The year 2025 has been marked by rapid and transformative developments in large language models (LLMs), yielding significant advancements. The following analysis outlines six key paradigm shifts that have notably altered the technological landscape. **1. Reinforcement Learning with Verifiable Rewards (RLVR)** By early 2025, the standard LLM production stack—comprising pre-training, supervised fine-tuning, and reinforcement learning from human feedback (RLHF)—underwent a fundamental change. RLVR emerged as a core technology, involving the training of LLMs in environments with automatically verifiable rewards, such as mathematics and programming. This approach enables models to develop problem-solving strategies that resemble reasoning, including decomposing tasks into intermediate steps. Unlike previous fine-tuning stages, RLVR involves extended optimization against objective reward functions, consuming substantial computational resources previously allocated to pre-training. This shift has resulted in models of comparable scale but with significantly longer reinforcement learning training durations. A notable feature is the new ability to control model capability by adjusting “thinking time” or the length of reasoning trajectories during inference. The release of models like OpenAI’s o1 and o3 demonstrated the qualitative leap enabled by this technology. **2. The Spectrum of Intelligence: ‘Ghost’ vs. ‘Jagged’** A clearer conceptual understanding of LLM intelligence emerged in 2025. The consensus is that LLMs represent a form of “ghost” intelligence, fundamentally distinct from biological, “animal” intelligence due to differences in architecture, training data, and objectives. LLMs are optimized for mimicking text and achieving rewards in verifiable domains, leading to a “jagged” performance profile. They can exhibit expert-level capability in specific, trained areas while displaying profound limitations in others, making traditional benchmark evaluations increasingly problematic. As benchmarks typically represent verifiable environments, they are susceptible to targeted optimization via RLVR and synthetic data, leading to a prevalent practice of “training on the test set.” **3. Cursor and the New LLM Application Layer** The rapid ascent of Cursor highlighted the emergence of a distinct “LLM application” layer. Such applications specialize in orchestrating LLM calls for specific vertical domains. Their core functions include context engineering, managing complex call graphs, providing domain-specific user interfaces, and offering adjustable autonomy sliders. This development sparks debate on the future division of labor: while LLM platforms may produce generalist capabilities akin to “university graduates,” specialized applications will be crucial for fine-tuning, integrating private data, and connecting to sensors and actuators to create deployable “expert teams” for specific tasks. **4. Claude Code: The Local AI Agent** Claude Code presented a compelling vision for AI agents by effectively combining tool use and reasoning in an iterative loop for complex problem-solving. Its most distinctive feature is its operation on a user’s local machine, deeply integrating with private environments, data, and context. This local-first approach contrasts with cloud-centric agent deployments, offering a more pragmatic and tightly coupled collaboration model for the current transitional phase of AI capability. It redefines the interaction paradigm, positioning AI not as a service to be accessed but as a persistent local entity. **5. The Rise of ‘Vibe Coding’** In 2025, AI crossed a critical threshold, enabling the creation of functional programs through natural language description alone—a practice termed “Vibe Coding.” This paradigm dramatically lowers the barrier to programming, empowering non-specialists while simultaneously augmenting professional developers. It allows for the rapid prototyping of ideas, the creation of disposable code for specific tasks (like debugging), and the implementation of projects that would otherwise be impractical. Vibe Coding is poised to reshape software development ecosystems and redefine professional boundaries. **6. Nano Banana and the LLM Graphical Interface** Google’s Gemini Nano banana project signaled a major shift in human-LLM interaction. The current primary mode of text-based dialogue is analogous to command-line interfaces in early computing. The future lies in LLMs communicating through human-preferred mediums like images, infographics, animations, and interactive applications. Early steps include the use of visual text formatting (e.g., Markdown, emojis). Nano banana serves as an early prototype of this future graphical interface, notable not just for its image generation but for the integrated capability blending text generation, image creation, and world knowledge within a single model. *Disclaimer: This article summarizes the views expressed by Andrej Karpathy. It is for informational purposes only and does not constitute investment advice.*
