Fri, 20 July 2018
My guest today is Roman Yampolskiy, computer scientist and AI safety researcher. He is the author of multiple books, including Artificial Superintelligence: A Futuristic Approach. He is also the editor of the forthcoming volume Artificial Intelligence Safety and Security, featuring contributions from many leading AI safety researchers.
We discuss the nature of AI risk, the state of the current research on the topic, and some of the more and less promising lines of research.
Tue, 14 June 2016
When I think of emulation, I think of retro gaming. My Android phone can easily emulate a Super Nintendo, a gaming console from the 1990s, and it can do that because the phone is much more powerful than the Super Nintendo and because we know exactly how a Super Nintendo works. My guest for this episode, Robin Hanson, argues that we may one day be able to emulate human brains. His book, The Age of Em, provides a detailed analysis of what a society made largely of emulated humans would be like.
Whole brain emulation is unlike my emulated Super Nintendo in many ways. With the brain, we're trying to emulate something that we couldn’t build ourselves. The challenge is in developing a sufficiently accurate model of each part of the brain that is necessary for it to function. If we knew how each node in the brain worked, if we could model it such that our node would take the same inputs, produce the same change in its internal state, and send the same outputs as biological brain cells, then all that would remain would be to find the precise network of cells in a biological brain. This could be achieved by scanning an actual human brain. The brain could then be emulated by a sufficiently powerful computer. The emulated brain would have precisely the same memories and thought processes as the person who was scanned. Hanson calls these emulated individuals "ems."
Hanson applies standard theoretical tools to the analysis of this em economy. Here are some of the implications:
1. Ems will be able to operate much faster or much slower than normal human brains.
The cost of running an emulation faster or slower is roughly linear in the speed. That means that for ems working on time-sensitive tasks, a race to develop some new technology first for example, they will likely work many times faster than biological humans, perhaps experiencing weeks or months in the blink of an eye. Ems that work alongside biological humans, for instance those engaged in services, would likely run at the same speed as we do. Ems could also run at slower-than-human speeds, which might be used as a sort of low-cost retirement for ems who have completed their working lives.
2. Most ems will probably live at subsistence.
We live in a world where the supply of human labour is limited by biology. Ems will not be so limited. Once a single em exists, making a copy will only be as costly as the processing power needed to run that copy. This means that the value of em labour will fall to the marginal cost of running an em. The em economy is a Malthusian economy, where the em population can vary instantaneously to keep up with the need for em labour.
However, subsistence might not be as bad for an em as it has been for most humans through history. Ems need not fear starvation or disease. Their consumption goods will all be simulated, and in a world of extremely cheap processing power, simulated luxuries would be cheap as well.
3. An em can work 99 percent of the time and go on vacation for 99 percent of the time, too.
This may seem paradoxical, but it follows from the possibility of creating and deleting copies at will. Suppose you have one em plumber. Each day he can make 99 copies of himself, in order to perform 99 plumbing jobs while he relaxes on a simulated beach, deleting the copies at the end of the day. While 99 percent of his processing power is being used to complete plumbing jobs, each em experiences a life of leisure followed by a single day of work.
4. Biological humans will be a true rentier class.
In a world populated by ems, the value of human labour will fall to near zero. An em brain can do anything a human brain can do, and ems will be produced until their marginal value falls to the cost of processing them. Biological humans won't be able to count on the value of their labour to sustain them, but they will earn vastly more from the wealth they already own. An em economy will grow very quickly, and thus will be able to give very high returns to the owners of capital.
5. The age of em might only last a few years before the next major change.
Robin compares the development of an em economy to three past changes in our society: The evolution of our latest non-human ancestors into humans, the move from a hunter-gatherer society to a farming society, and the birth of our modern industrial society. He observes that with each transition, the growth rate (measured as the increase in brain size before the evolution of humans and as economic growth thereafter) has increased and the period between transitions has shrunk. As ems will be able to experience far more time than we do, and since an em economy will be capable of extremely high growth, it won't take long for em society to produce the next radical shift. Perhaps just a year or two.
What will that shift entail? Robin declines to speculate, as there are too many degrees of freedom to predict with any degree of accuracy.
Read Scott Alexander's review of the book, which I mentioned during the interview.