Do leading LLM developers possess a proprietary ``secret sauce'', or is LLM performance driven by scaling up compute? Using training and benchmark data for 809 models released between 2022 and 2025, we estimate scaling-law regressions with release-date and developer fixed effects. We find clear evidence of developer-specific efficiency advantages, but their importance depends on where models lie in the performance distribution. At the frontier, 80-90% of performance differences are explained by higher training compute, implying that scale--not proprietary technology--drives frontier advances. Away from the frontier, however, proprietary techniques and shared algorithmic progress substantially reduce the compute required to reach fixed capability thresholds. Some companies can systematically produce smaller models more efficiently. Strikingly, we also find substantial variation of model efficiency within companies; a firm can train two models with more than 40x compute efficiency difference. We also discuss the implications for AI leadership and capability diffusion.

We develop a quantitative task-based model of automation in which deploying machines for a task, in addition to variable (marginal) costs, requires fixed costs (e.g., fine-tuning for AI models). A task is automated when its production has sufficient scale to amortize the fixed cost. Applying the model to the context of computer vision automation, we discipline the fixed and maringal costs using AI scaling laws relating performance to training costs. Estimating this relationship in a fine-tuning vision AI experiment, we additionally rely on LLM-generated measures of task complexity and error tolerance to infer the computational costs of automation for all vision-related production tasks in the economy. Calibrated to the U.S. firm-level adoption of computer vision AI in 2023, and fed with current estimates of the trend in falling computing costs, the model projects rapid diffusion, reaching 23% of firms (60% employment-weighted) by 2035, and shows that scale advantage accounts for roughly three-quarters of cross-task variation in comparative advantage. As computing costs fall, real output rises by about 7% by 2075, real wages increase throughout, and the labor share follows a U-shape as the aggregate elasticity of substitution between labor and AI declines with deeper automation.

Understanding the intent behind scientific citations is critical for advancing scholarly search and knowledge mapping. This paper reflects on the methodological use of large language models (LLMs) for multi-class citation intent classification. Our experiments evaluating a diverse range of models and approaches reveal striking disagreement among state-of-the-art (SotA) systems. This inconsistency suggests that citation intent classification remains a challenging task for LLMs raising questions about the robustness, reliability and replicability of current methods. Moreover, our findings highlight a concerning dependency on proprietary LLMs that, even with access to compute resources, were necessary to achieve sufficient accuracy. This introduces new challenges, as silent updates, lack of versioning, and opaque training pipelines pose threats to methodological transparency and long-term reproducibility in LLM-enabled research.