Pretty much the most interesting blog on the Internet.— Prof. Steven Landsburg

Once you get past the title, and the subtitle, and the equations, and the foreign quotes, and the computer code, and the various hapax legomena, a solid 50% English content!—The Proprietor

Wednesday, November 11, 2015

George Washington’s Brain

George Washington's Axe

This is George Washington’s axe, the curator said with pride, gesturing toward a worn and rugged thing, locked inside of a glass case, held in the very hands of the first President of our great nation.

Of course, he said, we are much more careful about it now than we have in the past. The room we used to store it in wasn’t climate controlled, and the wood rotted, so we had to replace the handle. And one of our interns dropped it one day, and the blade cracked, so we had to replace that, too.

But, a young boy asked, if the handle has been replaced, and the blade has been replaced, it isn’t really George Washington’s axe anymore, is it?

This old joke, which entertained even the ancients, also shows an alternate route to the singularity, one on which Google this week made a small, but crucial, first step.

The standard route to the singularity—the point at which technological enhancement of the human mind starts to enable to so many further enhancements that progress deviates from the slow, exponential path we have observed so far and reaches a singularity in the mathematical sense—involves whole brain emulation. Under this process, a human brain is scanned at a sufficient granularity as to capture all mental processes and this scan is then used as the initial state of a computer simulation of the human brain.

This process has the advantage of regularity: To achieve it would require orders-of-magnitude improvement in brain scanning and processing power. But the additional required technological progress is no larger than what we have already observed. If we can merely continue on the present path of progress—likely, but not certain—then we will within a lifetime or so reach the capacity for whole brain emulation.

However, there is an alternate approach, that may reach a similar target and sooner: Piece by piece replacement of mental function with computational components of ever-increasing sophistication and autonomy. This process began long ago, no later than we ceased to memorize telephone numbers and relied on computers which are so much better suited to the task. Other functions have started to follow: map reading and navigation. Soon so will driving automobiles.

Some lament this as dumbing down. And indeed, we probably lack the mental training to do some tasks our parents, or we ourselves, could have performed years ago. But this is a cause of celebration, not concern. All of us have only limited mental capacity, only so many neurons, only so many interconnections. It is a far wiser choice to dedicate these limited, precious resources to mental tasks which cannot (yet) be replicated by mere cheap machines. This renders us no more dumb than the replacement of human labor by animals and machines in agriculture and industry has rendered us weak. To the contrary, both types of technology render us in the aggregate much stronger than we would otherwise be.

As computers, data sets, and algorithms become ever more powerful there is no absolute limit to what constitutes a lower mental task that can be delegated to a computer. Rather the current limitation is that, while optimal behavior was relatively easy to define for the lowest functions—there is only one correct cell number for any contact one wants to look up, only one fastest driving route between two points given any traffic pattern—even the lower remaining functions are ones in which people are idiosyncratic. What would be the optimal behavior of a computer delegate for one person would be regarded as rather sub-optimal by another.

The solution to this problem is big data. A computer delegate who can observe the behavior of its human principal in countless similar situations will often be able to take over the task to the satisfaction of the principal.

Consider the simple problem of an e-mail program sending a standard response to a routine e-mail. No one standard response is always appropriate. But a computer delegate which can learn both from how people generally respond to certain e-mails, how the particular principal differs from others, and what other factors (the recipient? the time of day? the status of the principal’s calendar?) affect the response, may be able to do a creditable job.

An early version of this is what Google has just offered as a public trial: Computer, respond to this email. A few early tries show promise. Google currently suggests one or more answers each of which deftly falls between the generic and specific. In every case, one of the suggested responses was appropriate, even though sometimes manual explanations at great detail had to follow.

There is no definite limit to how far this process can advance. Feeding ever more personal information into such delegates’ databases and equip them with more and more sophisticated statistical mechanism, and there is no reason that the delegate would not be able to handle ever more complex interactions with ever less human supervision. Someday, one may authorize your delegate to answer your e-mails when you sleep or are on vacation with little fear of major mishap. As one’s mental faculties decline with senescence, one may give ever larger authority and autonomy to the delegate. And, finally, when one’s mind fails entirely, the delegate may prove both one’s continuation beyond death and the first intelligence not based on a neuron substrate.