For years, many AI researchers dismissed large language models (LLMs) as mere stochastic parrots, insisting they lacked any real understanding of meaning or context. However, by 2025, this perspective shifted significantly. The introduction of chain of thought (CoT) techniques has become a key factor in enhancing LLM performance. CoT allows models to sample relevant information more effectively, improving their ability to generate coherent responses. This internal search mechanism, combined with reinforcement learning, helps models better structure their outputs by adjusting based on prior tokens. The limitations previously associated with scaling LLMs no longer hold true, thanks to advances in reinforcement learning that incorporate verifiable rewards. While we're not yet at a breakthrough moment like AlphaGo’s famous move 37, the potential for LLMs to tackle complex tasks over extended periods is promising. Tasks like optimizing code can provide clear feedback, suggesting that future improvements in reinforcement learning could lead to significant advancements in AI capabilities. Resistance from programmers toward AI-assisted coding has decreased. Many have begun to accept LLMs as valuable tools, even if they sometimes generate errors. The programming community remains divided, with some using LLMs as collaborative partners and others treating them as autonomous coding agents. Meanwhile, a number of AI researchers are exploring alternatives to the Transformer architecture, seeking models that might leverage symbolic reasoning. Despite these explorations, the belief persists that LLMs, as they currently exist, could still lead to artificial general intelligence (AGI) without needing entirely new paradigms. The ARC test, once thought to be a major hurdle, now appears more manageable. Smaller, task-optimized models and larger, CoT-enhanced LLMs are showing promising results on different ARC benchmarks. What was initially seen as a challenge to LLMs has shifted to a validation of their capabilities. Amid these advancements, the overarching concern remains the existential risks associated with AI development over the next two decades.