Redis creator rebuts the argument that "Chinese large models are strengthened by distilling American models": API answers cannot one-click replicate cutting-edge models

According to Dongcha Beating's monitoring, Redis creator Salvatore Sanfilippo has refuted the claim that Chinese large models rely on distilling American models to become stronger. He believes that a regular API can only return a text answer, without accessing the probability distribution and internal state of the model when generating an answer, making it impossible to replicate a cutting-edge model's core capability through a small number of external calls.

The true logical reasoning and thinking ability of large models are hidden in extremely complex neural networks. Through an API interface, external users can only receive the final text answer, unable to access the complete thinking path and probability calculation process the model undergoes when generating an answer. This is akin to only seeing the answers to a few exam questions, unable to reverse-engineer the vast knowledge system in the teacher's mind. Chinese large models like DeepSeek, currently on the rise, rely on solid groundwork in data pretraining and reinforcement learning, rather than taking shortcuts.

The academic community has further divided large model distillation into "soft distillation," which relies on probability distributions, and "hard distillation," which relies only on text answers. Soft distillation itself is a common post-training technique that cannot be easily achieved through API calls. The current debate mainly focuses on "hard distillation." What major companies are striving to prevent is the misuse of APIs to circumvent service terms, whereby adversaries leverage jailbreaking and prompting to force the model to externally output the derivation sketches, validation steps, and self-correction processes that the product intended to keep hidden. While this detailed step data remains in text form and is not the model's underlying probability distribution, it significantly helps competitors avoid billions of dollars in blind exploration costs in reinforcement learning. Major companies find it challenging to completely block jailbreaking because outputting detailed derivation processes is a core feature that enables reasoning models to maintain high intelligence. If this is forcefully blocked for anti-theft purposes, the model's performance will severely degrade.

Major companies are quick to define service breaches and API misuse as "attacks" on a security level, with the underlying motive stemming from gaps in copyright law. Under the current legal framework, AI-generated text is not protected by copyright, preventing major companies from stopping adversaries from legitimately using this generated text as corpus data. This copyright gap leaves industry leaders vulnerable to the "burden of bearing the high cost of reinforcement learning exploration themselves while being unable to prevent competitors from legally leveraging data for catch-up." Therefore, major companies tend to frame competitive actions as "distillation attacks" in public relations and policy efforts, seeking ethical sympathy and legislative protection to uphold their competitive advantage.