Don't sign up casually. I've done the machine learning class on line. There is a lot of homework. Expect to spend at least 8 hours a week on the class. Also, the videos consist of Andrew Ng writing math on a chalkboard. An actual chalkboard. In a weird notation where indices are superscripts, rather than subscripts.
As I point out occasionally, many, if not most, of the problems with web spam, phishing, etc. on the web are because Google doesn't verify the identity of the business behind a web site.
Businesses don't have any right to anonymity. Even in Europe. In the European Union, businesses come under the European Directive on Electronic Commerce.: "Member States shall ensure that the service provider (defined as "any natural or legal person providing an information society service" i.e. a web site) shall render easily, directly and permanently accessible to the recipients of the service and competent authorities, at least the following information: (a) the name of the service provider; (b) the geographic address at which the service provider is established... (c) his electronic mail address...". The European Privacy Directive is only for individuals. If the search end of Google took a hard line on that, search would be much less spammy. Currently, they can't even keep totally fake business locations out of Google Places. Yes, "Illusory Laptop Repair is still in Google Places, right in the middle of the railroad crossing. So are so many phony business locations that it's been covered at length in the New York Times. Legitimate local businesses are screaming about this; customers try to find them and end up calling some outsourced lead-generation service, thinking it's a local company.
Google wants to use Google+ for "crowdsourcing" recommendations. They used to use Citysearch and Yelp for that, but those became too polluted with fake recommendations. The trouble with "crowdsourcing" is that crowds can be sourced. You can buy "likes", "recommendations", and "+1"s in bulk on any of the black hat SEO forums.
Recommendation systems only work in three situations - when the number of reviewers is huge compared to the number of items being reviewed, as with movies, when the reviewer is known to have bought the product, as with eBay and Amazon, and when the reviewer's identity is verified and their reputation is known. Google seems to be trying for #3. To make that work, they have to tighten the screws on "Google+" users. Tightening the screws on businesses would be more productive.
Isn't this really more about publishers than bookstores, though?
It is. We'll probably still have a publishing industry, but it will publish the few mass-market titles that appear in racks at non-bookstore retailers, like drug stores and supermarkets. Everything else will be remote-order or on line only. For those books, the publisher has a very limited role and function.
What competing bookstores? There's Barnes and Noble, and a few remaining independents. Borders is in bankruptcy liquidation. ("Everything must go! 40-60% off! Store fixtures for sale.) Barnes and Noble is in financial trouble. When they go, there won't be much left.
When the big guys give it up, the distribution channel dries up.All the warehousing and shipping needed to service little bookstores isn't profitable if the volume is too small.
Bookstores are going the way of record stores and video rental stores. Gone. It's sad, but probably inevitable.
Migration of work from one physical machine to another is a useful application of visualization. But to some extent it reflects that few OSs today are real network operating systems. UCLA Locus and Tandem NonStop were operating systems that could migrate jobs from one machine to another. No mainstream OS today does that, although virtualization systems do.
Yet another unneeded thousand page book of examples. The Python library documentation isn't bad, as such things go. Nor is the library that complex. Most of the library modules are independent of other modules, and have APIs of modest size. Another giant cookbook is unnecessary.
There are too many of these morbidly obese programming books out there. Little pamphlet-sized books would be more useful. Printing long code examples on paper is just silly; programmers who want to cut and paste need an online version.Those plastic cards which cram what you really need to know on one sheet would be even more useful for many programs.
Recorded human history is about 3000 years long, but technological civilization is only about 200. Only in the last 100 years have humans been able to make a significant dent in the planet's resources.
Most of the major resources that make our industrial civilization go run out within 50 to 200 years. There hasn't been a new energy source in 50 years. Absent new physics, industrial civilization is more than half over. If we're lucky, there's a post-industrial bio-based society ahead. If not, it's regression to pre-industrial society or worse.
That's probably why we haven't had any contact with other civilizations. Each industrial civilization has its brief moment, and then the planet reverts to normal, but with all the concentrated resources mined out, so there's no second chance. Put 400 years as a civilization lifetime ("L") into the Drake equation.
Google used to have a problem with malware and phishing sites being hosted on their own Google Sites. Once they plugged that hole, the malware moved to Google Spreadsheets. Because you can put HTML in Google's spreadsheets, it can be used as a free hosting service. Google hadn't anticipated this, and their abuse operation couldn't handle it.
Google seems to have plugged the spreadsheet hole now.
I noticed recently that Google has disappeared from our major domains being exploited by active phishing scams. There were pages hosted by Google which were in Google's own "this site may harm your computer" blacklist. So their hostile-site detection wasn't coupled to their abuse department.
That was kind of embarrassing, but until we publicized it, it didn't get fixed.
Basic truth - run a free hosting service, a free "forms" service, a free "poll" service, or a free URL direction service, and you will end up hosting phishing sites and related annoyances. If you run a free service, you must have an automatic check against the major phishing blacklists, or you will be pwned. This
week's big sites being abused by phishers are "piczo.com" (social networking for teenage girls), "webs.com" and "moonfruit.com" (free hosting), and "t35.com" (which, the last time we contacted them, has one poor abuse guy trying to deal with a daily tide of phishing pages by hand).
Those sites are used by the bottom-feeders of the malware/phishing world. The big guys buy hosting with stolen credit card numbers, use botnets, and contract with "bullet-proof hosting services.
There is progress. "Open redirectors" are more or less gone from major sites. Over the last three years, MSN, Yahoo, eBay, and Facebook have all had open redirectors. Publicity and nagging put a stop to that. Slowly, too slowly, IE6 is dying out.
So, the only way to guarantee a native code program on current machines is... equivalent to solving the halting problem.
Which isn't that hard for most useful programs. The Microsoft Static Driver Verifier, used on all signed Windows 7 drivers, can decide safety about 97% of the time within its standard time limit. If your driver's memory safety is undecidable, or even hard to decide, it's broken.
Since the x86 has deprecated segmentation (which could be used for this)...
That's how the Google Native Client works, and that's why it supports 32-bit x86 code only.
No, this is how engineering moves forward if you have enough money. In the 1940s and 1950s, a huge number of experimental aircraft and rockets were built. Some worked, some didn't, and some went through a large number of prototypes before they worked. There were terrible problems getting early jet fighters to work right. A lot of test pilots died. Even the successful military planes weren't that safe; in the 1950s, a Navy pilot had about a 1 in 5 chance of dying in a crash, without help from the enemy.
In the early days of rocketry, a huge number of rockets were launched unsuccessfully. About 600 V-2 rocket launches were attempted in the R&D phase, before they were able to hit London. ICBM development in the US and USSR had dozens of launch failures. Frequent launches were expensive, but projects were completed faster.
You have a program you'd like to run.
You make a list of the files and other resources it should have, and the types of access
You hand that list to the program, which uses the operating system to access them, strictly limited to the items on the list.
This prevents unwanted side effects, and removes the bizarre (to me) need to trust programs at all.
You've just described IBM mainframe Job Control Language:
Classic JCL tells the program what files, etc. to access, and the program is limited to the resources the JCL gives it. The syntax is clunky, but there's something to be said for the concept. If phone apps worked that way, they'd be much easier to control and limit.
The article perpetuates the myth that native code has to be "unsafe". That's an artifact of C and C++. It's not true of Pascal, Ada, Modula, Delphi, Eiffel, Erlang, or Go.
Nor does subscript and pointer checking have to be expensive. Usually, it can be hoisted out of loops and checked once. Or, for many FOR loops, zero times, if the upper bound is derived from the array size.
One of the sad facts of programming is that there should have been a replacement for C/C++ by now. But nothing ever overcame the legacy code base of the UNIX/Linux world. Every day, millions of programs crash and millions of compromised machines have security breaches because of this.
Wireless devices ought to be wireless. They already have several radios, including Bluetooth. Headphones and docking should be wireless. So should charging. which should be inductive. Then you can have a sealed, waterproof unit with no annoying connector holes.
I'm surprised Apple hasn't already gone that way on aesthetic grounds. Why should those perfect forms have holes in them?
The embedded world doesn't have much trouble with this. For QNX, there's the kernel, which is the same for all CPUs with the same instructions set, and a "board support package", which has the driver programs for a given board or variant.
Linux is a monolithic kernel, and so it has to be hacked all over the place to deal with architecture variations. Linux lacks a clean conceptual model of operating system vs. board support.
S&P dropped their rating on Google stock from "buy" to "sell" after the Motorola acquisition, and knocked $200 off their one-year predicted price for Google stock. That's very unusual.
Google's track record with hardware is not good. They were in the direct sales phone handset business for only a few months before they had to exit it. Customers insisted that the hardware work, and wanted customer service when it didn't. Google couldn't handle that. Their approach to the "Google Search Appliance" (Mini size) is weird. There's no phone support for this rack-mounted enterprise device. If it breaks, they FedEx you a new one. After three years, the Google Search Appliance stops working and you have to buy a new one. Really. That's Google's approach to enterprise support. That won't fly with Motorola's customer base.
It's nice that they're doing this, but it's a bit late to be doing a planar walking/running machine. (It's supported so that it can't fall sideways, so it's a 2D balancing system instead of a 3D one, and is considered "planar", although it's going in a circle.) Compare Spring Flamingo, 1996-2000 at MIT.
This group doesn't seem to be addressing slip control or hills (which I've worked on), or twist control (first addressed by Jessica Hodgkins at Georgia Tech and by Honda). As soon as you go beyond flat, high-friction floors (these guys use rubber mats), slip control starts to dominate the problem. Take a look at Big Dog videos.
Twist control involves not inducing torso rotation, or at least cancelling it on the next half-stride, and is an additional problem bipeds face more than quadrupeds. Honda had a lot of trouble with that in the early days of Asimo. Planar bipeds can't twist, so this group gets to ignore that problem.
It's easier to work in this area than it was in the 1990s. Simulators are much better. CPUs are much faster. The theory is better. Control algorithms are better. Motors are better. Funding is better. There's no need to work on an oversimplified version of the problem any more.
This is another one of those "top N, one per page, ads on every page" ad farm trolls.
Their list isn't too impressive, either. Bell Labs, yes. IBM Watson, yes. PARC, yes. But where's the Moore School of Electrical Engineering, from which came ENIAC, and the beginnings of UNIVAC, the first commercial electronic computer to go into production? Also, Bletchly Park wasn't that influential because nobody knew about it until the 1970s.
What we call a "computer" today is properly a stored-program general purpose digital computer. There were machines built before that which had some, but not all, of those attributes.
Bletchley Park's machines fall into that category). The WWII US crypto operation was at Arlington Hall, which did more hardware development than Bletchley Park. were developed. They were using punched cards where Bletchley used people and filing cabinets, and they seem to have developed digital magnetic tape, although the history there is cloudy. NSA is the direct descendant of Arlington Hall.
Another major pre-computer computing company
was Teleregister, which was a spinoff from Western Union in 1949. They pioneered "remote computing" for stock quotations, railroad ticketing, and airline ticketing. Their Magnetronic Reservisor was the first big remote-access system, with magnetic drums holding the reservation data.
They're on a fixed schedule, available for a limited time, and require registration. You'd think they'd just upload them to the Internet Archive and be done with it. Talks for previous years are on "video.linux.com".
"Google ['s game revenue cut] is at 5% because they don't have any [gaming] users,"
Does Google+ require that games on their system use their billing system? Facebook does, but it's not clear if Google does. With Google's Schmidt commanded to testify before the Senate Judiciary Committee on September 21 on Google's antitrust issues, I'd be surprised if Google did require that. 5% is a reasonably competitive price for a payment service. At 30%, you look into handling your own credit card payments.
It's a special-purpose mail system, much like a corporate system. Content stays around forever, and it's only for sending to and from units of the Government.
The real issue is winding down the postal system. Postal systems worldwide are sending fewer letters each year.
Space-X may be the future of space travel. They designed that thing. It's not a NASA design, and it didn't go through NASA's process of spreading everything out among contractors spread across the US.
San Francisco has a lot of crazies, and sorting out the harmless ones from the dangerous ones is hard. Here's one case. A guy in a wheelchair was slashing the tires of city vehicles with a rock. He was apparently shot by cops with a beanbag gun, and he and one cop were taken to a hospital with non-lethal injuries. There's video from someone across the street. Excessive force? Perhaps, but subduing someone with a sharp object without getting cut is very tough.
Here's another case, of a known mentally ill woman shot in her home after threatening a social worker. This probably was excessive force; the cops could have waited for backup from the guys in heavier protective gear.
Here's the aftermath of the shooting the protests are about. This doesn't seem to have been a crazy, but a parolee from Portland who ran from cops after trying to ride a bus for free and being caught at it. Whether he had a gun remains an open question.
The SFPD is at best a mediocre department. NY and LA have really tried to clean up their act, but SF hasn't.
Slashdot Mirror
Don't sign up casually. I've done the machine learning class on line. There is a lot of homework. Expect to spend at least 8 hours a week on the class. Also, the videos consist of Andrew Ng writing math on a chalkboard. An actual chalkboard. In a weird notation where indices are superscripts, rather than subscripts.
As I point out occasionally, many, if not most, of the problems with web spam, phishing, etc. on the web are because Google doesn't verify the identity of the business behind a web site.
Businesses don't have any right to anonymity. Even in Europe. In the European Union, businesses come under the European Directive on Electronic Commerce.: "Member States shall ensure that the service provider (defined as "any natural or legal person providing an information society service" i.e. a web site) shall render easily, directly and permanently accessible to the recipients of the service and competent authorities, at least the following information: (a) the name of the service provider; (b) the geographic address at which the service provider is established ... (c) his electronic mail address...". The European Privacy Directive is only for individuals. If the search end of Google took a hard line on that, search would be much less spammy. Currently, they can't even keep totally fake business locations out of Google Places. Yes, "Illusory Laptop Repair is still in Google Places, right in the middle of the railroad crossing. So are so many phony business locations that it's been covered at length in the New York Times. Legitimate local businesses are screaming about this; customers try to find them and end up calling some outsourced lead-generation service, thinking it's a local company.
Google wants to use Google+ for "crowdsourcing" recommendations. They used to use Citysearch and Yelp for that, but those became too polluted with fake recommendations. The trouble with "crowdsourcing" is that crowds can be sourced. You can buy "likes", "recommendations", and "+1"s in bulk on any of the black hat SEO forums.
Recommendation systems only work in three situations - when the number of reviewers is huge compared to the number of items being reviewed, as with movies, when the reviewer is known to have bought the product, as with eBay and Amazon, and when the reviewer's identity is verified and their reputation is known. Google seems to be trying for #3. To make that work, they have to tighten the screws on "Google+" users. Tightening the screws on businesses would be more productive.
Told you so on February 6, 2009.
Back in 2009, it was just a Windows autorun problem. Since then, Google and Apple have been able to screw up in the same way.
Coming soon, I suppose, attacks on appliances via "smart meter" data links. Not everything should have a data link.
Isn't this really more about publishers than bookstores, though?
It is. We'll probably still have a publishing industry, but it will publish the few mass-market titles that appear in racks at non-bookstore retailers, like drug stores and supermarkets. Everything else will be remote-order or on line only. For those books, the publisher has a very limited role and function.
What competing bookstores? There's Barnes and Noble, and a few remaining independents. Borders is in bankruptcy liquidation. ("Everything must go! 40-60% off! Store fixtures for sale.) Barnes and Noble is in financial trouble. When they go, there won't be much left.
When the big guys give it up, the distribution channel dries up.All the warehousing and shipping needed to service little bookstores isn't profitable if the volume is too small.
Bookstores are going the way of record stores and video rental stores. Gone. It's sad, but probably inevitable.
Migration of work from one physical machine to another is a useful application of visualization. But to some extent it reflects that few OSs today are real network operating systems. UCLA Locus and Tandem NonStop were operating systems that could migrate jobs from one machine to another. No mainstream OS today does that, although virtualization systems do.
Yet another unneeded thousand page book of examples. The Python library documentation isn't bad, as such things go. Nor is the library that complex. Most of the library modules are independent of other modules, and have APIs of modest size. Another giant cookbook is unnecessary.
There are too many of these morbidly obese programming books out there. Little pamphlet-sized books would be more useful. Printing long code examples on paper is just silly; programmers who want to cut and paste need an online version.Those plastic cards which cram what you really need to know on one sheet would be even more useful for many programs.
Recorded human history is about 3000 years long, but technological civilization is only about 200. Only in the last 100 years have humans been able to make a significant dent in the planet's resources.
Most of the major resources that make our industrial civilization go run out within 50 to 200 years. There hasn't been a new energy source in 50 years. Absent new physics, industrial civilization is more than half over. If we're lucky, there's a post-industrial bio-based society ahead. If not, it's regression to pre-industrial society or worse.
That's probably why we haven't had any contact with other civilizations. Each industrial civilization has its brief moment, and then the planet reverts to normal, but with all the concentrated resources mined out, so there's no second chance. Put 400 years as a civilization lifetime ("L") into the Drake equation.
Google used to have a problem with malware and phishing sites being hosted on their own Google Sites. Once they plugged that hole, the malware moved to Google Spreadsheets. Because you can put HTML in Google's spreadsheets, it can be used as a free hosting service. Google hadn't anticipated this, and their abuse operation couldn't handle it.
Google seems to have plugged the spreadsheet hole now. I noticed recently that Google has disappeared from our major domains being exploited by active phishing scams. There were pages hosted by Google which were in Google's own "this site may harm your computer" blacklist. So their hostile-site detection wasn't coupled to their abuse department. That was kind of embarrassing, but until we publicized it, it didn't get fixed.
Basic truth - run a free hosting service, a free "forms" service, a free "poll" service, or a free URL direction service, and you will end up hosting phishing sites and related annoyances. If you run a free service, you must have an automatic check against the major phishing blacklists, or you will be pwned. This week's big sites being abused by phishers are "piczo.com" (social networking for teenage girls), "webs.com" and "moonfruit.com" (free hosting), and "t35.com" (which, the last time we contacted them, has one poor abuse guy trying to deal with a daily tide of phishing pages by hand).
Those sites are used by the bottom-feeders of the malware/phishing world. The big guys buy hosting with stolen credit card numbers, use botnets, and contract with "bullet-proof hosting services.
There is progress. "Open redirectors" are more or less gone from major sites. Over the last three years, MSN, Yahoo, eBay, and Facebook have all had open redirectors. Publicity and nagging put a stop to that. Slowly, too slowly, IE6 is dying out.
So, the only way to guarantee a native code program on current machines is... equivalent to solving the halting problem.
Which isn't that hard for most useful programs. The Microsoft Static Driver Verifier, used on all signed Windows 7 drivers, can decide safety about 97% of the time within its standard time limit. If your driver's memory safety is undecidable, or even hard to decide, it's broken.
Since the x86 has deprecated segmentation (which could be used for this)...
That's how the Google Native Client works, and that's why it supports 32-bit x86 code only.
This is how science moves forward.
No, this is how engineering moves forward if you have enough money. In the 1940s and 1950s, a huge number of experimental aircraft and rockets were built. Some worked, some didn't, and some went through a large number of prototypes before they worked. There were terrible problems getting early jet fighters to work right. A lot of test pilots died. Even the successful military planes weren't that safe; in the 1950s, a Navy pilot had about a 1 in 5 chance of dying in a crash, without help from the enemy.
In the early days of rocketry, a huge number of rockets were launched unsuccessfully. About 600 V-2 rocket launches were attempted in the R&D phase, before they were able to hit London. ICBM development in the US and USSR had dozens of launch failures. Frequent launches were expensive, but projects were completed faster.
Do Apple products go through the US to go to Australia, or are they shipped directly from Hon Hai in Shenzen, where they're made?
You have a program you'd like to run.
You make a list of the files and other resources it should have, and the types of access
You hand that list to the program, which uses the operating system to access them, strictly limited to the items on the list.
This prevents unwanted side effects, and removes the bizarre (to me) need to trust programs at all.
You've just described IBM mainframe Job Control Language:
Classic JCL tells the program what files, etc. to access, and the program is limited to the resources the JCL gives it. The syntax is clunky, but there's something to be said for the concept. If phone apps worked that way, they'd be much easier to control and limit.
The article perpetuates the myth that native code has to be "unsafe". That's an artifact of C and C++. It's not true of Pascal, Ada, Modula, Delphi, Eiffel, Erlang, or Go.
Nor does subscript and pointer checking have to be expensive. Usually, it can be hoisted out of loops and checked once. Or, for many FOR loops, zero times, if the upper bound is derived from the array size.
One of the sad facts of programming is that there should have been a replacement for C/C++ by now. But nothing ever overcame the legacy code base of the UNIX/Linux world. Every day, millions of programs crash and millions of compromised machines have security breaches because of this.
Wireless devices ought to be wireless. They already have several radios, including Bluetooth. Headphones and docking should be wireless. So should charging. which should be inductive. Then you can have a sealed, waterproof unit with no annoying connector holes.
I'm surprised Apple hasn't already gone that way on aesthetic grounds. Why should those perfect forms have holes in them?
The embedded world doesn't have much trouble with this. For QNX, there's the kernel, which is the same for all CPUs with the same instructions set, and a "board support package", which has the driver programs for a given board or variant.
Linux is a monolithic kernel, and so it has to be hacked all over the place to deal with architecture variations. Linux lacks a clean conceptual model of operating system vs. board support.
S&P dropped their rating on Google stock from "buy" to "sell" after the Motorola acquisition, and knocked $200 off their one-year predicted price for Google stock. That's very unusual.
Google's track record with hardware is not good. They were in the direct sales phone handset business for only a few months before they had to exit it. Customers insisted that the hardware work, and wanted customer service when it didn't. Google couldn't handle that. Their approach to the "Google Search Appliance" (Mini size) is weird. There's no phone support for this rack-mounted enterprise device. If it breaks, they FedEx you a new one. After three years, the Google Search Appliance stops working and you have to buy a new one. Really. That's Google's approach to enterprise support. That won't fly with Motorola's customer base.
It's nice that they're doing this, but it's a bit late to be doing a planar walking/running machine. (It's supported so that it can't fall sideways, so it's a 2D balancing system instead of a 3D one, and is considered "planar", although it's going in a circle.) Compare Spring Flamingo, 1996-2000 at MIT.
This group doesn't seem to be addressing slip control or hills (which I've worked on), or twist control (first addressed by Jessica Hodgkins at Georgia Tech and by Honda). As soon as you go beyond flat, high-friction floors (these guys use rubber mats), slip control starts to dominate the problem. Take a look at Big Dog videos.
Twist control involves not inducing torso rotation, or at least cancelling it on the next half-stride, and is an additional problem bipeds face more than quadrupeds. Honda had a lot of trouble with that in the early days of Asimo. Planar bipeds can't twist, so this group gets to ignore that problem.
It's easier to work in this area than it was in the 1990s. Simulators are much better. CPUs are much faster. The theory is better. Control algorithms are better. Motors are better. Funding is better. There's no need to work on an oversimplified version of the problem any more.
This is another one of those "top N, one per page, ads on every page" ad farm trolls.
Their list isn't too impressive, either. Bell Labs, yes. IBM Watson, yes. PARC, yes. But where's the Moore School of Electrical Engineering, from which came ENIAC, and the beginnings of UNIVAC, the first commercial electronic computer to go into production? Also, Bletchly Park wasn't that influential because nobody knew about it until the 1970s.
What we call a "computer" today is properly a stored-program general purpose digital computer. There were machines built before that which had some, but not all, of those attributes. Bletchley Park's machines fall into that category). The WWII US crypto operation was at Arlington Hall, which did more hardware development than Bletchley Park. were developed. They were using punched cards where Bletchley used people and filing cabinets, and they seem to have developed digital magnetic tape, although the history there is cloudy. NSA is the direct descendant of Arlington Hall.
Another major pre-computer computing company was Teleregister, which was a spinoff from Western Union in 1949. They pioneered "remote computing" for stock quotations, railroad ticketing, and airline ticketing. Their Magnetronic Reservisor was the first big remote-access system, with magnetic drums holding the reservation data.
The postal service is a push system, a pigeon hole is a pull system.
That depends on how far away your mailbox is. For me, FedEx and UPS are push systems, but the USPS mailbox is a quarter mile away.
They're on a fixed schedule, available for a limited time, and require registration. You'd think they'd just upload them to the Internet Archive and be done with it. Talks for previous years are on "video.linux.com".
"Google ['s game revenue cut] is at 5% because they don't have any [gaming] users,"
Does Google+ require that games on their system use their billing system? Facebook does, but it's not clear if Google does. With Google's Schmidt commanded to testify before the Senate Judiciary Committee on September 21 on Google's antitrust issues, I'd be surprised if Google did require that. 5% is a reasonably competitive price for a payment service. At 30%, you look into handling your own credit card payments.
It's a special-purpose mail system, much like a corporate system. Content stays around forever, and it's only for sending to and from units of the Government.
The real issue is winding down the postal system. Postal systems worldwide are sending fewer letters each year.
Space-X may be the future of space travel. They designed that thing. It's not a NASA design, and it didn't go through NASA's process of spreading everything out among contractors spread across the US.
San Francisco has a lot of crazies, and sorting out the harmless ones from the dangerous ones is hard. Here's one case. A guy in a wheelchair was slashing the tires of city vehicles with a rock. He was apparently shot by cops with a beanbag gun, and he and one cop were taken to a hospital with non-lethal injuries. There's video from someone across the street. Excessive force? Perhaps, but subduing someone with a sharp object without getting cut is very tough.
Here's another case, of a known mentally ill woman shot in her home after threatening a social worker. This probably was excessive force; the cops could have waited for backup from the guys in heavier protective gear.
Here's the aftermath of the shooting the protests are about. This doesn't seem to have been a crazy, but a parolee from Portland who ran from cops after trying to ride a bus for free and being caught at it. Whether he had a gun remains an open question.
The SFPD is at best a mediocre department. NY and LA have really tried to clean up their act, but SF hasn't.