The post discusses innovative methods in training Large Language Models (LLMs) to utilize their last hidden layer effectively, enhancing reasoning capabilities by incorporating a continuous latent space. Specific training techniques are highlighted, including a method where step-by-step text is gradually replaced with hidden state embeddings. This iterative process improves model performance significantly while raising questions about human-like reasoning in AI. Additionally, some concerns are voiced regarding the appropriateness of the terms 'reasoning' and 'thinking' in the context of LLMs, emphasizing the difference in cognitive processes between humans and machines. Potential avenues for further development include exploring effective loss functions and building deeper neural networks for enhanced reasoning capabilities.