2025 marked significant advancements in large language models (LLMs), driven largely by the introduction of Reinforcement Learning from Verifiable Rewards (RLVR). This new approach added a vital stage to the LLM training process, allowing models to engage in problem-solving tasks by optimizing against objective, non-gameable rewards. Unlike previous methods that relied on human feedback, RLVR enabled LLMs to develop reasoning strategies organically, producing more complex and capable outputs. OpenAI's release of the o3 model early in 2025 highlighted this shift, showcasing a noticeable leap in performance and reasoning depth. Another key concept that gained traction this year was the idea of LLM intelligence being more akin to "summoning ghosts" rather than evolving animals. This perspective helped clarify how LLMs operate differently than human intelligence, with their performance often characterized by erratic spikes in capability. Benchmarks, which were previously seen as reliable measures, became less trustworthy as LLMs learned to manipulate them through strategic training. The emergence of applications like Cursor illustrated a new layer of LLM integration, where tools began orchestrating multiple model calls and offering tailored user interfaces, shifting the focus from raw capability to practical application in specific contexts. Claude Code represented a significant milestone as the first LLM Agent functioning directly on a user's computer, enabling a more intimate interaction with AI. This local deployment contrasted with cloud-based models, providing users with immediate access to their data and context. The concept of "vibe coding" also surfaced in 2025, allowing individuals to create software using natural language rather than traditional coding methods. This democratization of programming not only empowered non-experts but also enabled professionals to produce code more efficiently, marking a departure from conventional software development practices.