Towards artificial general intelligence
Tony J. Prescott · 2024
This chapter looks at the path from single-purpose AIs to artificial general intelligence (AGI) defined as an AI that is as flexible as human intelligence in its ability to work across different domains and to adapt to new challenges. AI’s capacity to understand is discussed in relation to Frege’s theory of meaning as comprising sense—how words relate to other words—and reference—how words relate to the external world. Large language models are claimed to already have meaning in relation to sense but not in relation to reference. To obtain a better grasp of meaning, AIs will require a greater capacity for unmediated two-way interaction with the world through robotic bodies. This can be evaluated through a “total Turing test” that requires AIs to match humans in terms of both their linguistic and robotic abilities. The broader class of generative AIs is discussed in relation to theories of predictive processing in the human brain, and an argument is made for a nuanced view of multiple intelligences, including that humans combine fast, pattern-recognition intelligence with a slower sequential reasoning capacity. The chapter concludes by arguing that both robots and humans can be considered to belong to the wider class of “cyber-physical systems”.