The next big breakthrough will be AIs learning on the job (www.dwarkesh.com)

Recent research suggests that training AIs in diverse reinforcement learning (RL) environments to perform millions of tasks may be the key to achieving artificial general intelligence (AGI). Proponents argue that this approach can overcome fundamental issues in current AI models, such as data inefficiency and the need for continual learning, by simply increasing the scale of training. As AI models become more capable of solving complex tasks over extended periods, there’s optimism that novel architectures, such as transformers with larger context windows, can allow models to effectively learn in real-time and adapt without traditional updates to their underlying weights. However, a significant challenge remains in the practical application of this approach, particularly in domains that are less easily replicable, such as business strategy or complex decision-making. The article raises concerns about whether RL-trained AIs can truly generalize their training to real-world tasks due to inherent inefficiencies in learning from sparse, unstructured data. The future of AI models may hinge on bridging these gaps through architectural advancements that capture knowledge efficiently while allowing for continual learning—a critical area for research as the industry navigates the complexities of training AIs to perform tasks that humans excel at naturally.