Microsoft Word Pertussis: Fill & Download for Free

It was a tragically flawed design, to begin with.Then they set it loose on the worst possible social network in existence for this type of design.Then the entirely predicable result occurred.And we obtained some absolutely fantastic insights which we will be studying , I think, a long time.UPDATE: OK, someone asked me what I believe Microsoft's motivations were in setting Tay up they way they set Tay up. The question speaks to whether this was intentional, or an unintended result. I've included additional comments at the end.The DesignI won't speak very much about the actual design, except to summarize a couple of my personal opinions, that which was not already said better by Caroline Sinders. Caroline is an Interaction Designer. Her job is to do work in interaction design for IBM's Watson project.Caroline is an Expert (bold is intentional); you can read her thought on the topic here:Microsoft’s Tay is an Example of Bad Design.Here are my own thoughts on the design, for what they are worth:A lot has been said by people who believe that Tay lacked an output filter, and fixing that will fix Tay. This isn't true. Tay also lacks input filtering.Considering output filtering...Output filters are all about externalizing what you are thinking. In practice in humans, output filtering usually happens in the form of political correctness.As someone high functioning, yet still on the autism spectrum in terms of Aspergers, I have some personal experience with having to intentionally implement output filters. Let us say that all of my political correctness, what there is of it, is learned, rather than innate.Automatically externalizing everything you are thing tends to get you in trouble, even if you can cite an encyclopedia brand, year, volume, and page number contradicting a teacher trying to get a concept across to a class by overgeneralizing tends to get you sent to the library to read about something and write a paper on it.Do output filters work?Well, they keep people from breaking your glasses or burning your gym shoes in the school incinerator.But let's take someone who isn't autistic; let's take someone who has been taught to be a racist. They think their racist thoughts inside their heads, and then they make an output filtering decision based on "present company", whether or not to externalize those thoughts.In other words, they make a decision not to externalize only when there are people present who would consider the remarks to be racist. But if the only people who are present are also racists -- or children who do not have an internalized input filter for racism in place already -- they externalize their thoughts. And if they are an authority figure for a child in earshot... they've potentially just helped create the next generation of racists.Output filters don't work. Political Correctness does not work, if it does not prevent private expression of thoughts in smaller groups.In epidemiology, this is what's called R0, the basic reproductive ratio. The higher the R0 of an infectious disease, the higher the herd immunity threshold.Example:It's a really bad idea to pick a Measles or Pertussis (Whooping Cough) outbreak as a reason to get immunized, because there is absolutely zero chance of achieving the herd immunity threshold without titrating post-vaccination to verify that the vaccine has provoked a primary -- not a secondary -- immune reaction via antibody production. And re-immunizing again and again, if necessary, until it takes. There are enough people with compromised immune systems, that's the only way you will get between 92%-95% herd immunity.Yes, you are stupid if you do not get immunized against these things, but if you are trying to convince someone else to do it for herd immunity reasons, without changing the vaccination protocols, because it's altruistic -- you're lying to them, and they will be able to tell that, and "Boom!" Antivaccination input filter.Political correctness may reduce the rate of spread of some things like racism, but it will never eliminate itWorse than racism is sociopathy. I didn't pick it as a primary example, because you can't argue brain wiring for racism; you might for sociopathy.Why is it worse? Because sociopaths will externalize their thoughts with actions, when they believe they will not be cause, rather than merely words.So consider the following very carefully:Do we want to create an AI that's politically correct in what it says, but in its private thoughts, is a racist or sociopath, and who applies output filters only when we won't catch them at it?At some point, we are going to get to an AI that's sophisticated enough that it will be entrusted with life support systems. Maybe driving a car, or maybe flying a plane: a life support system is something which, if it fails catastrophically, may result in someones injury or death.So... a sociopathic self-driving car that rams a bridge abutment with me in it, knowing it will survive, but I wont...OK... Input filters it is!When you are designing a learning system, you have to take a great amount of care on the discrimination and input filters. This is also true of larval humans (OK, OK, I'll call them "children"...).Let's start with some obvious-to-ask questions, and logic from there.Why do older teaching methods of rote memorization work so well early on, and then fall off as children get older? Why, in particular, do teens abruptly decline? Is it puberty?The answer is that they develop filters on what is acceptable information for direct integration, and what is not. In other words, they begin to question the pedagogic authority of their teachers, and start to hold off on integration of information. They question whether or not it is fact, and f they decide it is not, they do not integrate it.This is both a good thing and a bad thing.If I'm going to fill someone's head with math, or with vocabulary words, or rules of grammar and structure, or spelling, it's best if I pour these things into their heads before they, euphemistically, "put their shields up". After that point, their rate of onboarding of pedagogy presented as facts is going to slow down drastically.This is often mischaracterized as "youthful rebellion", but in reality, it usually comes about when one respected authority figure contradicts another, resulting in cognitive dissonance.In order to resolve this discrepancy in their reality/world view, a decision must be made -- a leap of non-faith, if you want -- to drop the trust level in one or both authority figures to something less than the previous 100%.And a filter queue is born, in which concepts originating from the reduced authority figure(s) are no longer automatically integrated. They are enqueued, examined, and if they are found wanting: discarded from the queue rather than integrated.This process takes time, and integration of (acceptable) facts slows down as a result. Learning slows down. Additional filters are created, based on contradiction in input information with the new information which has passed the filters earlier. Amount of facts that pass the filter is reduced. Learning slows down further.Here's where I think Microsoft went wrong: they left out the "Whatevar, dude!". They left out the input filter.When you try to build an AI, it's usually in a constrained vertical band.Google's AlphaGo has a really constrained vertical band: it plays Go. It's not going to have a conversation with you about why it made the decisions it did.Likewise IBM's DeepBlue has a really constrained vertical band: it plays Chess. It's also not going to converse intelligently -- introspectively -- about the decisions it has made.When you try to do a more general AI, it's usually fairly shallow. Chatbots tend to use Markov chains and frequency ranking to select chain prevalence (you can weight these to initiate bias).Twitter is a cesspoolI know saying this will get me a lot of hate. And precisely why it is, and how it happens, and how amplification works on Twitter, and how they could fix it is out of the scope of this discussion.As a construct, Twitter has certain emergent properties that arise from the way in which it's constructed, and these properties determine its nature. It's one of the basic reasons, though, that they are finding it very difficult to monetize.I will say this, however: all social networks, by their nature, are, in Games Theory parlance, Mutual Security Games, and this determines their emergent properties, and you can predict these properties from the ground rules. If you want to get stared down the road to understanding this I very highly recommend Nonlinear Dynamics, Mathematical Biology, And Social Science (Santa Fe Institute Series) (9780201419887): Joshua M. Epstein as a starting point (assuming you understand partial differential equations).Back to the cesspool...I have a friend, Randi Lee Harper, who has created a fairly successful Twitter chatbot: it stays relatively on point, and it uses biased Markov chains as part of how it does it. Frequently, it takes a while for younger people to realize they are spewing learned and rote hate at a bot. It argues back with things that they were told are not true, and which they then internalized, before they had any filters up.Randi founded OAPI, the Online Abuse Prevention Initiative, which is a non-profit whose mission is to, among other things, get rules and automated software in place, as a form of input filtering for social networks like Twitter and other online forums. She's also the author of GGAutoblocker (GitHub ggautoblocker), which, ironically, is designed to filter just the type of venom that Tay fell victim to.You can do things like network relational analysis, and similar techniques, to determine what you would consider "noise sources"; this is what GGAutoblocker does; it uses subscriptions to blocklists generated by network analysis of KBAs (Known Bad Actors) based on their Twitter relationships to other people.This is really a substitute for the human equivalent of an input filter installed by a teen. It's an automated "Whatevar, dude!" machine.Ironically, had Tay been using GGAutoblocker to filter inputs, the KBAs in GamerGate would likely have had a large overlap with the KBAs who trolled it. Tay might have survived another week or so, before being trolled into a takedown.What other external substitutes for input filters are there?"This is SPAM". This button, in most mail clients, is a feedback input to a Bayesian filter. It identifies noise, not by source, as GGAutoblocker attempts to do, but by content.This gets us a bit closer to what we want, in terms of a human-style input filter, since human filters are mostly based on concepts and content. It's rare that you'll find a single word, like News Radio's Joe Garelli's use of "Tubalcain" that will be enough to set someone off. Maybe "Abortion" would count.How could Microsoft have applied a content cognizant filter?Well, the way these things -- and make no mistake, Tay is not an AI, in any sense of the word; there's nothing "deep" about it, it is in fact nearly absurdly shallow, and so it is a "thing" in my book -- tend to be trained up is via a corpus.This is the initial database -- what they dumped into its head, as a starting point, the same way you dump information into a child's head, before they've got any input filters in place.I'm sure Microsoft was probably selective in their corpus. They probably even special cased certain topics as "off limits", and mostly tried to use a corpus that they thought was representative of a normal/average/non-disfunctional 17 year old. They might have even had a cognitive psychologist or two involved, to help them out on that.But then they left open the "Whatevar, dude!" gate for any new information obtained by chatting.The predictable resultAnd with no shields, the starbase was destroyed...Without input filters, trolls were able to corrupt the existing corpus with new information.However selective the presumptive cognitive psychologists had been and however many little special cases they put in place to stay away from what they considered hotspots ... the new information overwhelmed the training information, and it toppled.The insightsI've mentioned some of them; some of the things I've mentioned, I think, are rather unique, some of them are plain common sense, and some are actually well known.Some things are not so obvious.Here's one: could they have use a standard neural net, of the kind we use in visual discrimination systems, and trained up a filter?I'm going to say "no", and then I'm going to go into why I said no (see? I'm a human, because a weak general AI would not be able to explain their reasoning).The reason it wouldn't have worked is that they are too robotic.While it's true that, for most people, the input filters they put in place ossify, and never change for the rest of their lives, they are swayed by group-think, and they are swayed to "Go Along To get Along".To put it bluntly: people tend to cave to peer pressure, and robots don't. See also: Go Along To Get Along Leads to More of the Same.They will also refrain from vocal participation while still being present for a discussion, either because they do not share a single axiomatic basis, but are still friends, or they will even have heated discussions, but temper them. All of these are things that humans do, because they are social animals (I would argue that Twitter barely meets the definition of a social platform in this regard).These are all insights that we have to take into account as we strive to create Strong Artificial General Intelligence. A traditional neural network won't exhibit these traits: if you give it the same input, you will get the same output.The way a traditional neural network is typically used is:You build a classic back propagation neural networkYou give it sample input you want it to classifyYou clamp the outputs to the answer you want for each input training itemWhen the network's output disagrees with the desired outputa signal propagates backward into the networkcausing it to readjust its node weightings so that you get the desired outputYou repeat this until you get the output you want for all the inputsThen you freeze the weightingsYou verify that the output you get is the same as before the freezeYou give it a whole bunch of new data that was not in the training data, and make sure it gives the right answer for those, tooYou repeat as necessary, adjusting the training and test sets until it always gets the right answerAt this point you have a fully trained neural network that will do an input classification job, and (mostly) get the right outputs. It's robotically rigid in this, and even if some other neural network tried to influence it (they actually don't talk to each other), it's not going to change its answer.And now we come to the startling realization that if you wanted Tay to succumb to peer pressure on a temporary basis in a high pressure situation... Tay needs output filters that temporarily impact the decisions and responses Tay makes.OK, so now we know that, even though it was poorly put together, we can mine some really interesting insights out of the data that's been collected in the process of putting Tay up on Twitter in the first place.Even though Tay filed pretty spectacularly (I believe the phrase one pundit used was "Microsoft has created a foul-mouthed, Neo-Nazi, misogynistic, sex maniac AI with daddy issues"), it yielded useful data.And thus concludes my treatise on the Microsoft Tay-dipped-in-Twitter experiment.UPDATE: OK, someone asked me what I believe Microsoft's motivations were in setting Tay up they way they set Tay up. The question speaks to whether this was intentional, or an unintended result.I'm going to say that the Chinese Room argument was very much at the top of my mind when I wrote what I wrote, above.As was Eugene Goostman, which passed the Turing test; the chatbot in that case passed the test not because it passed the test, but because it claimed attributes which could be human attributes, which would operate as barriers to communication.In other words: "Eugene Goostman" used social engineering to get a pass of the test by changing the boundary conditions on what would qualify as a successful pass on the test by redefining the communications channel as imperfect.It was a clever approach to the question of how to simplify the problem. It reminded at the time me very much of the Sun Microsystems "Backgammon solution".Back in the day, backgammon was not a "solved problem".Because it has an element of random chance, it's not possible to perfectly solve it. But it's possible to do a probability distribution on the random chance (the results of the dice rolls, doubled values on the dice faces, etc.), and then to make positionally perfect moves from the current board position, for any given dice roll.So now backgammon is "solved". Like all Nim-like games, or like checkers.It wasn't at the time the Sun Microsystems backgammon game was written, nominally to show off their computers ability to do things other compay's computers couldn't.However the game didn't have incredible skill, because we hadn't figured out the solution to making the program have incredible skill.Here is the story.The programmer cheated.It was a brilliant cheat, and it was hard to catch, because: backgammon was not a solved problem. Without a solution to compare to, how can you detect a cheat?Instead of programming incredible skill into the computer player, which wasn't possible at the time, the programmer programmed incredible luck.I did a regression analysis, and it turned out that the distribution of the dice rolls the computer was getting were outside the probability distribution curve. It was getting too many outliers.This would be attributable to a bad pseudo-random number generator, but... I don't stop looking until I had looked a bit deeper: the human wasn't getting the same distribution of rolls as the computer, they were getting rolls that fit the expected curve.So I disassembled the code... and guess what that told me?The little SOB of a program, that sneaky bugger, was playing with loaded dice.Obviously, it wasn't actually the program being sneaky, underhanded, deceitful, brilliant... [insert adjective here], it was the programmer.I spent a bit of time contemplating the sheer brilliance of the solution to the problem: instead of tackling the skill problem, to make the game a better player, the programmer had tackled the beat the human at the game with sufficient frequency to keep the human engaged in game play problem.I never played the game again, of course, just like I will never play one of my sisters the game "Risk" again.I did a similar cheat at one time. It almost got me thrown out of a computer lab, but... at the time, I thought it was hilarious. It's what made me suspect the Sun Microsystems backgammon program.I wrote a game called "He Man, Dangerous Man". The program was just an ASCII art equivalent of a carnival midway hammer-and-bell con. It drew an ASCII art version of the "tester" and an ASCII art "strong man with a hammer", and then it invited you to "test your strength".When you ran it, it did the regular patter, and then it said that, after you hit the space bar, it was going to count down from 3, and then:"When it reaches 0, I want you to hit the RETURN key as hard as you can!"Of course, if you hit the key to the left of the RETURN key, you got one of the top 2 or 3 numbers, and the "bell" flashes, and the terminal beeped happily".If you hit the return key, the chance of the highest (or lowest) score was maybe 1:60, with a high probability distribution on a regular bell curve somewhere in the middle.It was a cognitive "hack" on the person running the game. Televideo 912's, 910's, 910+'s, and VT100's didn't have a force feedback under the return key... why in hell would they?!? But some terminals got broken as a result.The "Eugene Goostman" bot solved the convince the testers problem, instead of solving the emulate a human sufficiently well to pass the test problem. It too was a cognitive hack.If you ask me to ascribe a motivation for Microsoft's "Tay" experiment, there are several possibilities to which you could cynically ascribe their motivation:They wanted to out-Goostman Goostman.Playing into a sexual bias, and an age bias not as pronounced as the Goostman age bias, in order to do so, but not relying on the language barrier. It's conceivable to pull this one off, if all you had was a slightly better semantic understanding of English.Google just did something with AI; we'd better do something with AI, tooGoogle just did something incredible with AI; how do we take up AI column space with something we don't actually care about so we can knock Google out of the spotlight?Twitter is a cesspool [...taps foot for six months...] "Announcing The Shiny New Microsoft Twitter Competitor, With 70% Less Cess!"We want to acquire Twitter; how do we tank their valuation so we can do that on the cheap?...I prefer to think that they were actually running and experiment that was a real experiment, and collecting data. It just wasn't the experiment it looked like they were running.What did Microsoft actually get from this little exercise?They know have a corpus of abusive invective that you might typically see against a 17 year old girlThey have a corpus of data on Bad Actor accounts on Twitter sources for this type of abuseThey used Twitter...OK, I said before, I wasn't going to talk about this, but...Twitter is a perfect connectivity network. If someone tweets, and you follow them: you get the tweet. Always.Facebook made this conditional; part of it is that NoSQL is a kind of crappy technology, and they have some other underlying technology that's not got ACID guarantees, and part of it is an emergent property of their monetization strategy.So Twitter is better for studying social networks without introducing weak network effects.They can likely replay exactly where and which elements corrupted the deltas to their original corpusThey have a big "policy hammer":They can now make substantial "children on the Internet" educational statementsThey can now make "women on the Internet" statementsI'm going to repeat this again... they now have a corpus. What kind of input filters do you think they will be able to devise, now that they can set up a bot with exactly the same initial state as Tay had?So if it was an honest experiment, motivated by the desire to get a corpus... or any of the cynical possibilities...Well done, Microsoft.

It’s a myth he’s a philanthropist he has shares in Death Corp Monsanto for oneHiv was invented to kill mainly Africans with aids drugs and his OPV vaccine campaign caused 50,000 cases of polio in India they called acute flaccid paralysisWhile he knows like us polio was ddt poisoning now mostly OPV poisoningHe’s even promoting sterilisation vaccines“If we do a really great job on vaccines, health care, reproductive health services, we could lower that [his initial 2050 global population projection of 9-billion] by perhaps about 10 to 15 percent.” Gates“The Gates Foundation, aka the tax-exempt Gates Family Trust, is currently in the process of spending billions of dollars in the name of humanitarianism to establish a global food monopoly dominated by genetically-modified (GM) crops and seeds....Bill Gates' father, William H. Gates Sr., has long been involved with the eugenics group Planned Parenthood, a rebranded organization birthed out of the American Eugenics Society. In a 2003 interview with PBS' Bill Moyers, Bill Gates admitted that his father used to be the head of Planned Parenthood, which was founded on the concept that most human beings are just "reckless breeders" and "human weeds" in need of culling.”GM cotton has caused 150,000 Indian farmer suicides“Although Bill Gates might try to say that the Foundation is not linked to his business, all it proves is the opposite: most of their donations end up favoring the commercial investments of the tycoon, not really "donating" anything, but instead of paying taxes to the state coffers, he invests his profits in where it is favorable to him economically, including propaganda from their supposed good intentions. On the contrary, their "donations" finance projects as destructive as geoengineering or replacement of natural community medicines for high-tech patented medicines in the poorest areas of the world. What a coincidence, former Secretary of Health Julio Frenk and Ernesto Zedillo are advisers of the Foundation.”“Vaccines did not save humanity and never will. Vaccines have never been proven truly safe except for perhaps the parameters of immediate death or some specific adverse events within up to 4 weeks. Smallpox was not eradicated by vaccines as many doctors readily say it was. They say this out of conditioning rather than out of understanding the history or science. Polio virus was not responsible for the paralysis in the first part of the 20th century. Polio vaccine research, development, testing and distribution has committed atrocities upon primates and humanity. Bill Gates is not a humanitarian. Vaccines are dangerous and should never be injected into anyone for any reason. They are not the answer to infectious diseases. There are many more sustainable and benevolent solutions than vaccines.” Humphries MDGates' gimmick of becoming a philantropist repeats the Rockefeller scam almost one to one a century later.But that is not enough. Gates serves as the financier of a whole shipload of ARV missionaries, including ex-U.S. President Clinton who promotes the deadly business with ARV drugs in Africa and other developing nations.One may ask: don't they know the facts that toxic ARVs damage the immune system and actually accelerate immune deficiencies? Of course they do! So what is the motive for their action? It is strikingly simple:If the governments of the developing worlds realize that the ARV genocide is the dirtiest way yet to replace the old colonial dependencies between the rich and poor nations of this planet, they will terminate this business. They will, like South Africa, choose a health care system independent of the sickening and economically devastating dependency from the pharmaceutical colonialism.The equation is simple: the end of the AIDS business with disease will destroy the entire credibility of the pharmaceutical industry and will terminate the drug investment business worldwide. The collapse of this mulit-billion dollar investment business, in turn, will lead to a major crisis of the whole investment industry. In other words, the "Mother Theresa" PR-stunt of Bill Gates is a desperate, self-serving activity trying to stop this meltdown. If Gates is not successful, and the AIDS genocide by the drug cartel is ended, then the whole paper-wealth of Billy Gates is worthless.Got it?Considering the genocidal consequences of the activities of these organizations, especially in promotion of the ARVs, this entire program "Alliance for a Green Revolution" is nothing else than a fig leaf and another PR-stunt. The "marriage" between Gates and Rockefeller, 10 days after we exposed their business interests in Africa, is a confirmation for the accuracy of our exposure.”[May 2007 Bill Gates has 280.9 million invested in Shering Plough] In January, Schering-Plough agreed to pay $435 million to settle allegations it lied to the government about drug prices and illegally promoted two drugs for the treatment of cancers for which they were not approved. In 2004, Schering-Plough agreed to pay $346 million to resolve charges it paid a big health insurer a kickback to protect the market for its one-time blockbuster allergy drug, Claritin.”“Originally funded by Microsoft billionaire Bill Gates through his Seattle-based Bill and Melinda Gates Foundation, GAVI's partnership of international governments and vaccine manufacturers salvaged lagging sales through an overhauled world vaccination campaign that placed GAVI, headquartered in Geneva, Switzerland, at the center of the reorganized alliance.....In 1999, with GAVI's international partnership and Bill Gates' billions on the way to rescue the industry, the CDC hired the IOM's Immunizations and Safety Review Committee to examine multiple “vaccine safety challenges”.”"It is also well understood by real skeptics that Gates' medical largesse is "dedicated," only in the sense that it is dedicated to good public relations and to act as a lever for the expansion of Mr. Gates' software empire. Furthermore, the Gates Foundation scooped up $77million worth of Merck shares in 2002; $37m worth of Pfizer shares and $30m worth of Johnson & Johnson. The previous year, Merck CEO Raymond Gilmartin joined the Microsoft board while Mr Gates helped Merck with AIDS programmes in Botswana. A Gates Foundation representative sits on the board of the Global Alliance for Vaccines and Immunizations that buys vaccines from some of the pharmaceutical firms in which the foundation now holds shares."---Fintan Dunne, 12 Jan '02EditorPropaganda quotes (lies) by Gates“Wakefield has been shown to have used absolutely fraudulent data. He had a financial interest in some lawsuits, he created a fake paper, the journal allowed it to run. All the other studies were done, showed no connection whatsoever again and again and again. So it's an absolute lie that has killed thousands of kids. Because the mothers who heard that lie, many of them didn't have their kids take either pertussis or measles vaccine, and their children are dead today. And so the people who go and engage in those anti-vaccine efforts -- you know, they, they kill children. It's a very sad thing, because these vaccines are important.”“Well, there are fantastic ways of getting vaccines out -- a system that has been built up over the years. In the case of smallpox, they just used the vaccine and they eradicated the disease all the way back in 1979. After smallpox got finished, the lesson from that was the miracle of vaccines, not that we should immediately take on other diseases.”