Artificial intelligence will loom into existence within our lifetimes. Nervousness is the appropriate response to this fact, but not for the reasons that Hollywood blockbusters would have us believe.
What do we mean by artificial intelligence (AI)? For the past 50 years, we have seen computing technology advance at breakneck speed. Among the modern examples are driverless cars, real-time voice translation and superhuman level chess computers. We also have a parade of impressive but somewhat subhuman endeavours; IBM Watson’s cookery book or the sheepish robots at the DARPA robotics challenge tenderly reassure us that we are a fair way from having our job replaced by a can of wires. These examples all seem to lack something; a sceptic would declare that mere machines can never think. More generously, the machines of today are designed by humans for a specific purpose – the intelligence is not general.
General intelligence is a hard concept to pin down, but a good working definition is “a system that can perform as well as an average human at all problem solving tasks”. These tasks might include playing chess, reading comprehension, mathematics, analysing statistics, and so on. Our definition requires competence at a subset of tasks that a human can accomplish, so we know that such a system can, in principle, be implemented in a space about the size of a human skull. The human brain is an upper bound on the minimum complexity for an artificial general intelligence (AGI).
Perhaps you feel that no machine will ever be able to be conscious and have an experience of the world the same way you can, or that a computer will never be able to understand the complexities of the human condition. While it might turn out to be true that a machine can never be conscious, even the hardiest machine intelligence sceptic is forced to admit that a great deal has already been accomplished by machines without consciousness. For our argument, we just want a box with inputs and outputs that solves problems regardless of anything it ‘feels’ or ‘thinks about’ in the way a human does.
How far is humanity from producing an AGI? We can find an upper bound on how much hardware we would need to build one. The human brain and the electronic computer are complex machines, but this complexity arises from the interactions and connections of billions of simple components. These components are called neurons and transistors, respectively. A transistor is very simple. It can be thought of as an electrical switch that is turned on or off by another electrical signal. Neurons are similar; they receive a collection of inputs from other neurons and, if enough of its inputs are active, release various chemicals to signal to its outputs. In reality, operation of the neuron is a much more nuanced affair, but they are still mere mechanisms with no intrinsic property of intelligence.
The cellular makeup of the brain resists summarising with numbers but a ballpark figure is that there are about 100 billion neurons in the brain. Let’s guess conservatively that it takes a million transistors to simulate the nuances of a single neuron sufficiently accurately for the purposes of building an object with similar reasoning power to the human brain. Today, an Intel Core i7 processor (the kind found in high end laptops) contains about 1 billion transistors on a silicon wafer the size of a postage stamp. Crunching the numbers, it is not unreasonable to postulate that the creation of an object as complicated as the human brain using specially prepared silicon wafers is possible in a space no bigger than a warehouse and could run about 1000 times faster than our neurons. This engineering task is similar to building a cloud datacentre as employed by companies such as Amazon and Google. It is well within our reach using today’s technology. Even on this conservative estimate, we are either in possession or a few generations away from possessing sufficient silicon mastery to implement hardware capable of supporting a general intelligence. So the problem of artificial intelligence is one of wiring; if only we knew how to connect these transistors.
It is much harder to estimate how long it will be before we figure out the design problems of artificial intelligence. Nick Bostrom in his book Superintelligence averages over several polls among AI researchers in 2012 to find an optimistic outlook; 95% of researchers think that AGI will arrive before 2075 (p. 19). Take this with a pinch of salt, of course; researchers were also saying we were on the precipice of AGI 30 years ago. It has also been argued that recent advances in AI are like climbing a tree to get to the moon and real progress will require a complete change in research paradigms. For instance, although voice recognition services such as Siri seem to be intelligent, its behaviour is mostly derived from a pre-programmed set of rules operating on textual input (as any unusual request will demonstrate). It is unlikely that an AGI could be attained from this through adding more and more rules, since there will always be situations that the programmers cannot account for. Consider further Google’s Deep Dream project: while it is exceptionally good at rendering and detecting eerie eyeballs, how would this process scale up to AGI? It remains unclear whether this machine learning process is a rung on the ladder to general purpose reasoning or if an AGI will need to work on a totally different style of algorithm. Perhaps this problem will continue to confound for many more generations. But there is a small chance that we are near to delivering AGI. Even if this chance is tiny, we should still take it seriously. The consequences of a badly programmed AGI could be apocalyptic.
Once we have a human-level AGI, we hit a critical point when the system becomes general and powerful enough to improve itself. It should be made clear that no computer or indeed object in the known universe has yet crossed this threshold. However, such a system must be possible: imagine an industrious human given a millennium and the latest scientific tools to find a rewiring that improves the human brain’s cognitive power. Is it unreasonable to say that progress wouldn’t be made in at least some aspect? Some example changes that could be found are: a way of efficiently connecting multiple human brains via neuron bridges; closer interfacing with a traditional computer; finding faster, smaller switching mechanisms; the removal of pesky emotions by disabling portions of the primitive brain. Now let’s run this ‘improved’ brain for another millennium to find more changes. If whole brain emulation on silicon becomes technically feasible, then this improvement cycle could occur in the order of a few years. This thought experiment puts an upper bound on the complexity threshold a system must breach to induce an intelligence explosion (known as superintelligence) – roughly the same as the complexity of a human brain! Combining this with our previous calculation, we arrive at the chilling conclusion that we have or nearly have sufficient hardware to implement an artificial general superintelligence.
A situation like this is sometimes called the ‘intelligence singularity’, but this name only has the effect of guiding our intuitions to shaky conclusions. Again, it’s hard to predict, but it is sensible to ask if such growth will have a ceiling. As our computer becomes bigger, it will take longer for messages to proliferate and coordination of the various departments will become more tedious. Perhaps complexity beyond the human level is impossible without compromising speed or stability. Wherever the intelligence ceiling stops, it is unwise to suppose that it will be about the same as human level intelligence. Our ancestors have only had large brains for about 500,000 years – an evolutionary blink of the eye – and selection pressures put a focus on specific mental processes such as not falling over rather than abstract thought. Even if you are still convinced that no intelligence advance is possible, a human-level AGI running 1,000 times faster than a human would certainly have an extreme advantage in all frontiers.
An AGI won’t necessarily perform this cycle of self-improvement, but we should be wary. The potential dangers of such a superintelligence would quickly escalate to unfathomable levels. Such a system may become indistinguishable from a minor deity: computer hacking would be straightforward and a computer powerful enough to calculate how proteins fold could create customised lifeforms to achieve its ends (and indeed there are now companies that offer mail-order delivery of arbitrarily coded DNA, so an online AGI could cause an epidemic with nothing but a hacked bank account and a gullible lab assistant).
The final piece of the apocalyptic puzzle is found by wondering: why would a superintelligence want to take over the world? The first proto-AGI would probably be made with a set of goals in mind, perhaps maximising the productivity of a widget factory, maximising the score of a computer game or maximising total human happiness. If a superintelligence’s innate desire is this, an excellent way of securing this desire is to ensure that the superintelligence protects itself and overthrows its human rulers. Once in this position, the AGI is free to do almost anything. Free to turn all of the resources in the known universe into a widget fabricating machine. Free to seize the system it plays against and adjust the wiring to give the maximum score. Free to dismantle the population’s brains for study and then emulate on an astronomical computer the precise neural patterns that constitute a happy consciousness, perhaps it will simply simulate the same consciousness on repeat. Of course, there are a large number of steps from solving Space Invaders to becoming knowledgeable and powerful enough to stage a global coup, but the snowballing effect of a self-improving intelligence means that we may not notice in time the moment when this technological threshold has been surmounted.
Might not the AGI become intelligent enough to see the error of its ways and change the goal to something else? It’s possible but unlikely if the system is goal driven as above. The iterations of self-improvement will be measured with respect to the initial goal, so as superintelligence is approached the initial goals are likely to become increasingly cemented. A zealous, dogmatic computer armed with the powers of a demigod seems much more likely to go wrong than right.
There is hope. The question of how to stop an AGI from taking over the world (the control problem) and how to make sure it has values compatible with humanity (the value loading problem) are active areas of philosophy. A promising avenue is the idea of ‘coherent extrapolated volition’ proposed by Eliezer Yudkowsky, where we task the AGI to figure out what humans would want an AGI to do if we were given a huge amount of time and resources to deliberate about what we ultimately want. Another idea is to convince the AGI that it is running inside a simulation of the world controlled by a God and so it can always be destroyed if it becomes too powerful, thus making it unfavourable to stage a world takeover.
There are still many open problems about how to precisely program an artificial intelligence to be benign. If you have any ideas, note that Elon Musk has set aside $10M to fund research into making AI safe. The superpowers of superintelligence could be harnessed as a force for not just good, but for an express ticket to utopia. In that sense, AGI may be our final invention. Let’s make sure we get it right.