Photo by Pavel Danilyuk on Pexels

A Reddit user has ignited discussion within the AI community with a novel reinforcement learning (RL) approach aimed at achieving Artificial General Intelligence (AGI). Criticizing the current RL paradigm’s dependence on human-designed environments as a limiting factor, the user proposes incorporating a self-supervised learning component. Inspired by biological reward mechanisms, the system would learn its own reward proxies, analogous to a dog learning the association between a clicker and a treat. This, they argue, could foster the development of abstract interests and complex relationships within the model, significantly reducing reliance on initial human intervention. The complete proposal can be found on Reddit: [https://old.reddit.com/r/artificial/comments/1p6tm0o/my_take_on_ilyas_interview_a_path_forward_for_rl/]