It is exceedingly unlikely that the results don't overlap after the first few, but if you can produce a copy of the two sets of results, I will forward them to someone on the Google Search team for debugging.
People hugely overestimate the effect of personalization -- it is a ranking tweak not a complete change to the search engine. It does not make economic sense to have personalized whole-web indexes.
This seems to be the most accurate post on the topic, yet carries the lowest score. This is about making suggestions, not auto-sending messages.
Typing on a phone is annoying, so if I can say "on my way home" with fewer clicks, instead of having to retype the message all the time, I would be happy to do so. Of course, I could set up some kind of macro, but an automatic system is far easier for normal folks to use.
Of course, whether this should be patentable is an entirely different matter, but the feature is an entirely reasonable thing to try.
If you assume a linear traction-limited model[1], information only needs to be sent backwards. Specifically: (a) current velocity (b) current acceleration/deceleration. (c) maximum deceleration.
The immediately following car just needs to measure the distance[2], and know its own state and parameters. Then it can calculate how much space it needs to stop if the car in front immediately starts breaking at the maximum rate. You can incorporate communication & reaction delays easily too, as well as any bounded noise on the state variables. This would work for an arbitrarily long chain of cars, with each car just needing to monitor the one in front of it. It even works for autonomous cars following human-driven cars[3].
I used to be a robotics researcher, with a focus on high performance navigation. The lane-free full 2D generalization of the problem above was a chapter in my thesis (from 2007).
[1] or just make sure your actual model fits within a conservative linear envelope. [2] obviously you can estimate a&b, but it potentially introduces additional delay to get the noise down, in particular for acceleration since that is second order. [3] this is an ideal case *IF* that driver is paying attention, since the human driver has a better perception system.
I'm afraid the idea, often expressed in this discussion, of "that's what most people want" sells us short. The whole point of a smart search engine is to give me what I want. What I want is not what most want.
When Google tries to do this, the same people start complaining about filter bubbles[1] and either turn off personalization in their search settings, or turn to DDG, where a primary selling point is that they don't personalize. You really can't have it both ways, although Google comes very close with a simple toggle button for personalized results[2].
So as a monopoly it has started to ignore its users. It has even wound back features that were previously useful. Most of us could quickly list 10 things it could do to improve its service.
I don't believe you appreciate the difficulty of search given the current state of advanced [black hat] SEO; things that worked in the past (such as plain pagerank) would not work at all today. All search engines must run to keep in place. Also, economics plays a role -- can those 10 things be implemented in a practical way that scales and is cost effective.
I can 10 things on my car that I'd like, such as better fuel economy, more horsepower, better crash safety, better visibility, more convenience features, and a lower price. Unfortunately many of those things conflict, so in a practical sense it is likely that the car company had to strike a balance. From my armchair I am unlikely to know all of the things that went into those trade-offs. ~~
[1] "filter bubbles" don't really apply to multi-answer ranking problems or are trivially broken with standard techniques from reinforcement learning to manage the "explore-vs-exploit" tradeoff. As far as I've been able to determine, the person who coined the bubble term has no formal background in statistics (in particular ranking problems) or machine learning (in particular reinforcement learning).
[2] An oberservation from a long-time logged-in user: In my search results, personalization hardly ever effects more than two results out of the first 10. In search at least, filter bubbles do not exist for me, and I've taken no steps to avoid them. They do happen when I listen to a music service for a while (where unlike search, only one song can be chosen as the next to play).
Maybe I'm a curmudgeon, but I would rather tweak the search to narrow down crap results than try to outsmart the 'smartness' any day of the week. I understand that this isn't necessarily what John Q. Internetuser is looking for in search, but at least having the option there would be a big help.
It's really not that private stuff; here's four categories from my profile: Business & Industrial Business News Computer & Video Games Computer Components
The idea of a super-detailed profile is something with no original source, it has just been copied around the internet long enough that everyone accepts it as true. Of course you can claim that I'm not trustworthy, so below is an argument using only economics and public information.
There is no economic justification for a hyper-detailed profile. Here is why:
(1) Advertisers don't write ads for demographics so specific that there are only one or a few people in it. It is only in your interest to show to categories where many people apply, otherwise you are wasting your effort for no gain. Thus the worth of a profile is only in generalities. (2) Specific keywords can be handled when the query is made or the page/email is shown. Just about all internet advertising is just-in-time like this, since anything else involves lots of serving-accessible storage which costs money. Even then, if the keyword only applies to a few people, the advertiser is wasting time as per #1. (3) Every computation costs money. In advertising, if the cost to compute > incremental profit, you don't do it. The worth of a profile is only in its generalities as per #1, so that's the only thing worth computing, storing, and retrieving. (4) If having a detailed profile on everyone was the holy grail of advertising, facebook would be making a lot more money per page view.
Go back and read the section titled Relative position of the Sun to the center of the Galaxy and 14 pulsars, which has this sentence in particular:
If the plaque is found, only some of the pulsars may be visible from the location of its discovery. Showing the location with as many as 14 pulsars provides redundancy so that the location of the origin can be triangulated even if only some of the pulsars are recognized.
Given the distance of the pulsars, it is a pretty good bet that at least 4 would be visible by a hypothetical finder of the plaque.
For a moving spacecraft, you could easily seed it with these 14 pulsars, and run a SLAM[1] algorithm to add new ones and fix their position as you move. Localization with an initially unknown set of point beacons is well studied.
Now, there might indeed be new stuff on top of this in the paper, but the slashdot summary certainly isn't revealing it.
The system is not aware of what is happening around it.
Here's a video from two years ago, linked to the location in the talk about pedestrian, car, cyclist, and traffic light detection: https://www.youtube.com/watch?v=YXylqtEQ0tk&t=390 Around 9 minutes it shows how it all comes together to make a left turn at an intersection with many cars and pedestrians.
That was all two years ago. You're right that there's a long way to go, but describing it as an "auto-braking system" is extremely inaccurate.
ask your employers and coworkers how you can improve.
I doubt they'd admit that they'd want me to be a worse programmer, so as not to draw attention to their own laziness/incompetence and (statistically) lack of engagement at work.
A software engineering job is between 25% and 50% programming / programming ability. First, there are design and specifications to meet client requirements. If you want any but the lowest-paying programming jobs you need to be a part of that specification process. Same goes with milestone setting, scheduling, and assignment of developer resources. Yes, various parts of management will do a lot of that, but they cannot do that in a vacuum -- guidance from technical jobs is needed to keep things realistic. Finally, as part of designing, planning, implementing, and documenting, you'll need to communicate with peers so that they can understand and make use or your work, and work with management to understand the need and importance of each part.
For the past six years I've worked at a company that does yearly peer reviews, which I've found quite helpful. In none of those reviews has someone stated that I need to program better or more skillfully. Yet in all of those reviews I've gotten constructive feedback -- on how to improve in the *other* skills that a more senior software engineer will need.
I hope you are able to find the stable job you seek. While I can't claim to understand the details, a small change in attitude toward supporting skills may be what is needed to get you closer to your goal.
137 acre plant != 137 acres of solar panels Large solar setups need roads to access the panels, and if they are tilted it'll need space between panels to avoid wasted panel area from shadows.
Here's a similar but older plant: http://en.wikipedia.org/wiki/Nellis_Solar_Power_Plant
14 MW, 140 acres, 30 GW*h/year, built in 2007 Note in the photos how much sun hits the dirt (i.e. not on panels within the 140 acre plant).
So it's the right ballpark for a newer plant of the same size (but with better solar panels or packing) to be 18MW in 137 acres. I think you are right that it isn't optimistic, 43 GW*h/year sounds perfectly reasonable for a plant built 6 years later.
You got it all wrong. Big-O is indeed about the tight upper bound, and the complexity of the input size. And as the number of operations increase, you bet your ass that it will be particularly useful. Oh you bet your ass.
I work at Google, and have no idea where you came up with your claims.
i hear from acquaintances who work in Google that the algorithms they run on emails do something much like this. among other things, they know when you are thinking of taking another job almost before you do.
While I cannot disprove that HR is running sentiment analysis, we have company-wide surveys every year that they could use, biannual reviews by co-workers, and quarterly short reviews from managers. All of those probably have much higher signal/noise ratio than rummaging through peoples' email. Also, they type of people who can do that kind of NLP are probably better off working on NLP-related areas that help the company such as Android, Search, or Ads.
word is, among the things you must not say on the phone inside the pure-freedom, do-no-evil world of Google, is "let's take this offline" or anything else indicating you don't want to talk about something on the phone, since that's an instant tip that you want to say something unsurveilled. coming soon to our entire society!
This is not at all true. First of all, for internal communications hardly anyone uses phones anymore -- chat, voice chat, and hangouts are a simpler and faster options. The only people frequently on the phone are those talking to external people (sales folks, customer reps, etc). I guess those "calls could be monitored for quality" but that has little to do with the panopticon-like claim you are making.
Also, in the context of meetings, people say the phrase "let's take this offline" all the time, indicating that they don't want to start some (possibly long) side discussion in a meeting with multiple participants.
It works like this: Google makes you scan several of your friends in several outfits and tag them.
Now Google has a database of, your friends and social circle; your friends faces; your friends cloth shopping habits for direct ad targeting
The original sources in the TFA don't agree with you: This fingerprint is constructed by a smartphone app which snaps a series of photos of the user as they read web pages, emails or tweets. It then creates a file – called a spatiogram – that captures the spatial distribution of colours, textures and patterns (vertical or horizontal stripes, say) of the clothes they are wearing. This combination of colour, texture and pattern analysis makes someone easier to identify at odd viewing angles or over long distances.
Usefully, in terms of protecting people's privacy, the fingerprint changes every time you change your clothes, so you can be anonymous again whenever you wish.
"A person's visual fingerprint is only temporary, say for a day or an evening," says Nelakuditi.
And you have nothing because this feature will probably only work 5% of the time
Also from the TFA: In early tests using 15 volunteers, the team identified people 93 per cent of the time, even when they had their backs to the headset user.
I guess it is easier to make stuff up than to read.
The example using a Windows drive letter was probably not added by an engineer, since the use of Windows in engineering is very close to zero. Regardless of who added it, it still serves its purpose as a short example, so there wasn't a reason to change it.
FWIW, Google's most common internal date format is:
YYYYMMDDxHHMMSS Where x is "s" for standard and "d" for daylight savings time. It sorts correctly even across DST shifts.
Of course, using dates for any reason other than informational hints is not considered a good idea -- time and date will vary over the globe at any given point. All data files will have internal records with unambiguous timestamps, but the name gives hints to someone doing maintenance work. If need be, it is also good enough for very coarse cleanup rules (delete stuff > 30 days).
If the government started raiding your bank account, the correct reaction would not be to store your money in your mattress, nor to ask that banks hold less money.
The government can't search the inside of my locked car just because it is parked on the street, but when effectively the same thing happens electronically it's open season.
Property laws need to be updated to reflect the reality that many people want to store their email, photos, and other account data on a server somewhere.
The lawmakers have quite purposely dragged their feet on updating the relevant laws, while law enforcement uses outdated legal analogies to exploit loopholes in laws that didn't foresee the present reality.
One thing that I think trips up people who used web search for a long time is that you drop words you don't think are important for keyword searches, but that actually hurts now that search engines use more than keywords. Keyword spam killed keyword search a decade ago, and regular people could not use pure keyword search anyway; so now (whether we like it or not) all search engines try to operate at more of a semantic level. If you go in with that mindset you can still find almost anything within a few tries.
As the web and search engines both evolve, you may need to change the way you search to get the same information. Something that worked before may not work now, and the critical words or phrases to get the best results are still there but they aren't the same as what they were in the past.
In your particular example, the exclusion is far too weak, as "women's blazers" matches [blazers -ladies], and once you start using unusual queries (very few people will use exclusions) the search engine will tend to be more literal since it doesn't have the statistics of many previous searchers using those terms to go on.
Google Now is an example of a product that could not exist without data sharing. The premise is that it cross references data to make timely suggestions, such as letting you know when you should leave for the airport if you have a flight, and if your flight is on time. It can do this even though you never explicitly told it you have a flight or made a calendar entry.
Well, the 90s are over too, and we have larger datasets now. With "large scale" SVMs still being measured in 10s of thousands of examples, you can see why companies with 4 orders of magnitude more *users* (let alone data items to classify) would need to use better scaling techniques. The older algorithms, when coupled with more modern minimizers, tend to fare well in comparison to the much smaller models you can train with more advanced techniques.
Also, as a researcher, you should recognize the adage about the actual order of importance for getting machine learning to work: (1) picking the right features. (2) getting enough data (3) the learning algorithm
People love to talk at length about picking "the best" #3, when really you need to consider answers for #3 that let you do well on #2 and #1.
While I was a bit surprised to hear this Google project used networks (though not backprop trained NNs btw, which was the 80s fad), Andrew Ng is on the author list and he's a pretty smart guy (if you've done anything with reinforcement learning in the past 10 years you've probably run across his work). So I'm pretty sure they considered various options before they built something to run on 16K cpu cores.
Slashdot Mirror
The Goobuntu part is incorrect. Both my laptop and desktop have GCC 4.6.3 installed right now.
It is exceedingly unlikely that the results don't overlap after the first few, but if you can produce a copy of the two sets of results, I will forward them to someone on the Google Search team for debugging.
People hugely overestimate the effect of personalization -- it is a ranking tweak not a complete change to the search engine. It does not make economic sense to have personalized whole-web indexes.
Btw, if you don't like personalization ever, it is pretty easy to turn off:
https://support.google.com/accounts/answer/54048?hl=en
Just remove web history and uncheck private results.
Decisions must be made based upon vigorous consideration of real world conditions and forces not abstract philosophy.
This is a very good argument against government intervention. A government made of lawyers is not a good substitute for scientists and economists.
This seems to be the most accurate post on the topic, yet carries the lowest score. This is about making suggestions, not auto-sending messages.
Typing on a phone is annoying, so if I can say "on my way home" with fewer clicks, instead of having to retype the message all the time, I would be happy to do so. Of course, I could set up some kind of macro, but an automatic system is far easier for normal folks to use.
Of course, whether this should be patentable is an entirely different matter, but the feature is an entirely reasonable thing to try.
If you assume a linear traction-limited model[1], information only needs to be sent backwards. Specifically:
(a) current velocity
(b) current acceleration/deceleration.
(c) maximum deceleration.
The immediately following car just needs to measure the distance[2], and know its own state and parameters. Then it can calculate how much space it needs to stop if the car in front immediately starts breaking at the maximum rate. You can incorporate communication & reaction delays easily too, as well as any bounded noise on the state variables. This would work for an arbitrarily long chain of cars, with each car just needing to monitor the one in front of it. It even works for autonomous cars following human-driven cars[3].
I used to be a robotics researcher, with a focus on high performance navigation. The lane-free full 2D generalization of the problem above was a chapter in my thesis (from 2007).
[1] or just make sure your actual model fits within a conservative linear envelope.
[2] obviously you can estimate a&b, but it potentially introduces additional delay to get the noise down, in particular for acceleration since that is second order.
[3] this is an ideal case *IF* that driver is paying attention, since the human driver has a better perception system.
(I work at Google, but not on search)
I'm afraid the idea, often expressed in this discussion, of "that's what most people want" sells us short. The whole point of a smart search engine is to give me what I want. What I want is not what most want.
When Google tries to do this, the same people start complaining about filter bubbles[1] and either turn off personalization in their search settings, or turn to DDG, where a primary selling point is that they don't personalize. You really can't have it both ways, although Google comes very close with a simple toggle button for personalized results[2].
So as a monopoly it has started to ignore its users. It has even wound back features that were previously useful. Most of us could quickly list 10 things it could do to improve its service.
I don't believe you appreciate the difficulty of search given the current state of advanced [black hat] SEO; things that worked in the past (such as plain pagerank) would not work at all today. All search engines must run to keep in place. Also, economics plays a role -- can those 10 things be implemented in a practical way that scales and is cost effective.
I can 10 things on my car that I'd like, such as better fuel economy, more horsepower, better crash safety, better visibility, more convenience features, and a lower price. Unfortunately many of those things conflict, so in a practical sense it is likely that the car company had to strike a balance. From my armchair I am unlikely to know all of the things that went into those trade-offs.
~~
[1] "filter bubbles" don't really apply to multi-answer ranking problems or are trivially broken with standard techniques from reinforcement learning to manage the "explore-vs-exploit" tradeoff. As far as I've been able to determine, the person who coined the bubble term has no formal background in statistics (in particular ranking problems) or machine learning (in particular reinforcement learning).
[2] An oberservation from a long-time logged-in user: In my search results, personalization hardly ever effects more than two results out of the first 10. In search at least, filter bubbles do not exist for me, and I've taken no steps to avoid them. They do happen when I listen to a music service for a while (where unlike search, only one song can be chosen as the next to play).
Maybe I'm a curmudgeon, but I would rather tweak the search to narrow down crap results than try to outsmart the 'smartness' any day of the week. I understand that this isn't necessarily what John Q. Internetuser is looking for in search, but at least having the option there would be a big help.
There already is such an option, called "verbatim":
https://support.google.com/websearch/answer/1734130?hl=en
... crash ... thunk ....
Boooooooooooooom!
I work at Google. You can read (and edit) your own profile right here:
www.google.com/settings/ads
It's really not that private stuff; here's four categories from my profile:
Business & Industrial
Business News
Computer & Video Games
Computer Components
The idea of a super-detailed profile is something with no original source, it has just been copied around the internet long enough that everyone accepts it as true. Of course you can claim that I'm not trustworthy, so below is an argument using only economics and public information.
There is no economic justification for a hyper-detailed profile. Here is why:
(1) Advertisers don't write ads for demographics so specific that there are only one or a few people in it. It is only in your interest to show to categories where many people apply, otherwise you are wasting your effort for no gain. Thus the worth of a profile is only in generalities.
(2) Specific keywords can be handled when the query is made or the page/email is shown. Just about all internet advertising is just-in-time like this, since anything else involves lots of serving-accessible storage which costs money. Even then, if the keyword only applies to a few people, the advertiser is wasting time as per #1.
(3) Every computation costs money. In advertising, if the cost to compute > incremental profit, you don't do it. The worth of a profile is only in its generalities as per #1, so that's the only thing worth computing, storing, and retrieving.
(4) If having a detailed profile on everyone was the holy grail of advertising, facebook would be making a lot more money per page view.
Go back and read the section titled Relative position of the Sun to the center of the Galaxy and 14 pulsars, which has this sentence in particular:
If the plaque is found, only some of the pulsars may be visible from the location of its discovery. Showing the location with as many as 14 pulsars provides redundancy so that the location of the origin can be triangulated even if only some of the pulsars are recognized.
Given the distance of the pulsars, it is a pretty good bet that at least 4 would be visible by a hypothetical finder of the plaque.
For a moving spacecraft, you could easily seed it with these 14 pulsars, and run a SLAM[1] algorithm to add new ones and fix their position as you move. Localization with an initially unknown set of point beacons is well studied.
Now, there might indeed be new stuff on top of this in the paper, but the slashdot summary certainly isn't revealing it.
[1] http://en.wikipedia.org/wiki/Simultaneous_localization_and_mapping
The system is not aware of what is happening around it.
Here's a video from two years ago, linked to the location in the talk about pedestrian, car, cyclist, and traffic light detection:
https://www.youtube.com/watch?v=YXylqtEQ0tk&t=390
Around 9 minutes it shows how it all comes together to make a left turn at an intersection with many cars and pedestrians.
That was all two years ago. You're right that there's a long way to go, but describing it as an "auto-braking system" is extremely inaccurate.
ask your employers and coworkers how you can improve.
I doubt they'd admit that they'd want me to be a worse programmer, so as not to draw attention to their own laziness/incompetence and (statistically) lack of engagement at work.
A software engineering job is between 25% and 50% programming / programming ability. First, there are design and specifications to meet client requirements. If you want any but the lowest-paying programming jobs you need to be a part of that specification process. Same goes with milestone setting, scheduling, and assignment of developer resources. Yes, various parts of management will do a lot of that, but they cannot do that in a vacuum -- guidance from technical jobs is needed to keep things realistic. Finally, as part of designing, planning, implementing, and documenting, you'll need to communicate with peers so that they can understand and make use or your work, and work with management to understand the need and importance of each part.
For the past six years I've worked at a company that does yearly peer reviews, which I've found quite helpful. In none of those reviews has someone stated that I need to program better or more skillfully. Yet in all of those reviews I've gotten constructive feedback -- on how to improve in the *other* skills that a more senior software engineer will need.
I hope you are able to find the stable job you seek. While I can't claim to understand the details, a small change in attitude toward supporting skills may be what is needed to get you closer to your goal.
137 acre plant != 137 acres of solar panels
Large solar setups need roads to access the panels, and if they are tilted it'll need space between panels to avoid wasted panel area from shadows.
Here's a similar but older plant:
http://en.wikipedia.org/wiki/Nellis_Solar_Power_Plant
14 MW, 140 acres, 30 GW*h/year, built in 2007
Note in the photos how much sun hits the dirt (i.e. not on panels within the 140 acre plant).
So it's the right ballpark for a newer plant of the same size (but with better solar panels or packing) to be 18MW in 137 acres. I think you are right that it isn't optimistic, 43 GW*h/year sounds perfectly reasonable for a plant built 6 years later.
You got it all wrong. Big-O is indeed about the tight upper bound, and the complexity of the input size. And as the number of operations increase, you bet your ass that it will be particularly useful. Oh you bet your ass.
GP is being an ass, and doesn't seem to understand what "asymptotic complexity" means. However, you are incorrect about big-O, which does not need to be a tight bound. You're thinking of big-theta. Wikipedia has a concise summary:
https://en.wikipedia.org/wiki/Big_theta#Family_of_Bachmann.E2.80.93Landau_notations
A little too late for my tastes.
I work at Google, and have no idea where you came up with your claims.
i hear from acquaintances who work in Google that the algorithms they run on emails do something much like this. among other things, they know when you are thinking of taking another job almost before you do.
While I cannot disprove that HR is running sentiment analysis, we have company-wide surveys every year that they could use, biannual reviews by co-workers, and quarterly short reviews from managers. All of those probably have much higher signal/noise ratio than rummaging through peoples' email. Also, they type of people who can do that kind of NLP are probably better off working on NLP-related areas that help the company such as Android, Search, or Ads.
word is, among the things you must not say on the phone inside the pure-freedom, do-no-evil world of Google, is "let's take this offline" or anything else indicating you don't want to talk about something on the phone, since that's an instant tip that you want to say something unsurveilled. coming soon to our entire society!
This is not at all true. First of all, for internal communications hardly anyone uses phones anymore -- chat, voice chat, and hangouts are a simpler and faster options. The only people frequently on the phone are those talking to external people (sales folks, customer reps, etc). I guess those "calls could be monitored for quality" but that has little to do with the panopticon-like claim you are making.
Also, in the context of meetings, people say the phrase "let's take this offline" all the time, indicating that they don't want to start some (possibly long) side discussion in a meeting with multiple participants.
Wrong.
It works like this: Google makes you scan several of your friends in several outfits and tag them.
Now Google has a database of, your friends and social circle; your friends faces; your friends cloth shopping habits for direct ad targeting
The original sources in the TFA don't agree with you:
This fingerprint is constructed by a smartphone app which snaps a series of photos of the user as they read web pages, emails or tweets. It then creates a file – called a spatiogram – that captures the spatial distribution of colours, textures and patterns (vertical or horizontal stripes, say) of the clothes they are wearing. This combination of colour, texture and pattern analysis makes someone easier to identify at odd viewing angles or over long distances.
Usefully, in terms of protecting people's privacy, the fingerprint changes every time you change your clothes, so you can be anonymous again whenever you wish.
"A person's visual fingerprint is only temporary, say for a day or an evening," says Nelakuditi.
And you have nothing because this feature will probably only work 5% of the time
Also from the TFA:
In early tests using 15 volunteers, the team identified people 93 per cent of the time, even when they had their backs to the headset user.
I guess it is easier to make stuff up than to read.
...and it's a damn shame. I miss this aspect of what the Democrats used to stand for. Now, both parties are led by hawks.
The example using a Windows drive letter was probably not added by an engineer, since the use of Windows in engineering is very close to zero. Regardless of who added it, it still serves its purpose as a short example, so there wasn't a reason to change it.
FWIW, Google's most common internal date format is:
YYYYMMDDxHHMMSS
Where x is "s" for standard and "d" for daylight savings time. It sorts correctly even across DST shifts.
Of course, using dates for any reason other than informational hints is not considered a good idea -- time and date will vary over the globe at any given point. All data files will have internal records with unambiguous timestamps, but the name gives hints to someone doing maintenance work. If need be, it is also good enough for very coarse cleanup rules (delete stuff > 30 days).
If the government started raiding your bank account, the correct reaction would not be to store your money in your mattress, nor to ask that banks hold less money.
The government can't search the inside of my locked car just because it is parked on the street, but when effectively the same thing happens electronically it's open season.
Property laws need to be updated to reflect the reality that many people want to store their email, photos, and other account data on a server somewhere.
The lawmakers have quite purposely dragged their feet on updating the relevant laws, while law enforcement uses outdated legal analogies to exploit loopholes in laws that didn't foresee the present reality.
There seems to be very little misunderstanding if I just type your actual question:
https://www.google.com/search?q=convert+from+a+WPF+Visual+to+a+Windows+Metafile
One thing that I think trips up people who used web search for a long time is that you drop words you don't think are important for keyword searches, but that actually hurts now that search engines use more than keywords. Keyword spam killed keyword search a decade ago, and regular people could not use pure keyword search anyway; so now (whether we like it or not) all search engines try to operate at more of a semantic level. If you go in with that mindset you can still find almost anything within a few tries.
Try this query instead:
https://www.google.com/#q=men's+blazers
The entire first page is full of items that are exactly what you are looking for.
As the web and search engines both evolve, you may need to change the way you search to get the same information. Something that worked before may not work now, and the critical words or phrases to get the best results are still there but they aren't the same as what they were in the past.
In your particular example, the exclusion is far too weak, as "women's blazers" matches [blazers -ladies], and once you start using unusual queries (very few people will use exclusions) the search engine will tend to be more literal since it doesn't have the statistics of many previous searchers using those terms to go on.
Google Now is an example of a product that could not exist without data sharing. The premise is that it cross references data to make timely suggestions, such as letting you know when you should leave for the airport if you have a flight, and if your flight is on time. It can do this even though you never explicitly told it you have a flight or made a calendar entry.
Well, the 90s are over too, and we have larger datasets now. With "large scale" SVMs still being measured in 10s of thousands of examples, you can see why companies with 4 orders of magnitude more *users* (let alone data items to classify) would need to use better scaling techniques. The older algorithms, when coupled with more modern minimizers, tend to fare well in comparison to the much smaller models you can train with more advanced techniques.
Also, as a researcher, you should recognize the adage about the actual order of importance for getting machine learning to work:
(1) picking the right features.
(2) getting enough data
(3) the learning algorithm
People love to talk at length about picking "the best" #3, when really you need to consider answers for #3 that let you do well on #2 and #1.
While I was a bit surprised to hear this Google project used networks (though not backprop trained NNs btw, which was the 80s fad), Andrew Ng is on the author list and he's a pretty smart guy (if you've done anything with reinforcement learning in the past 10 years you've probably run across his work). So I'm pretty sure they considered various options before they built something to run on 16K cpu cores.
You can read the ICML paper here:
http://research.google.com/pubs/pub38115.html
Here's an example where you can really tell the difference when compared to the earlier lower-bitrate sample.