Recent discussions highlight advancements in enhancing the reasoning capabilities of Large Language Models (LLMs), particularly through a paper from Meta. The research focuses on Critical Token Analysis, which aims to improve LLMs' performance in complex reasoning tasks, such as mathematics. The paper suggests that instead of forcing AI to communicate in human language, it could potentially think directly in neural patterns, akin to how the human brain processes thoughts. This approach could significantly boost the reasoning power of LLMs. Experts express excitement over the potential of reasoning in a continuous latent space, questioning whether text or pixel representations are optimal for reasoning tasks. The consensus among researchers is that unlocking the full potential of LLMs is essential for the future of AI.