100万亿Token实证AI研究_从交互到推理的范式转移_36页_10mb

报告摘要

Shift to Reasoning Models: The release of models like o1 in 2024 marked a move from single-pass generation to multi-step deliberation, accelerating real-world applications and usage patterns.
Open vs. Closed Source: A balanced ecosystem emerged, with proprietary models dominating enterprise tasks and open-source models excelling in roleplay and programming. Chinese open-source models saw rapid growth, contributing significantly to token volume.
Usage Categories: Programming and creative roleplay are the dominant use cases, accounting for most open-source token usage. Geographically, usage is globalizing, with Asia's share rising sharply.
Agentic Inference: LLMs are increasingly used for multi-step reasoning, tool invocations, and complex interactions, signaling a shift from single-turn to persistent agent-like workflows.
Retention Patterns: Early adopters ("Glass Slipper effect") show strong stickiness, forming foundational cohorts that resist model churn, while later cohorts have high churn.
Cost vs. Usage: Lower costs drive volume in some areas (Jevons Paradox), but high-value tasks retain pricing power. Open-source models compete on efficiency, while proprietary models focus on quality and reliability.
Overall Implications: The study highlights a fragmented, multi-model AI landscape where empirical data drives understanding, countering assumptions about AI deployment in productivity and creative contexts.