OpenAI recently introduced its advanced AI system, o1, which boasts a new level of capability and claims to closely mimic human thought processes. This release reignited debates on the timeline for achieving artificial general intelligence (AGI)—a machine capable of performing the full range of human cognitive tasks, such as reasoning, planning, and learning from the environment. AGI, if achieved, could solve complex challenges like climate change and diseases but also poses significant risks if misused or uncontrollable, warns Yoshua Bengio, a deep-learning expert. Despite the advances in large language models (LLMs) like o1, experts like Bengio argue that these alone cannot achieve AGI due to missing elements in their architecture and functioning. LLMs have transformed AI by adopting transformer architectures that allow them to learn language patterns in a manner reminiscent of human cognition. This has enabled them to perform complex tasks like solving mathematical problems and generating computer code. However, their ability to adapt and recombine learned knowledge to tackle entirely new tasks—a hallmark of AGI—is limited. Some researchers have noted that while LLMs like o1 incorporate advanced methods, such as chain-of-thought prompting for improved performance in problem-solving, they still face constraints in handling tasks with extensive planning or abstract reasoning requirements. The success of transformers in processing diverse data types suggests a potential path toward AGI.

Yet, challenges remain, including the finite availability of data for model training and diminishing returns from model scaling. Moreover, more adaptive models that can construct and use world representations might be necessary for AGI, resembling human neural systems with internal feedback mechanisms for perception and planning. While some researchers are beginning to explore new architectures that incorporate feedback loops and efficient data usage, the journey towards AGI remains nascent. The need to ensure AI safety through regulation and inherent model safeguards is critical, according to Bengio and others. There is general agreement among scientists like Melanie Mitchell and George that AGI is theoretically achievable, evidenced by human intelligence. However, predictions on its arrival vary, with some estimating AGI could be a few to ten years away. Its true impact might unfold gradually rather than through an abrupt breakthrough.