I have argued that we are at a unique point in the development of technologies that are critical to the realization of artificial minds. I have even gone so far as to predict that human-level intelligence and physical ability will be achieved in about fifty years. I suspect that for many familiar with the history of artificial intelligence such predictions will be easily dismissed. Did we not have such predictions over fifty years ago? Some have suggested that the singularity will occur by 2030 (Vinge 1993), others by 2045 (Kurzweil 2005). There were suggestions and significant financial speculation that AI would change the world economy in the 1990s, but this never happened. Why would we expect anything to be different this time around?
In short, my answer is encapsulated by the specific technological, theoretical, and empirical developments I have described above. I believe that they address the central limitations of previous approaches to artificial cognition, and are significantly more mature than is generally appreciated. In addition, the limitations they address—such as power consumption, computational scaling, control of nonlinear dynamics, and integrating large-scale neural systems—have been more central to prior failures than many have realized. Furthermore, the financial resources being directed towards the challenge of building artificial minds is unprecedented. High-tech companies, including Google, IBM, and Qualcomm have invested billions of dollars in machine intelligence. In addition, funding agencies including DARPA (Defense Advanced Research Projects Agency), EU-IST (European Union—Information Society Technologies), IARPA (Intelligence Advanced Research Projects Agency), ONR (Office of Naval Research), and AFOSR (Air Force Office of Scientific Research) have contributed a similar or greater amount of financial support across a wide range of projects focused on brain-inspired computing. And the two special billion dollar initiatives from the US and EU will serve to further deepen our understanding of biological cognition, which has, and will continue, to inspire builders of artificial minds.
While I believe that the alignment of these forces will serve to underpin unprecedented advances in our understanding of biological cognition, there are several challenges to achieving the timeline I suggest above. For one, robotic actuators are still far behind the efficiency and speeds found in nature. There will no doubt be advances in materials science that will help overcome these limitations, but how long that will take is not yet clear. Similarly, sensors on the scale and precision of those available from nature are not yet available. This is less true for vision and audition, but definitely the case for proprioception and touch. The latter are essential for fluid, rapid motion control. It also remains to be seen how well our theoretical methods for integrating complex systems will scale. This will only become clear as we attempt to construct more and more sophisticated systems. This is perhaps the most fragile aspect of my prediction: expecting to solve difficult algorithmic and integration problems. And, of course, there are myriad other possible ways in which I may have underestimated the complexity of biological cognition: maybe glial cells are performing critical computations; maybe we need to describe genetic transcription processes in detail to capture learning; maybe we need to delve to the quantum level to get the explanations we need—but I am doubtful (Litt et al. 2006).
Perhaps it goes without saying that, all things considered, I believe the timeline I propose is a plausible one.[1] This, of course, is predicated on there being the societal and political will to allow the development of artificial minds to proceed. No doubt researchers in this field need to be responsive to public concerns about the specific uses to which such technology might be put. It will be important to remain open, self-critical, and self-regulating as artificial minds become more and more capable. We must usher in these technologies with care, fully cogniscent of, and willing to discuss, both their costs and their benefits.
Acknowledgements
I wish to express special thanks to two anonymous reviewers for their helpful feedback. Many of the ideas given here were developed in discussion with members of the CNRG Lab, participants at the Telluride workshops, and my collaborators on ONR grant N000141310419 (PIs: Kwabena Boahen and Rajit Manohar). This work was also funded by AFOSR grant FA8655-13-1-3084, Canada Research Chairs, and NSERC Discovery grant 261453.