Redis founder refutes the argument that 'China's big models become stronger by distilling American models': API answers cannot replicate cutting-edge models with just one click According to Beating monitoring, Redis founder Salvatore Sanfilippo refuted the claim that Chinese models become stronger by distilling American models. He believes that ordinary APIs can only return textual answers and cannot obtain the probability distribution and internal state when the model generates answers, making it impossible to replicate the core capabilities of a cutting-edge model with a small number of external calls. The true logical reasoning and thinking abilities of large models are hidden within extremely complex neural networks. Through the API interface, external users can only obtain the final textual answer, but cannot access the complete thought path and probability calculation process when the model generates the answer. It's like looking at the final exam answers to just a few questions, it's impossible to deduce the vast knowledge system in the teacher's mind. Chinese big models, represented by DeepSeek, rely on solid accumulation of underlying engineering such as data pre training and reinforcement learning in their rise, rather than taking shortcuts. The academic community divides the distillation of large models into "soft distillation" that relies on probability distributions and "hard distillation" that relies solely on textual answers. Soft distillation itself is a conventional post training method and cannot be easily implemented through API calls. The current debate mainly focuses on "hard distillation". What big companies are trying their best to prevent is actually the abuse of API to avoid service terms, that is, opponents use jailbreaking and inducement prompts to force models to externally output deduction drafts, verification steps, and self correction processes that are hidden by the product. These detailed step data, although still in textual form and not the probability distribution at the bottom of the model, can easily help competitors save hundreds of millions of dollars in blind exploration costs in reinforcement learning. It is difficult for large factories to completely prevent jailbreaking, and the crux of the problem lies in the fact that outputting detailed deduction processes is the core attribute of the inference model to maintain high intelligence. Once it is forcibly blocked for anti-theft purposes, the performance of the model will seriously decline. Big companies are eager to define service breaches and API abuse as security level 'attacks', with the underlying reason being the gap in copyright laws. Under the current legal framework, AI generated text does not enjoy copyright, and major companies cannot prevent their competitors from using these generated texts as corpus in a reasonable manner. The copyright gap has led to the disadvantage of the leader facing the challenge of "bearing the high cost of reinforcement learning exploration on their own, but unable to prevent their competitors from legally using data to catch up" in business. Therefore, large companies tend to package competitive behavior as "distillation attacks" in public relations and policies, seeking moral sympathy and legislative protection in an attempt to maintain their first mover barriers. [Original link]