Woofun AI reports that Redis creator Salvatore Sanfilippo has refuted assertions that Chinese large language models derive their strength from distilling American counterparts. He contends that standard API responses provide only final text outputs, withholding the critical probability distributions and internal states necessary to replicate a model's core reasoning capabilities. This limitation prevents external users from reverse-engineering the complex neural networks that underpin advanced AI systems, akin to seeing exam answers without understanding the teacher's knowledge base.

The debate centers on 'hard distillation,' where competitors attempt to extract detailed derivation steps through jailbreaking and prompting techniques. While this data aids in avoiding costly reinforcement learning exploration, it does not constitute the model's underlying architecture. Sanfilippo notes that major firms frame these actions as security attacks due to copyright gaps, as AI-generated text lacks legal protection, leaving them unable to prevent rivals from using such data for competitive catch-up.