The authors define artificial general intelligence by comparing it with human intelligence in terms of the breadth of tasks they can accomplish, and that they should be almost similar. The field of artificial intelligence (AI) is developing at tremendous rates, but in order to measure its progress toward AGI, we must first identify what intelligence and general intelligence really mean.

On pages 1-2, in support of the aforementioned context, the authors argue that general intelligence is a subset of intelligence based on one’s ability to learn skills and apply them to complex tasks and unfamiliar environments. That is, if one picks up the skills of cutting an apple, they wouldn’t have issues cutting a pear. By virtue of this, the authors extend that merely the ability to do specific tasks at greater speeds or with better arithmetic capabilities does not determine general intelligence. It is ‘cognitive flexibility’ that counts towards this intelligence, and this is an attribute prevalent in humans and other components of nature. An example in nature includes learning and communication in bees. This implies that bees understand which flowers have better nectar to optimize their work, and are also able to communicate to their fellow bees the location of food sources. A few more examples provided include “navigation in migratory birds, or the astonishing memory and tool use found in corvid birds” (Shevlin 2). Machines, on the other hand, are only able to achieve specific tasks, and cannot exhibit adaptability outside of certain domains. By that argument, AI in today’s time has narrow intelligence. For example, if an algorithm is trained with pictures of cats and dogs, but we present it with a horse, it would not know what to do.

Assuming the above definition of general intelligence holds true, while progress is being made in the field of transfer learning, we are yet to make any great progress in artificial general intelligence. But, an extended question to this is if we were to train AI agents in such a fashion that we taught them how to apply learnings to new environments, would they then be considered generally intelligent? For this to be possible, we would have to assume that teaching cognitive flexibility is truly possible. By virtue of this, we would understand emotional and moral beings more deeply, because this would make it clear whether it is a learned skill to be able to learn and adapt, or if it is inherent to humans and other animals. In other words, we would shed light on whether humans and animals possess a threshold intelligence that allows them to have this cognitive flexibility attributed to general intelligence. The idea of a necessary threshold intelligence would in fact halt our progress for artificial general intelligence entirely because we would have no attainable goal to work towards. To conclude, the likelihood of developing artificial general intelligence is based on whether we can teach AI how to be cognitively flexible.

Works cited: Shevlin, H., Vold, K., Crosby, M., & Halina, M. (2019). The limits of machine intelligence: Despite progress in machine intelligence, artificial general intelligence is still a major challenge. EMBO Reports, 20(10). https://doi.org/10.15252/embr.201949177