Written for The Nature of Mind (Philosophy 3), UC Berkeley, Fall 2019. On David Chalmers’ argument in The Singularity: A Philosophical Analysis.

In this paper, I will deconstruct Chalmers’ argument for the Singularity that he presents in his work ‘The Singularity : A Philosophical Analysis’ as well as providing an argument countering his claim of the inevitability of the Singularity. However, with due cognisance of the unpredictability of the future, I will also quantify the probable risk to benefit ratio of the Singularity, with the hypothesis that the risks outweigh the benefits. Finally, I will attempt to build upon Chalmers’ suggestions of how we can ensure our safety if we deem the Singularity to be unavoidable.

The Singularity refers to the idea that there exists a point in time when the creation of an Artificial Intelligence (AI) entity will result in an ‘intelligence explosion’ and whereby an intelligence far superior to that of human intelligence will be created. An ‘intelligence explosion’ can refer to any recursive mechanism that establishes such a superintelligence. The argument is that if we design an AI that is more intelligent than us, it will possess the capacity to design an even more intelligent AI, which in turn can design even better AI and so on. Similarly, a ‘speed explosion’ is a logically independent idea, that relies on a version of Moore’s Law that holds that faster processing can lead to AI that can design superior transistors that allow for even faster processing. Intuitively one can see how both intelligence and speed explosions are complementary as the manifestation of one improves the likelihood of the other.

Chalmers formalises the argument of the singularity into three chronologically dependent premises that are represented below -

1) Equivalence Premise - There will be AI before long that is equivalent to human intelligence

2) Extension Premise - If there is AI, there shall be AI+ that is more intelligent than any human

3) Amplification Premise - If there is AI+, there shall be AI++ that is superintelligent

4) Conclusion - There will be AI++

Chalmers further qualifies each premise with a timeframe as well as adding potential ‘defeaters’, that he defines as “anything that prevents intelligence systems from manifesting their capabilities to create intelligence systems” [Chalmers, 2010]

I reject Chalmers’ argument for the inevitability of a Singularity. I will allow that it is a possibility within a long enough time frame and if certain technological and philosophical advances occur. However, I will argue that from our current state of knowledge it seems unlikely that we shall ever create AI++. My rejection stems from a primary objection to the equivalence premise, a skepticism of the amplification premise and a belief in the likelihood of a defeater manifesting before these objections can be dealt with.

First, I reject the equivalence premise and the idea that we will create AI before long that is equivalent to human intelligence. Chalmers argues that there are four primary ways we could create such an AI, through direct programming, emulation, evolution or machine learning. Chalmers admits that direct programming is unlikely to result in AI but delineates how the other three methods could be successful. For emulation, he argues that a) The human brain is a machine, b) We will have the capacity to emulate the machine and c) If we emulate this machine, there will be AI. I contest the first premise and would argue that viewing the human brain as a machine is at best, merely a metaphor steeped in physicalism that helps us in conceptualising the brain, and at worse, a crass oversimplification of the complexity of our brain. My claim would be that we first need to understand how the brain works, a task still very much in its nascent stage, before we can think of emulating it. Moreover, as Chalmers notes, emulation is contingent on a functionalist notion of human cognition and it being substrate independent, a claim that I would dispute. For evolution, Chalmers proposes the following argument, that a) Evolution produced human level intelligence b) If evolution did so, so can we c) There will be AI. Similar to the emulation argument, I will argue that our limited understanding of the sheer complexity involved in the 3.8 billion years of evolution prevents us from the lofty task of designing a similar evolutionary process for creating AI. Furthermore, biological evolution involves processes like embryogenesis that are theorised to play a crucial role in the robustness of the evolutionary process and such processes are not yet replicable physically. Hence, both emulation and evolutionary methods seem unlikely to work till leaps in scientific progress are made. Finally, the multitude of problems that machine learning experts are grappling with and the vast chasm between the kind of crystallised intelligence that allowed for AlphaZero to beat humans in chess and Go and the general intelligence that human beings possess does not seem like it will be bridged anytime soon.

Secondly, I am doubtful of the possibility of the amplification premise, that the creation of AI+ will lead to AI++. As Chalmers notes, the premise depends on the proportionality thesis that states increases in intelligence always leads to increases in capabilities to design intelligence. He further notes that this thesis fails if either the limit of intelligence space is reached in the process of amplification or if the gains in intelligence follow a model of diminishing returns. I see both problems as possibilities, especially considering the fact that we already seem to be approaching a model of diminishing returns when it comes to Moore’s Law. Moreover, I am not convinced that the proportionality thesis is backed up by evidence and it seems likely that physical limits that cannot be overcome will act as constraints and prevent an amplification process. Finally, I realise that my objections, especially against the equivalence premise are refutable by allowing for a longer time frame. However, I would argue that a situational defeater - namely the extinction or the regression of the human species - is more likely to occur in an extended time frame than the creation of AI++. This statement is based on the consideration of the state of global climate change as well as the high possibility of nuclear war.

However, it is prudent to allow for the possibility of the Singularity and a consideration of the risk/benefit ratio of superintelligence is paramount to guiding us on how to align our collective decisions. It is fair to assume that a superintelligent entity would be able to solve most of humanity’s problems - find cures for all our diseases, end warfare and bring about a utopian world. However, execution does not follow from ability and it does not seem to be implausible that a superintelligent entity would choose to simply enslave us or even get rid of us all together. Definitionally, a superintelligent entity would relate to us similar to how we relate to mice and it seems dubious to trust that it would have our best interests at heart. This might be solved by carefully designing the creation of AI++ such that their values coincide with our goals. I will briefly go through possible ways we could achieve this below. However, I would overall conclude that the risk inherent in creating such a superintelligent being are too great, despite the attractiveness of a utopian future.

As noted above, our decision on whether to proceed with the creation of superintelligence should hinge on whether we could apply suitable constraints on its behaviour. Chalmers discusses two general categories that such constraints could fall under, internal constraints that could be applied to the very design of AI and external constraints that focus on restricting the influence of AI in our world. The latter proposition primarily consists of what he terms as the leakproof singularity, wherein we deny AI access to the physical realm and restrict it to the virtual. As Chalmers notes, such denial would be fragile as a superintelligent entity would be able to exercise influence and ‘get out’ as long as there is some input into the virtual world. I believe that applying internal constraints is a more promising approach with our best bet being imbibing human values in the design process as goals the behaviour of AI. As Chalmers notes, if Kant’s view on the correlation between morality and rationality is true, superintelligent beings will be super moral. While super morality does not necessarily mean humans will fare better, an AI that is motivated by human values should behave in a manner that suits our interests. The question then is how can we ensure that we are imbibing our values in the design process of AI such that our values remain uncorrupted. I believe our best answer will have roots in the concept of embodied cognition. Embodied cognition is a theory that key qualities and features of mind are shaped by the structure of the overall body. Accordingly, human values could best be distilled into AI by embodying the AI in some form, possibly through integration with our own human bodies or some form of uploading of our consciousness.

References

  • Chalmers, D. J. (2010). The singularity: A philosophical analysis. Journal of Consciousness Studies, 17(9–10), 7–65. ↗