2025 到底是 LLM 的「什么年」？

Group 1 - The year 2025 is characterized as the "Year of LLMs," with significant advancements in technology, application paradigms, ecosystem dynamics, and risk governance, summarized by Simon Willison in 27 key themes [1][5]. - The focus on "Reasoning" and "Agents" highlights the evolution of LLM capabilities, where reasoning models are now more stable in driving toolchains and agents are increasingly defined and utilized in coding and search scenarios [9][12]. - Willison's analysis indicates that 2025 will see LLMs capable of planning multi-step actions and executing external tool calls, thus enhancing task completion chains [9][12]. Group 2 - The "Year of Long Tasks" discusses how agents can now handle longer-term engineering tasks, transitioning from demonstration to delivery due to advancements in reasoning and planning capabilities [10]. - The "Year of Coding Agents and Claude Code" emphasizes the scalable delivery forms of coding agents, exemplified by Claude Code, which lowers implementation barriers through local CLI and cloud asynchronous delivery [10]. - The "Year of LLMs on the Command-Line" addresses the shift from command-line as a toolchain language to a natural language interface, enabling broader accessibility for developers unfamiliar with command-line scripting [10]. Group 3 - The article also covers competitive dynamics in the LLM market, discussing the fleeting nature of "MCP" and the emergence of top-ranked Chinese open weight models, reflecting changes in the ecosystem and associated security risks [11]. - The advancements in reasoning capabilities are driven by methods like RLVR, with nearly every major AI lab releasing at least one reasoning model in 2025, indicating a significant supply-side shift [12]. - Applications such as "AI Search" and "AI Coding" are expected to materialize in 2025, showcasing the practical implications of enhanced LLM reasoning abilities [13].