Artificial intelligence within our lifetime? No idle speculation

By Kaj Sotala
Version 1.24, 7th of November, 2009.
Original version posted on 23rd of March, 2007.

Comment on this article - Read previous comments

Introduction

In recent years, some thinkers have raised the issue of a so-called "superintelligence" being developed within our lifetimes and radically revolutionizing society. A case has been made (see, for instance, [Vinge, 1993] [Bostrom, 2000] [Yudkowsky, 2006]) that once we have a human-equivalent artificial intelligence, it will soon develop to become much more intelligent than humans - with unpredictable results. A true artificial intelligence is not a tool to be used for good or ill, as technology usually has been - it is an independently acting agent. The power of a new kind of intelligence can be seen in the way Homo Sapiens rose to dominate the Earth in the blink of an evolutionary eye. Cultural and technological evolution has far outpaced the biological, as can be seen from the way our bodies and minds often seem maladjusted to our modern surroundings. If the human mind could beat evolution that badly, how badly might a new form of intelligence beat human minds?

Often, the thought of artificial intelligences being superior to human minds is not the hard part for people to grasp - instead, they question whether we really are anywhere near of developing AI. For as long as the thought of true machine intelligence feels remote, people feel that thinking about its consequences is a pure philosophical exercise at best. But when considering the immense impact that an AI will probably have, it becomes obvious that even the tiniest chance of it being developed is a cause for serious concern and attention. For one, we should do everything we can to ensure that any new artificial intelligences employ motivational systems which we can understand, so we can try to guarantee their safety.

This text will attempt to argue that there are several different ways by which artificial intelligence may be developed in the near future, and that the probability of this happening is high enough that the possibility needs to be taken account in our long-term planning. We define "near future" as within the next 50 years, and "artificial intelligence" as a machine intelligence capable of learning and reasoning all the things that an average human could. Covering the consequences of such a development in more detail is beyond the scope of this essay - for that, the reader is encouraged to look at the three references given above, with [Yudkowsky, 2006] being the most recommended for its comprehensiveness.

A note on developing technology

In this essay, we will discuss technology as it is now and as it will be in 50 years. People easily and often make the objection that these technologies are still in their infancy, which really shouldn't be an objection at all - that they're in their infancy is obvious, otherwise we'd have AI today. Equally obvious should be that they won't be as primitive 50 years from now. The field of psychology hardly even existed a hundred years ago, nor did computer science fifty nor the Internet twenty-five years ago.

Ray Kurzweil has argued that technological progress is accelerating exponentially [Kurzweil, 2005], an argument which seems hard to refute - just think of the ways developments like the Internet are improving communication and the accessibility of information. Think of how computers automated countless of complex calculations that would've otherwise required humans to work them out. Think of how urbanization and the Industrial Revolution accelerated the development of science. Think of how the printing press did the same, or the scientific method. When Moore's Law doubles the amount of transistors on a chip, that change is most certainly not a linear one - it's a pace that keeps accelerating year by year. In part, the accelerating pace of technology is predicted by ordinary economic theory, which states that new innovations fuel the growth of the economy, which in turn fosters again new innovations.

There is cause for believing that this trend will keep up. Drugs designed to improve memory and concentration are currently under development and should be entering markets in the future - some, like Modafinil, are already being used off-label for these kinds of purposes. Search engines are constantly being improved, with new techniques like Semantic Web being developed to help navigating immense amounts of data. Computing power available for researchers to run simulations increases every year, and so does the amount of money that formerly third-world nations can spend on educating new scientists. The aggregate effect of - among others - all these factors cannot be anything short of immense. Later in the essay, we will discuss entirely new things that'll accelerate development further. Estimations of the progress we'll make in the next 50 years are bound to be immense underestimations if they're linear extrapolations on the progress we've made during the last 50.

Motivations for developing Artificial Intelligence

It is not difficult to imagine why people would want to develop artificial intelligence. Pure curiosity and intellectual interest is one factor, but ever since the industrial revolution, corporations have been spending huge sums of money to replace workers with machines. Machines can do more work than humans, more cheaply and without tiring. Imagine having a generic mind template that was programmed to only love working for its owner and that asked for no pay, that could be cheaply copied, and which could be made to learn all the knowledge required to do the job of a doctor, a lawyer or an engineer in a matter of months. Without hesitation, major corporations would spend billions to develop it. Even much simpler AI-related innovations will be of immense economic benefit - artificial intelligence research will greatly benefit general information technology, and at least in the United States, information technology was responsible for two-thirds of total growth in productivity between 1995 and 2002 and virtually all of the growth in labor productivity [Atkinson & McKay, 2007]. It is safe to say that AI work will in all likelihood receive immense amounts of funding.

Approaches to Artificial Intelligence

Intelligence research is a highly interdisciplinary field, and the foundations of intelligence are a problem which can be attacked from several angles (to name a few: general computer science/math, computer hardware, evolutionary algorithms, brain imaging and modeling, biotechnology and general psychology/cognitive science) Progress in one field is likely to benefit others. As a good example, the foundation for the field of cognitive psychology was laid by advances in computer science.

General computer science and mathematics are what most people think about when hearing the word artificial intelligence. Advances in these two fields will help practically everything related to AI research as better algorithms for sorting through data are developed. Current "pure" computer science seems to have a long way to go before artificial intelligence will come out of it, but it has developed plenty of ways for solving different problems, some of which are arguably better than those used by the human brain. An understanding of general problem-solving methods will also make it easier to recognize equivalent algorithms which the brain uses.
Evolutionary algorithms are a subset of general computer science, but they - as well as other brute-force methods - deserve a special mention. It must be kept in mind that the human mind was never designed by anyone - it evolved through genetic drift and selection pressure. It might not be strictly necessary for us to understand how a mind works, as long as we can build a system that has enough computing power to simulate evolution and produce an artificial mind optimized to the conditions we want it to perform in. (It needs to be noted that this would be a very dangerous strategy, since we might not be able to control the motivations or other relevant details of the mind we produced.)
There are several brain imaging techniques which will allow us to study the functioning of the brain in more detail, hopefully leading to a reverse-engineering of the brain's function. This work is already in progress: parts of the brain's visual and auditory systems are pretty well understood, and researchers are now learning to recognize and identify different brain activation patterns. In one recent study, researchers were capable of reconstructing images the test subjects had imagined using brain imaging data [Thirion et al., 2006].
Biotechnology provides an alternative for traditional brain imaging methods. With a sufficient understanding of this field, it may eventually become possible for us to understand how the brain is built up by studying the human genome (while we'll never be able to understand the brain's composition simply by reading the DNA itself, an understanding of the conditions a human develops in may allow a program in which both sets of data are combined to create a detailed simulation of the growing brain). Even if this stage of understanding was never reached, the study of biotechnology is likely to very much aid and complement conventional brain imaging. The Allen Brain Atlas, a three-dimensional map of gene expression in the mouse brain, was completed in 2006, and the Gene Expression Nervous System Atlas, launched by the US National Institutes of Health in 2003, is mapping gene expression in the mouse brain throughout development [Bril, 2006].
General psychology and cognitive science is proceeding as well, helping us gain a better understanding of the methods the brain uses to understand and function in the world. A better understanding of these will aid in designing intelligent systems and in interpreting the results from the fields listed above.

For artificial intelligence to fail to bear fruit in the near future, nearly all of these fields need to fail to produce the breakthrough needed. Some potential near-future uses of them are explored below.

A word on basic assumptions

The assumption is made here that it will be sufficient to simulate the human brain on either the cellular or computational level, without modeling all the details of the molecular interactions. In other words, it is assumed that as long as we know how the cells react to different sorts of input and what sorts of input they get, it is unnecessary to simulate all the details of the molecules actually mediating the input (somewhat analogously, it's not necessary to detail every atom in an airplane when designing a plane that will fly). It can also be argued that it won't be necessary to actually model every neuron in the brain, as long as we can extract the mathematical operations they perform and implement them on a different system.

These assumptions are not universally accepted, but there is evidence in their support. For one, computer programmers wishing to port a program to another sort of system do not need to replicate the physical behavior of the chips down to the atomic level - it is sufficient to find out what operations are performed on the higher code level. Much of brain's functioning can be interpreted as mathematical models - for instance, neurons are considered to carry out a mathematical technique known as Fourier analysis in interpreting the visual data received from the eyes. Also, some of the higher-level principles employed in the brain have been studied and then successfully applied to computer programs (for one recent example, see [Serre, 2007]).

Indeed, for us to be able to closely observe a physical process yet be unable to translate it to a higher-level mathematical model would be a phenomenon nearly unprecedented in modern science. There do exist phenomena whose behavior cannot be accurately predicted on their own, but can be tracked using probabilistic or statistical models - for instance, it is extremely difficult to try to predict the behavior of a single gas molecule, but the behavior of a large amount of gas in a single area can be predicted with sufficient accuracy. This is actually not too distant from the principle of building higher-level mathematical models of the brain's behavior - such models may be unable to say what a particular neuron does, but they will accurately model the overall behavior of the system. It is further assumed that individual variations between people will not be large enough to seriously hamper our chances to build working models, an assumption which seems to be well-supported by current research.

There have been suggestions that the brain relies on quantum activity, making it impossible to model using ordinary computers. However, this suggestion is controversial and not widely accepted. We'll disregard it for the purposes of this article - however, it needs to be noted that even if it did hold true, much of this paper's suggestions could still be achieved using quantum computers instead of classical ones.

Understanding: Not required...
NOTE: This section is outdated. For a more up-to-date and far more comprehensive summary of the current prospects of uploading, see the Whole Brain Emulation Roadmap by the Future of Humanity Institute at the University of Oxford.

History knows of several inventions that were developed by accident, before anybody understood how or why they worked. X-rays were discovered before their operating mechanisms were understood. Many kinds of medicine are used even today without a full understanding about how they work. The artificial neural nets used in solving problems often produce solutions so complex that their exact mechanisms for working are not fully understood. The human brain never evolved because there was somebody who "understood" the mechanisms necessary to create a general intelligence and implemented it - it evolved by chance as a result of certain selection pressures.

We can imagine a number of ways in which an artificial intelligence could be created without us fully understanding it. We could muster enough hardware to implement a full simulation of the human brain, then use advanced brain scanners to scan ("upload") a brain into a computer. While this wouldn't be an artificial intelligence in the truest meaning of the word, it would be a human brain transferred onto an artificial substrate. Given enough time and an appropriate interface to the system it was running on, it could potentially become much more intelligent than ordinary humans, especially if it could observe the processes behind its own thoughts and alter them in real time. Other researchers could also observe the processes going on in such a mind, and implement partial simulations of isolated parts of the brain, creating detailed models of their responses to different inputs.

Alternatively, we could simply look at the ways how different brains run, and then build a computer program or an artificial brain that closely mimicked the behavior of human brains, or simply pick interesting algorithms from the brain and implement them in new computer programs. In any case, we wouldn't need to understand everything about the way the brain ran - simple replication ("brute forcing" the brain) would already accelerate our AI development immensely.

Is it realistic to expect such brain-scanning methods within our lifetimes? The resolution of noninvasive brain-scanning devices is doubling about every twelve months [Kurzweil 2005, p. 158-160]. Currently existing methods do not monitor brain activity per se, however, but blood flow within the brain. As such, there is a limit to how detailed their scans can get. Destructive methods are much more efficient - an instrument known as a Brain Tissue Scanner is purported of being able to scan a sliced mouse brain at a high resolution in a month [McCormick, 2002]. IBM's Blue Brain project is at a stage where they are intending to build simulated replicas of brains at a cellular level, also using information derived from invasive methods [IBM, 2004]. They estimate they'll have a whole brain simulation up and running within 10 to 15 years (while they stress that their research is about biological research and not artificial intelligence, it would be absurd to assume that research into the brain's biological functioning wouldn't also contribute to the study of intelligence). As of April 2007, researchers had simulated half a mouse brain on the BlueGene L supercomputer, with the simulation taking ten seconds to run the equivalent of one second of real life [BBC, 2007].

Real-time, high-resolution non-destructive scans are likely to become possible when nanoscale manufacturing techniques are perfected. Molecular nanotechnology is the engineering of functioning machines at a molecular scale. The classical example - which also serves as a proof-of-concept for their functioning - are ribosomes, cell organelles which assemble proteins according to the instructions in a being's genes. Nanotechnology investment is currently increasing world-wide - for instance, the amount of scientific papers containing the word "nano" has gone up from about 10,000 in 1999 to almost 40,000 in 2005 [National Academy of Sciences, 2006]. Total nanotechnology research and development funding was a $10 billion market in 2004, and is expected to triple by 2008 [Lawrence, 2005].

Nanotechnology could make very powerful scanning possible by allowing tiny probes injected into the brain to scan the undergoing molecular interactions in real time. Such machines would have the immense benefit of being able to record activity in the brain all the time - current brain research is limited by the fact that brain-scanning is currently a time-consuming and rather inconvenient process. Nanoprobes would also replace the current technique of Transcranial Magnetic Stimulation (TMS), in which magnetic fields are applied to the brain to selectively reduce or increase neuronal activity in certain areas. Being able to make local adjustments, nanoprobes could selectively excite or inhibit the activity of any part of the brain and allow researchers to observe the effects. Nanoprobes are no idle speculation - there have already been proposals about both medical nanoprobes in general (with one group of scientists promising robots traveling in arteries by 2009 [Cole, 2007]) and nanowires built to record brain activity in particular [Llinas et al, 2005] [Physorg.com, 2005].

Even if we couldn't image the living brain in full detail, other alternatives exist. Low-resolution scans could be used to create a rough map of the brain's connections, with evolutionary techniques - effectively a huge amount of trial-and-error runs and simple brute computing force - then being applied to develop these models into more advanced, intelligent ones. It worked for evolution, which didn't even have a specific goal in mind, so there's little reason to assume that it couldn't work for us. Within 50 years, our knowledge of biology and genetics could advance to the point where we could obtain a rough map of the brain's workings simply by scanning through the human genome.

Understanding: ...but useful.

The previously mentioned potential of nanoprobes is increased when we consider that these local adjustments need not be limited to research. It would be trivial to replace all existing psychopharmacological drugs with nanobots and their ability to selectively target only a small group of chosen brain cells, avoiding the side-effects caused by the imprecise "carpet bombing" that current medicine uses. Our understanding of the brain is likely to increase as incremental progress towards perfect delivery is made. As each refinement improves the accuracy of the delivery, our understanding of how and why exactly the drugs work will improve as well.

Full-detail brain scans made on high-IQ people could reveal what factors in the brain make certain people more intelligent than others, with nanoprobes then moving on to inexpensively optimize these factors in anybody who wanted it. If an increased amount of brain mass would increase intelligence, like some studies [McDaniel, 2005] seem to indicate, then nanomanufacturing techniques could be used to build artificial neurons to increase anyone's intellect. If there were any minds running as uploads, similar but more radical changes could be made on them, like increasing the amount of their simulated brain cells ten-fold. In any case, weak superintelligences would follow. If nanoprobes are unavailable, more traditional biotechnological approaches could be used for intelligence amplification.

What we have seen so far demonstrates that brain and artificial intelligence research has the potential to drastically accelerate further scientific study. As bits and pieces of the brain's workings are understood, the algorithms we learn can be applied into general computer science, giving us slightly better programs we can use in the analysis of data. The same understanding can be used to optimize the brain's workings in the general population. No doubt, the immense amount of data becoming available will be hard to sort through at first, but each step towards understanding the brain is likely to accelerate the completion of the next one - technological development will accelerate immensely.

When talking about uploads and artificial intelligences, we have only been talking about being able to scan the brain, but we also need to get the hardware up to spec. Again, molecular nanotechnology seems the most powerful way of solving the problem - evolution gave the human brain a large amount of processing power using a rather limited amount of nanotechnological building blocks and ways of improving on them. Intelligent beings with a vastly larger amount of possible tools and manufacturing techniques can do better. As just one example, Eric Drexler [1992] demonstrated a purely mechanical nanocomputer with ~10^28 operations/sec-m^3 of processing power [Freitas 1999, 10.2.1]. This is five to nine orders of magnitude higher than the estimated processing power Blue Brain needs to run a cellular-level simulation of the human brain [Peck, 2007] - and the computer assumed in Nanosystems is purely mechanical and only the size of one cubic meter.

While these levels of computing power are still in the future, near-term proposals for increasing the processing power of current designs seem to keep popping up - some of them even besting Moore's Law. (See, for instance, [Itwire, 2007a] and [Itwire, 2007b]). Even ignoring the prospect of advanced nanotechnological computers, processing power development by more traditional means will give us the computing power necessary to replicate human brains within 25 years or so.

Limitations

Now that we have discussed some of the possible developments enabling the creation of AI, it is worth noting some factors that might prevent it from bearing fruit.

One, technology might simply not develop as fast as we've predicted, or deciphering the brain may be even harder than we assumed. This is always a possibility in making forecasts. Second, brain uploads and experimentation with them might be severely restricted by ethical and religious concerns. It must be noted that this is more likely to be the case in some Western countries - nations like China are much less likely to care about such issues.

Third, the so-called "Peak Oil" event or general resource exhaustion might curtail the speed of development. However, even global recessions typically do not affect the speed of research very much [Kurzweil 2005, p. 99]. Fourth, some of our basic assumptions might simply be wrong. Maybe the human brain relies on some bizarre way of operating that simply cannot be replicated in any other medium. This seems extremely unlikely.

Creating AI: Conclusion.

The breakthrough necessary for developing true artificial intelligence may be found via independent developments in general computer science, evolutionary algorithms, brain imaging, biotechnology, general cognitive science, or (most likely) a combination of the five. Due to the constant, exponentially accelerating development of technology, it is unsafe to say that any one of the fields would be too far away from such a breakthrough for us to concern ourselves with it - let alone saying this about each of them.

Understanding how exactly the brain functions is useful, but not necessary in order to create AI - and even brute-force methods are likely to be helpful. Every step towards the development of true AI is likely to speed up the general pace of research due to new forms of intelligence becoming available to researchers. It seems probable that full-brain uploads may become possible within our lifetimes, and true artificial intelligence will most likely be developed relatively soon after mind uploads become possible. Also, mind uploads have the possibility of making themselves at least weakly superintelligent. There are reasons for believing that these predictions might not hold, but none of them are convincing enough to dismiss the chance of developing artificial intelligence off-hand.

The immense impact of artificial intelligence has been discussed by several thinkers. While we cannot say with certainty that we will develop artificial intelligence within our lifetime, the combination of its non-trivial probability and the amount of effect it will have means that we must start paying attention to the issue as soon as possible. Particularly worrying is the possibility of us developing artificial intelligence without understanding the specifics of its functioning - resources should be spent on maximizing the chance of us first building AIs whose motivations (or goal systems) we can understand and that we can deem safe. (For further discussion on the consequences of creating AI, the author recommends his own Why care about artificial intelligence?, as well as Eliezer Yudkowsky's Artificial Intelligence as a Positive and Negative Factor in Global Existential Risk.)

Acknowledgements.

The author's thanks go to Andreas Carp�n and Aleksei Riikonen, as well as to one person who preferred to remain anonymous, for commenting on a draft version of this article.

References:

Atkinson, R.D. & McKay, A.S. (2007). Digital Prosperity: Understanding the Economic Benefits of the Information Technology Revolution. The Information Technology & Innovation Foundation. http://www.itif.org/files/digital_prosperity.pdf

BBC News (2007). Mouse brain simulated on computer. http://news.bbc.co.uk/2/hi/technology/6600965.stm

Bostrom, Nick (2000). When machines outsmart humans. Futures. Vol. 35:7, pp. 759 - 764. http://www.nickbostrom.com/2050/outsmart.html

Bril, David (2006). Mouse brain map is complete. Nature 443, 380-381. http://www.nature.com/nature/journal/v443/n7110/full/443380c.html

Cole, E. (2007). Fantastic Voyage: Departure 2009. Wired magazine. http://www.wired.com/news/technology/medtech/0,72448-0.html?tw=wn_technology_1

Drexler, K.E. (1992). Nanosystems: molecular machinery, manufacturing, and computation. John Wiley & Sons, New York. http://www.zyvex.com/nanotech/nanosystems.html Cited in [Freitas 1999, 10.2.1].

Freitas, R.A. (1999) Nanomedicine, Volume I: Basic Capabilities, Landes Bioscience, Georgetown, TX, 1999. http://www.nanomedicine.com/NMI.htm

IBM (2004). Blue Brain FAQ. http://bluebrainproject.epfl.ch/FAQs.htm

Itwire (2007a). HP beats Moore's Law with new chip architecture. http://www.itwire.com.au/content/view/8631/53/

Itwire (2007b). Intel to extend Moore's Law with new chips in 2007. http://www.itwire.com.au/content/view/8989/53/

Kurzweil, R. (2005) The Singularity is Near: When humans transcend biology. 1st edition, Penguin Books Ltd.