Andrej Karpathy: 2025 LLM Year in Review

Quick Overview

Andrej Karpathy's 2025 LLM Year in Review highlights that the major shift was not just in model size, but in the fundamental framework for building and interacting with LLMs, evidenced by the emergence of self-improving agents like RLRV (Reinforcement Learning from Verifiable Rewards) and the move towards local, specialized models over massive cloud-based ones.

Key Points: The most significant shift in LLM development in 2025 was a total realignment in the conceptual framework for building and interacting with these models, moving away from simple scaling. The paper by Carpathy introduces RLRV (Reinforcement Learning from Verifiable Rewards) as a crucial new ingredient that broke the previous formula of massive pre-training followed by supervised fine-tuning (SFT). RLVR models receive feedback tied to tasks (like solving math puzzles or coding challenges) that are verifiable, unlike subjective rewards used previously. The trend is moving towards local, specialized LLMs (like a 'banana demo' model) that run on user hardware, contrasting with the previous reliance on massive cloud-based models like those from OpenAI. The concept of 'agents' that can orchestrate complex LLM operations (like reading documents, summarizing, and generating reports) is becoming the standard deployment method. The shift means that programming is evolving from writing explicit code to developing sophisticated, specialized 'agents' that manage complex operational chains, often with a graphical user interface (GUI).

Context: Andrej Karpathy reviews the state of Large Language Models (LLMs) as of 2025, focusing on paradigm shifts that go beyond mere model scaling. The discussion centers on new training methodologies, particularly Reinforcement Learning from Verifiable Rewards (RLRV), and the practical implications for how models are deployed and used, emphasizing a move toward local, specialized intelligence over generalized, cloud-hosted behemoths.

Detailed Analysis

Andrej Karpathy summarizes that 2025 felt fundamentally transformative for Large Language Models (LLMs) because the industry shifted its conceptual framework for building and interacting with them, moving beyond the simple scaling of previous years. The major breakthrough discussed is Reinforcement Learning from Verifiable Rewards (RLRV), introduced by Carpathy, which replaces the older paradigm reliant on massive pre-training and supervised fine-tuning (SFT). RLRV models are trained using rewards derived from verifiable tasks, such as solving complex math problems or coding challenges, where the output can be objectively checked for correctness, unlike previous subjective reward signals. This is contrasted with models optimized only for subjective rewards or mimicking human thought processes. The practical result is that these new models, like the 'jaggies' (jagged) models, are becoming adept at reasoning and problem-solving without needing constant human feedback loops. Furthermore, there is a significant move away from massive, cloud-based models towards smaller, local models that run on consumer hardware, indicated by the success of models like 'banana demo' and the development of specialized GUIs for these local applications. The source suggests this shift has created a crisis of confidence in older benchmarks because the new models are demonstrating capabilities—like writing complex, multi-step code or orchestrating tasks across different domains (law, finance)—that were previously unattainable, effectively making the underlying code maintenance less critical than the orchestration layer.