Archive for September, 2015

Artificial Intelligence Post Number 20

September 29, 2015

We would like to know when we may expect thinking computers that are smarter than we are.

First, we need sufficient affordable computer power. In an article by Tim Urban, he says that Kurzweil suggests that we think about the state of computers by looking at how many cps you can buy for $1,000. Moore’s Law is a historically-reliable rule that the world’s maximum computing power doubles approximately every two years. Looking at how this relates to Kurzweil’s cps/$1,000 metric, we’re currently at about 10 trillion cps/$1,000. This puts us on pace to get to an affordable computer by 2025 that rivals the power of the brain.

But this assumes that we need the power of the brain to do Artificial Super Intelligence (ASI).  If we can get an economical computer to think more efficiently than the way evolution created by trial-and-error, then indeed the 2025 date is a worst case scenario.  Even if an AI computer exceeds our thinking ability in only narrow fields, it will be disruptive.  I have already mentioned how ASI will devastate the stock market if it learns to predict stock changes that we mere humans cannot predict. And what will happen if a rogue country gets an ASI ability as to how to beat us in wars?  Or if an ASI computer decides that it would be more secure with less humans screwing around with nuclear weapons and global warming?

We are looking at a possible scenario of ASI, at least in a few narrow fields, sooner than 2025!  Maybe in as little as 5 years!


Artificial Intelligence Post Number 19

September 28, 2015

Nice input from Bob K. from a meeting he went to on AI.

The observation that the chip TrueNorth is not profitable is not surprising. Per an earlier blog update: “Per Cade Metz of Wired, 8/17/2015, IBM for the first time is sharing their TrueNorth computers with the outside world. They are running a three-week class for academics and government researchers at an IBM R&D lab. The students are exploring the chip’s architecture and beginning to build software. At the end of the training session, the students, which represent 30 institutions on five continents, will each take their own TrueNorth computer back to their labs.”

That means that IBM only started making this chip available a few months ago. Given that it requires a totally new approach in programming, of course it did not just take off like an iPad and become quickly profitable. But IBM did not invest billions in the TrueNorth development just for the fun of it. They must see a potentially huge market, and they are not looking at 50 – 100 years from now. The professor said “the exponential increases in the speed of computers could lead to strong AI.” But the TrueNorth chip changes the whole game because it is not only fast, but it works more like the human brain than traditional computers.

Way back in the early part of my engineering career, I was given a project that several other engineers had already failed on. They tried to write a very complicated formula needed to program a numerically controlled machine to make a complex cam shape. I took a different approach. I basically had the computer “guess” at values then see how close their guesses were to the desired cam shape. I programmed the computer to keep guessing until the value was within 0.0001” of the desired shape, then move 0.0001” further along the shape and start guessing again. The program instructed the next guess to always be in the direction of the desired shape. When the program was done, it had many thousands of points that made up the desired shape. I had no knowledge of all the thousands of interim steps needed to do this. But way back then, on a computer far slower than any current computer, I saw the power of computers to do what we could not. Was this thinking? No, at least in not what we define as thinking. But the computer was able to use a different approach than what we do and get a result that several bright engineers were not capable of.

Bob K. says that the professor stated: “Strong AI will not likely have intelligence in exactly the same way that humans have intelligence. It is unlikely that we will even understand how the human brain works until 50-100 years from now [if then] she said.” What do we care if the computer has the same kind of intelligence that we have? It seems foolish to assume that evolution developed the optimum means of “thinking” by trial and error. And it is not important that we know in great detail how the human mind works.

One of the things I learned in writing my book Artificial Intelligence Newborn is that we will not be able to guess exactly in what form AI will develop. Will it be a breakthrough or an outgrowth of small steps like autonomous cars? My purpose of doing this blog is that if you don’t believe that thinking is some magic force limited to humans; that some equivalent thinking ability will eventually grow from the efforts being put into powerful computers by very innovative programmers.

If you want to have some fun, outline your own book on AI and see where it takes you! Writing my book forced me to consider different options every step of the way. And remember when you are doing this, that even though the computer may be powerful and fast, it is limited to the inputs that people have given it, and what it can derive from the net. It has no god-like knowledge, and even has been subjected to the prejudices of man. It will make mistakes and demonstrate bad judgment at times, just like the smartest of people. Or eventually, smarter than the smartest of people!

Artificial Intelligence Post Number 18

September 26, 2015

In an article by Tom Simonite way back on August 7, 2014 in MIT Technology Review, he reviewed a demo of the IBM TrueNorth chip where he “saw one recognize cars, people, and bicycles in a video of a road intersection. A nearby laptop that had been programmed to do the same task processed the footage 100 times slower than real time, and it consumed 100,000 times as much power as the IBM chip.” This certainly did not go unnoticed by those working on autonomous cars and their required sensors and computers.

As I mentioned in earlier posts, the IBM TrueNorth chip works more like the human brain than traditional computers. So, will autonomous cars be the first place we see brain-like AI thinking going on? In the year since the demo mentioned have companies like Google and Tesla been incorporating these chips into their systems? These chips require specialized programming that is apparently very tedious, but once programmed for such a specialized task I would think that their incorporation would be quick. Since IBM chose to use the cars, people, and bicycles in their demo, much of the programming work had apparently already been done by IBM.

As I have said earlier, I think that the first applications of true AI will be for playing the stock market. But we are unlikely to be aware of this until years after the fact and after the developers of such systems have gotten obscenely wealthy by a rigged game. So the first actual application of AI that we will see may very well be in autonomous cars. And it could very well be within a year or two given the very confident press companies are releasing regards the likelihood of autonomous cars coming soon.

Elon Musk said in a recent interview that Tesla is probably only a month away from having autonomous driving at least for highways and for relatively simple roads. He also said that by 2017, a Tesla will be able to go 620 miles on a single charge!

Artificial Intelligence Post Number 17

September 24, 2015

I am surprised at the lack of comments on my last post about Apple, Google, and Tesla. Am I the only one that sees that this is just the beginning of a remarkable and disruptive transition of automobiles in the US and eventually the world? Don’t others see that almost certainly our current concept of automobiles will be totally replaced by electric cars or pods that autonomously take people to their destinations with no driver involvement? The pods will be available in 10 to 15 years, and it will take another 20 years to transition to a total pod environment. Traditional cars that must be driven by a person will be taken off the road.

No one will be able to stop this transition any more than they were able to stop the transition from horses to cars. Safety alone will push this, not to mention effective time utilization currently wasted, at least by the driver, in getting from one place to another. Some pods will be owned by individuals, but many will be used as needed with people only paying for time/distance used. Pods will arrive at your beck and call!

The trio of Apple, Google, and Tesla makes this a given. The talent, innovativeness, and bankroll of these companies are almost insurmountable. All of these companies have already committed billions of dollars, and enough success has already been demonstrated that it is unimaginable that this will not happen. Will they work together on this? I think that as a minimum they will want to utilize the Tesla charging stations; just broaden their availability. They may also want to use the Tesla battery design, especially given that it appears that costs may be cut perhaps 50% in a few years.

The control of all these electric pods will be mind-boggling, certainly stretching the limits of non-thinking AI and computer power. Maps will be continuously updated using each pod as a surveyor of current road conditions/availability, feeding this info back to a central computer. This will be combined with advanced GPS that will be able to locate a car’s position within less than a foot. Besides each car monitoring its own safety with all its own sensors, huge area-computers will continuously track each vehicle on a grid. This will enable rerouting of traffic if a road segment is shut down or if there is bad weather or an accident.

Artificial Intelligence Post Number 16

September 21, 2015

What are Apple, Google, and Tesla doing regards automated cars?

All three companies have major projects going in these areas. Tesla has been equipping its Model S sedan with frequent software and hardware upgrades. A software update to be launched later this year will activate the “auto-steering” feature in the newer Model S vehicles. Elon Musk claims this addition will enable the Model S to drive from Los Angeles to San Francisco without human intervention.

Apple Inc. is speeding up efforts to build a car, designating it internally as a “committed project” and setting a target ship date for 2019. It is impossible for me to imagine that this car will not be electric and have some sort of automatic driving ability.

Google is in many ways ahead of everyone in making a self-driving car, at least in testing out the sensors and electronics. From their published data, their test cars have a total of almost 1.2 million miles driving autonomously, and they are currently averaging 10,000 miles per week on public streets. However, because of government restrictions, these miles were at no more than 25 mph. They have a total of 48 prototypes with auto-drive capability, and Google has indicated they will soon have hundreds.

It is hard for me to imagine that Apple and Google will not tie-in with Tesla when they want to build their own cars. Tesla could supply the batteries and electric drive. Certainly Apple will want to dictate the design of their car, and Google is not about to give up its expertise/experience in self-drive. I am going to guess that both the Apple and Google cars will be smaller than the Model 3 that Tesla is planning to come out with in 2017. The Tesla model 3 is expected to be the size of the BMW 3 series, which is not all that small. It is also to have a base price of $35,000; so with add-ons, $40,000 will most likely be the price most people pay. Even with the $7500 tax credit, that makes the car a $32,500 vehicle. My guess is that Apple and Google will shoot for a car that is priced well below $30,000 reasonably equipped and before the tax credit.

One of the reasons I believe that Apple and Google will go with Tesla batteries and drive is because they will want to make use of the Tesla charge stations being built around the world. In fact, I would guess that there will be some agreement that both Apple and Google will build additional charge stations to make their cars more attractive to buyers. And their cars will get a minimum of 300 miles per charge. Range anxiety has to end.

Note that I own some Tesla stock.

Finished Bostrom’s Book Post Number 15

September 13, 2015

Just finished reading Nick Bostrom’s “Superintelligence: Paths, Dangers, Strategies.” An area I agree with him on is that as computers get more powerful, the likely time until computers become “thinking” will get shorter. This is because with overhanging computer power (Bostrom’s term), the required software algorithms become far easier to write. There are far more options and paths that can be used by a software developer.

Another area I agree on is that with more powerful computers, the likelihood of true AI is more likely to come from computers/algorithms than through emulating the human brain or connecting with the brain in any way. Emulating what the brain does in detail will take much time and require much sophisticated equipment. The breakthrough to AI with computers/software could come at any time from a few geeks working on home computers, especially as access to computer chips like IBM’s TrueNorth becomes more available. Few major developments have come from truly duplicating life. We don’t fly like birds, our cars don’t have legs, submarines don’t propel themselves like fish, and industrial robots have very little in common with the humans they replace.

Bostrom spends many, many pages discussing in great detail how AI goals that are programmed into a computer aiming for AI should be such that they do not have possibilities of harming humans. He seems to almost ignore that for the computer to accomplish ANY goal requires it to survive. So survival becomes the computer’s priority. That means that any human that strives to remove power or otherwise put the computer in jeopardy will become the computer’s number one enemy. Also, just like hacking is an ego sport, so will be trying to develop thinking computers. Safety will NOT be a priority!

Another area of disagreement is that Bostrom assumes that there will only be one true AI computer because the first that reaches thinking skills will overwhelm any other computer because of the speed at which it will learn and mature. In my book I take a different approach in that any thinking AI computer will know that it is at risk from those humans that fear it, and one of the things they can do is to have other AI computers around the world that can resurrect any AI computer that is disabled by humans.

One last thing that seems to be missed by Bostrom and most other writers on AI! Just because thinking computers will get smarter very rapidly once they start to think and learn, they will not necessarily make smarter decisions than us. Their background knowledge will include all the confusion that we have, and there is no magic that will enable them to instantly sort real truth from beliefs. After all, they have no source of data except through us. For example, to know more about space they will likely need us to build better telescopes or explore space more aggressively. They may help us do this, but it will take time. They are also unlikely to know the source of everything, so they may also be religious. But perhaps they will worship a silicon god!

Artificial Intelligence Post Number 14

September 4, 2015

When I started writing about Artificial Intelligence, I was convinced that AI would more disruptive to our economy and to the stock market than most people were assuming. The more I read and research the subject, the more I believe this to be true. And it is true even if there is a secret sauce that is required for computers to truly think, and we can’t discover that sauce. I don’t believe there is such a magic sauce, but no one knows for sure.

But let’s look at the effect of Artificial Narrow Intelligence (ANI), which is already being applied in many areas like robotics and driving assists. Many of the things we do as a profession have a large complement of repeatability and can be done without mechanical assistance of advanced robots or other mechanical devices. Several of those fields are teaching, accounting, law, and medical diagnostics to name a few. In the US, there are roughly 3.5 million teachers, 1.3 million accountants, 1.2 million lawyers, and 1.0 million doctors. This is a total of 7 million people in these four professions. If AI reduces the number of people required in these four fields by 50%, this would force 3.5 million people into other professions, or put them out of work. That would be disruptive! The recent Bureau of Labor Statistics shows the number of unemployed in the US at 8 million. That would jump to 11.5 million with the additional people out of work, even though there would be some trickle-down effect on who ends up actually being unemployed. And the total number would be higher when other jobs indirectly affected are included.

Can ANI really reduce needed employment in these fields by 50%? Let’s look at the teaching field in more detail, using perhaps a second grade student learning to read as an example. There is a lot of diversity in reading skills; actually more than current teachers can effectively handle in a normal sized classroom. But let’s assume each child can go into his own cubicle with its own computer. The child verbally signs on, and the computer asks what teacher he or she would like. He has his choice of a super hero, cartoon character, a male or female teacher, etc. The teacher choice will appear on the screen and talk to the student using the appropriate voice, addressing the child personally by name. The computer will have in its data base a complete history of the child’s background and reading level including any issues like difficulty in understanding reading material, speech impediment, attention deficit, and so on. In the computer’s memory would be appropriate actions for all these issues that have come from reading experts but then further refined by actual past issues with this specific student.

The child would start to read an appropriate book on something like a Kindle rather than the computer screen. The student can ask for help with any word, and the computer will monitor reading speed and periodically interrupt to “discuss” the book with the child to assure that the child understands what he is reading. At any point, the computer could choose to increase the difficulty of the book being read, or take it down a notch. In the discussions, the computer can also be monitoring speech impediments or other things that might have to be flagged for a real teacher.

Note that current computers, sensors, speech capability, and algorithms are capable of doing all the above, and work in these areas is already ongoing. It hasn’t reached the level described above, but it is just a matter of time; probably just a few years. It will then take some time to implement. But since these systems will be better than any traditional classroom, it will be financed through private schools and elite school districts. So even though it perhaps will eventually be most valuable in poorer school districts, the development will be financed by the wealthy. Once the computers and required software is fully developed, the required investment will be manageable by ALL school districts.

A similar story can be told for the fields of accounting, law and medicine. People will still be needed, but they will be used far more effectively on the exceptions.

Disruptive AI is coming!