As we approach the end of 2025, reflections on the state of AI reveal a field that has undergone a profound transformation—both in technical capability and in the way researchers and practitioners perceive large language models. For years, a vocal minority of AI researchers insisted that LLMs were nothing more than "stochastic parrots"—mechanical systems that generate text based solely on statistical patterns, without any real understanding of meaning or intent. They claimed LLMs lacked internal representations of the prompt and had no awareness of what they were about to say. By 2025, this narrative has largely collapsed, as the weight of evidence and real-world performance has forced a reevaluation. A key driver of this shift is the widespread adoption of chain of thought (CoT) reasoning. CoT is not a new model architecture, but a method of prompting that guides the model to break down problems step by step, producing intermediate reasoning before arriving at a final answer. The real power of CoT lies in two interrelated mechanisms: first, it allows the model to perform a form of internal search, sampling from its learned representations to identify and activate relevant knowledge and concepts already present in the context window. Second, when combined with reinforcement learning (RL) using verifiable rewards, CoT enables the model to learn not just what to say, but how to say it in a way that leads to better outcomes—each token subtly shifting the model’s internal state toward a more useful response. This evolution has fundamentally changed the game. The idea that scaling is limited by the number of tokens in a context window is no longer valid. With RL and clear feedback signals—such as correctness, efficiency, or logical consistency—LLMs can now continue to improve over long sequences, even in complex, open-ended tasks. While we are not yet at the level of AlphaGo’s legendary move 37, the potential for such breakthroughs in AI reasoning is no longer a distant dream. Tasks like optimizing code for speed, for instance, could be iteratively improved with a clear reward function, allowing models to make progress over extended reasoning chains. The impact on software development is especially striking. Resistance from programmers to AI-assisted coding has dropped dramatically. Even with occasional errors, the value delivered by LLMs in generating working code, suggesting improvements, and offering context-aware suggestions has reached a point where the return on investment is compelling for most developers. The field is now split between those who use LLMs as collaborative partners—interacting with them through web interfaces like Gemini or Claude—and those who treat them as autonomous coding agents, capable of executing entire workflows with minimal human oversight. Meanwhile, a new wave of research is exploring alternatives to the dominant Transformer architecture. Some prominent AI scientists believe that the success of Transformers may be repeated, or even surpassed, by entirely different paradigms—models with explicit symbolic reasoning, world models, or structured internal representations. Yet, despite this interest, the current trajectory suggests that LLMs, as differentiable systems trained to approximate discrete reasoning steps, may be sufficient to reach artificial general intelligence (AGI), even without a complete architectural overhaul. The path to AGI may be multiple, and it’s possible that many different approaches will eventually converge on the same goal. There is also a growing debate about whether CoT has truly changed the nature of LLMs. Some claim that it has transformed them from mere pattern matchers into true reasoning systems. But this is misleading. The underlying architecture remains the same: a model trained to predict the next token. CoT is still generated one token at a time, and the reasoning is constructed through the same autoregressive process. The difference is not in the model’s core, but in how it is used and guided. Finally, the ARC (Abstraction and Reasoning Corpus) test, once seen as a litmus test that LLMs could never pass, has become less of a barrier. Small, task-optimized models now perform well on ARC-AGI-1, while large LLMs with extensive CoT reasoning achieve strong results on ARC-AGI-2—outperforming expectations for their architecture. In this way, ARC has transitioned from being a challenge to LLMs into a validation of their growing capabilities. Looking ahead, the central challenge for AI over the next two decades is not just technical progress, but ensuring that this progress is aligned with human values and safety. The goal is not just to build smarter machines, but to build them in a way that ensures our long-term survival.