Visit Your Local PBS Station PBS Home PBS Home Programs A-Z TV Schedules Watch Video Donate Shop PBS Search PBS
I, Cringely - The Survival of the Nerdiest with Robert X. Cringely
Search I,Cringely:

The Pulpit
The Pulpit

<< [ If You Build It, They Will Come ]   |  Put On Your Thinking Cap  |   [ I'll Gladly Pay You Tuesday ] >>

Weekly Column

Put On Your Thinking Cap: Chris McKinstry Wants to Build a Brain Accelerator

Status: [CLOSED]
By Robert X. Cringely
bob@cringely.com

We are right now at what some people see as the end and others see as the beginning of a scientific, cultural and business explosion called the Human Genome Project. Scientists around the world are finishing a map that will fully describe the millions of chromosomes that define a man and make him different from a hamster. It is a mammoth undertaking that, at its heart, is really more about computers than about biology. That's because a project of this immensity couldn't be accomplished without automation. Computers make it possible. No wonder the companies engaged in this work feel like Silicon Valley startups.

Another similarity these outfits share with their Silicon Valley brethren is a fixation on wealth. These folks are determined to get rich from their good work. Stock options all 'round! And there will be plenty of wealth to share, because what the Genome Project is doing is literally defining how life functions. The therapeutic fallout of this is profound, scary even. Once we know our genetic makeup, it isn't very far to the end of manipulating those genes to cure genetic diseases. Cancer, heart disease, multiple sclerosis, thousands of ailments promise to be reduced dramatically, if not eliminated. Then there are conditions that are genetic, but have generally not been thought of as diseases. Ready for a cure for aging? It is possible. Eliminate enough genetic conditions and we'll be stuck solely with violent deaths. Life expectancy will increase and suicide rates will soar.

It is going to be a very interesting time, not just because of the medical changes that genetic engineering will enable, but because of the inequitable distribution of those changes. Who can afford to live forever? The rich are going to get healthier and the poor, which means probably 99 percent of the world population, aren't going to change much at all. In fact, they might get just a bit angry. Only time will tell.

But this isn't normally a column about biotechnology, nor am I turning it into one. Today's topic is still computers, but I want to frame it in terms of genetics to gain the most impact. This is because I've found a guy who has a plan to define machine intelligence in much the same way that the Human Genome Project is defining our human machine. And this guy wants you and me to share in the resulting wealth.

Chris McKinstry is a Canadian who operates the world's largest telescope, the VLT, for the European Southern Observatory at Cerro Paranal, Chile. He is also a computer guy with impressive credentials having made his high tech fortune and retired by the age of 30. Chris' new project, this defining of machine intelligence that is supposed to make us all rich, came from his dissatisfaction with the Turing Test.

Alan Turing was a British mathematician who pretty much defined how a modern computer would operate before anyone knew what one would do with such a machine if it ever existed. Back then, people played solitaire with quaint paper cards. Today's computers, right down to the one I'm writing on and the one you are reading on, are Turing Machines.

The Turing Test first appeared in 1950, and was Turing's best shot at how to determine if a machine could actually think. Turing wasn't sure how to define computer thought, but was pretty sure he could measure it through his test. The Turing test involves one person at a computer screen communicating with another person and with a computer. The trick here is that the person conducting the test doesn't know which of its correspondents is the person and which is the computer. The test is passed — and the computer is deemed to be capable of thought — if the tester can't determine through questions and interactions which is the person and which is the computer.

Robert French, a cognitive scientists and prolific writer who teaches at the University of Liege in Belgium, had a problem with the Turing Test. French thought the Test didn't so much test a computer's thought capability as its grasp of human experience. A computer might be designed in such a way that it was perfectly capable of thought, yet fail the Turing Test simply because it didn't share the same culture as the tester. In French's words, "[O]nly a computer that had acquired adult human intelligence by experiencing the world as we have could pass the Turing Test."

Bingo! Fireworks went off in the mind of Chris McKinstry. Who cares if a computer is capable of thought? It is much more useful to have a computer that APPEARS capable of thought, which is to say a computer that experiences the world as we do. But how do you program a computer with the entire human experience? How do you tell a computer that a little salt is good but too much salt is too much? How do you teach a computer that "Kiss my butt" doesn't literally mean kiss my butt? And how, even if you could teach a computer these ideas, do you teach it all the ideas in between. That's when Chris McKinstry invented the mindpixel, and the Mindpixel Digital Mind Modeling Project.

"All great truths are composed of a multitude of minor truths, and the minor truths are composed of massive numbers of atomic truths," says Chris. The basic facts he hopes to elicit from you and me are the atomic truths, items of binary consensus fact or mindpixels.

A mindpixel is a binary stimulus-response pair in the form of a statement that says something about our experience of the world. Here are three mindpixels:

Water is wet — true.
It is difficult to swim wearing ski pants — true.
"Flugly" is a good surname for a glamorous female movie star — false.

Now a machine that used those three mindpixels as the basis for its view of the world would have a skewed and not very useful view to say the least. But what if the computer was programmed not just with those three mindpixels but with hundreds of millions of mindpixels, each unique. Eventually, on the basis of such programming, a view of the world would emerge that's not very much different from our own. And that's the point of McKinstry's new project, to obtain and validate over the next decade one billion mindpixels that can be used together to describe the human experience to a non-human machine.

But why? This is what Chris, who let's just allow is perhaps 20 times smarter than I am, tried to explain to me this week. And the answer is simple - to help us think. Chris wants to build thinking machines to do some of our thinking for us.

It's based on tomography, rather like a CAT scan or an MRI. Both imaging technologies collect hundreds of millions of samples of X-rays or magnetic fields passing through our bodies then apply statistical correlation techniques — not to find some unique solution — but just to paint a picture of what the rays and fields just went through, our bodies.

"That's the way we think," said Chris, who apparently covers all this and more in his upcoming book, Hacking Consciousnes. "Think of the human mind as a hyperdimensional object and human thought as hyperdimensional tomography. We have in our minds all these binary pairs defining our own experience and we make our way through the world by building in our minds a statistical model of that world. People don't like to think that the brain is binary and that we are all just Turing machines, but we are.

Basic facts, such as the sky is blue, grass is green or one is less than two, are validated through repeated experience or confirmation from our parents, teachers and peers. Over time, these facts are consolidated by the brain to form each person's consciousness. While consciousness varies from person to person, similarities exist among individuals sharing similar life experiences.

Back to those mindpixels. What do we do with them once they've been gathered? The first 900 million mindpixels will be used to train a neural net, a kind of computer program that in this case extracts linguistic data, performs a principal component analysis and eventually yields a self-programmed hierarchical network that somehow — nobody needs to understand how — creates the world described by all those pixels. At this point, Chris says he'll test the neural net to see if it can pass the Turing test by requiring answers that point to mindpixels among the 100 million the computer has yet to learn.

While other projects such as Statistical Natural Language Processing (SNLP) are concerned with building statistical models of human language, they do so by statistically analyzing large bodies of text, which are plentiful on the Net, but the statistics have no connection to "meaning." The Mindpixel Project is the first to connect the statistics of language to the statistics of meaning. Meaning is ultimately a pattern of connections between a given concept and all other concepts. By having a large group of Internet participants validate consensus facts submitted by others, the project will establish universal meaning for each question submitted.

It's complex, I know, but if the neural net can actually generate from the 900 million mindpixels used to program it some of those 100 million unlearned mindpixels, it will represent the world's first artificial thought.

But the neural net still won't know what to think about. So the next step is to use a random sentence generator to feed millions of new word sequences into the neural net. Nearly all of these sentences will make no sense at all, but a few will make sense, and they are the key to taking the neural net to the next level of thought. You see, at this point, the neural net will have a pretty good description of the world. It will understand what the rest of us understand and will be able to make a binary response, either agreeing or disagreeing with a statement. The magic starts when, responding to one of those random sentences, the neural net comes to a conclusion it not only believes specifically to be true or false, but which it also knows is outside the range of experience defined by the mindpixels with which it was programmed. This will be a machine thinking thoughts that have never been thought before, which is pretty much the whole purpose of this effort.

"We'll use that first sentence to prime the random sentence generator to probe further that particular area of thought," said Chris. "Ideally, this should lead to a whole train of thought. That's when we'll have our artificial thinker."

The artificial thinker is the end product of all those mindpixels and of the company of the same name — Mindpixel. It will be a program that does nothing on its own, but when you tell it to think about something in particular — say rocks or ice cream recipes — it will just go off and think until you tell it to stop. That's a lot of thinking, nearly all of it done while the person who makes the thinking request is catching up on classic episodes of "Gilligan's Island." The thinker will think and think about whatever you ask it to, in each case comparing its thoughts to the world previously defined by all those mindpixels and by its subsequent thoughts. When one of those thoughts doesn't fit the mindpixel world yet seems to be true, it is an original thought that extends the total corpus of human thought. And because it's your thinker, that thought is your property.

Tell your thinker to find some new thoughts about genetic engineering or astrophysics, then go on vacation. When you get back, the thinking will be done.

And that's it. We'll all want our own thinkers because none of us is smart enough. That's a few hundred million customers. Sure looks like a business to me. Chris's plan is to give 70 percent of Mindpixel's equity to the people who contribute the mindpixels, themselves (20 shares per mindpixel). He advises the best way to come up with unique mindpixels is to specialize. "Describe how the bottoms of your feet feel when you come out of the shower."

All of this takes place at a Web site which is, itself, supported by the comings and goings of all those mindpixel contributors who will be exposed to banner ads as they try to come up with unique contributions. In 2018, when the neural net turns 18 and reaches majority, Chris says he will hand over the rest of the Mindpixel equity, presumably to the neural net, itself. By that time, the neural net will have grown exponentially and may even have chosen a better name. Chris calls it GAC, but I like HAL.

Comments from the Tribe

Status: [CLOSED] read all comments (0)