Large language models (LLMs) like ChatGPT have taken the world by storm, demonstrating impressive abilities to generate human-like text, translate languages, and even write different kinds of creative content. But beneath the surface, these powerful AI systems still grapple with fundamental reasoning challenges. A new study dissects OpenAI's 'o1' model—designed specifically to enhance reasoning capabilities—to uncover whether it escapes the limitations of its predecessors. The surprising finding? While 'o1' excels in many areas, it still exhibits peculiar weaknesses tied to the way it’s trained. Traditional LLMs learn by predicting the next word in a sequence, making them sensitive to the probability of words and phrases. This 'autoregressive' approach, while effective for generating fluent text, can hinder true reasoning. For example, reversing a common word like 'hello' is easier for an LLM than reversing an uncommon word, even though the logic is identical. The researchers explored whether 'o1,' optimized for reasoning, overcomes this limitation. They tested 'o1' on various tasks like deciphering codes, manipulating word order, and solving simple math problems, focusing on both common and uncommon variations. While 'o1' consistently outperformed older LLMs, it still stumbled in low-probability situations, requiring more 'thinking steps' to arrive at an answer. This suggests that even when explicitly trained for reasoning, the underlying autoregressive nature of LLMs can’t be entirely ignored. These findings highlight the importance of understanding the limitations of current AI systems. While 'o1' represents a step forward, it reminds us that building truly intelligent machines requires more than just improving performance on specific tasks. The next generation of AI models will need to move beyond statistical prediction to incorporate deeper, more flexible reasoning mechanisms. Only then can we unlock the full potential of AI to solve complex real-world problems and understand the world around us.