Recent developments in artificial intelligence (AI) have seen AI exceed human performance on professional exams, global math olympiads, and coding competitions. Generative models are now capable of creating photorealistic images and videos as well as copying human voices with only three seconds of reference audio. These advances are astonishing, yet the fundamental ideas that underlie the most capable modern AI models have existed since the 1950s, so why has it taken this long for the technology to take off? If this scaling continues, what should we expect to see in the next few years? What does this mean for the job market, or the economy as a whole? Will this technology surpass human intelligence? In this talk, I will address these questions and more as we examine AI safety through a physics perspective. We will use concepts such as exponential scaling, information theory, and the importance of incentive structures to understand the risks and unknowns inherent to AI. Ultimately, bringing us to a discussion of how the vast promises of AI may be overshadowed by the existential risks to human civilization posed by the ongoing AI arms race.
