Where are the recent improvements in AI coming from mostly?

I understand that on the onset of 2023 it was scaling up mostly. Based on training data, parameters and compute. Then MOE. Recent updates are because of post training and fine tuning as well as reward model policy and reinforcement learning for frontier AI systems. But even they would hit a limit after some time and that would be that. Are there any future directions where AI can keep improving and how to know of the recent research in frontier systems and methods used. Especially for closed AI systems.