According to @CNBC, Apple is in talks with PrismML to compress large models for on device iPhone inference, cutting costs and latency for AI apps.

According to Jeff Dean, the steep drops observed in model performance graphs are likely due to AI model distillation, a process in which smaller student models are trained to replicate the capabilities of larger, more expensive teacher models. This trend demonstrates that distillation can significantly reduce computational costs and model size while maintaining high accuracy, making advanced AI more accessible for enterprises seeking to deploy efficient machine learning solutions. As cited by Jeff Dean on Twitter, this development opens new business opportunities for organizations aiming to scale AI applications without prohibitive infrastructure investments (source: Jeff Dean on Twitter, December 17, 2025).