agglomeration of varying distributions (you probably have clusters and pockets). The distribution of searches (or customers) is neither normal nor uniform. To understand some further complexity, it is important to remember that the burden of success is the requirement for further nuance. Pareto. It may take few resources to get to "80%"/"good enough" but significant to increase that another 10%. Think about a Taylor Series (Fourier Approximation, or pick your preferred example), the first order approximation will only take you so far. It may be good enough to start, but the need for higher order (and this more complex calculations) approximations are increasingly necessary for precision (non-linearly). So simplify where you can, but don't forget what you've done. If you do, you'll be stuck. Or worse, you will compound your error (also very common).

Suppafly 2y ago

>probably don't need to apply your "did you mean" logic to my searches I wish it had a yes/no prompt that you could click and when you click no it would refresh with the results you wanted, and possibly use those answers to update whatever algorithm does the suggestions.

godelski 2y ago
With how much they track us you think that returning to the search page and selecting a different result would result in their prior being updated. You'd think that variations of the same search in a small time frame would be recognized as a signal of bad results. I understand the problem is difficult, but it appears to me that the signals they have that can help are being ignored.
- Suppafly 2y ago
  They may actually do that sort of improvement, and just be somewhat bad at it. At least if they let you check yes/no, you'd feel better about it, even if they ignored the answers.

godelski 2y ago

I wish I could get it to respect advanced filters. I can understand how a problem I type in might be very similar to a simple one, but there needs to be a way to signal differentiation. I know they track me, and you're telling me that clicking on the first 5 links to only revisit the search in under a minute and try a new link is not a signal that I'm not getting what I want? Is visiting page 2 not further signaling? I don't need a different variation of the same result, I need a different result. While the former can be certainly useful, it is absurd to think this is all there is. If you ever search anything with any bit of nuance, you'll have experienced this. I'll give an example: try searching "out of distribution" and try to find a result that is not about ML based OOD detection. They do exist and you'll find some better queries, but that result still dominates (and frankly, they are poor results). For our example query `statistics out of distribution -"machine learning"` has a top result of mathworks on neural networks and all results on the front page work. "statistics" is ignored before variations of "Machine learning." Top few results for `statistics out of distribution -"machine learning" -"deep learning" -"artificial intelligence" -"neural network"` returns to me an ICLR paper who's title that ends in "Deep Neural Networks" and many results are still ML based... How do I tell Google (or any engine) that I want to exclude results!? And before someone says GPT/Claude or some other LLM, the problem persists. Yes, I've paid. Yes, for the most advanced models. Derailing these tools from the wrong track is difficult. Even harder the longer the conversation goes (i.e. if you don't recognize you're headed in the wrong direction). The nuance matters.