In 2017, the AI company DeepMind (now Google DeepMind) published an impressive paper detailing the capabilities of their algorithm, AlphaZero. This algorithm had the ability to teach itself various games without any instruction, achieving superhuman performance in chess, shogi, and Go within 24 hours of play. DeepMind's CEO, Demis Hassabis, noted that AlphaZero's moves were often surprising and did not resemble human or computer play, describing it as a third, almost alien, way of playing. While the researchers were not spooked by AlphaZero's capabilities, they were certainly surprised. Fast forward to the present, where OpenAI's powerful GPT-4 model is being experimented with. When posed with a scenario about a parrot in a covered cage, GPT-4 reasoned that Bob would believe the parrot to still be in the room, while Alice would believe it to be in the covered cage inside a cupboard. GPT-4 also noted that the parrot's perception and understanding would be limited to its immediate surroundings.

This interaction with the machine challenges the assumption that large language models like GPT-4 are incapable of reasoning. Researchers have been testing GPT-4's reasoning abilities and while it performs relatively well on established tests, there are still certain challenging tasks. However, the development of this technology is progressing rapidly. The question remains whether these models are a stepping stone towards artificial general intelligence (AGI) or superintelligent machines. While conventional wisdom says no, as these models lack knowledge of the world's complexity, there is no doubt that they are becoming increasingly capable. "GPT-4 can solve novel and difficult tasks in various domains without any special prompting and its performance is strikingly close to human-level, " according to a recent Microsoft study. It is important to closely monitor the advancements in this field.