FYI, there is a newer version at https://sites.google.com/a/brown.edu/sergei-treil-homepage/l... . Google seems to never have learnt about this site, which is weird as it is on a Google server.
I know there are Google engineers here, possibly search. Do you guys not also feel it? I mean you dogfood, right? I know it is a hard problem but boy has it become harder to use, especially in the last 6 months.
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.
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).
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.
I understand the problem is difficult, but it appears to me that the signals they have that can help are being ignored.
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.
https://www.math.brown.edu/streil/index.html redirects to https://sites.google.com/a/brown.edu/sergei-treil-homepage/h...
But OP's link, https://www.math.brown.edu/streil/papers/LADW/LADW.html does not redirect. So it seems like a problem that the site owner can fix.
> "Linear Algebra Done Wrong site:sites.google.com" finds the page you linked.
As in, if you didn't know a second newer version existed, would you find it? Certainly the new version should prioritize the older version.
FWIW, my query "Linear Algebra Done Wrong" yields the newer version (2017), but I've just interpreted OP's point as being illustrative and helpful to readers (and it looks like the link now does point towards the new version. Possibly this made my query work)