Redis creator pushes back on 'Chinese AI just distills US models' myth

Redis creator Salvatore Sanfilippo is pushing back on the narrative that Chinese large language models are only catching up by distilling knowledge from American models. His argument: a typical API only returns text answers — not the probability distributions or internal states that make a frontier model tick. You can't replicate a cutting-edge model's core capabilities with a handful of API calls.

A model's true reasoning and thinking ability is buried deep inside its complex neural network. Through an API, all you get is final text output; you don't get the full chain of thought or probability calculations that produced it. It's like looking at a final exam answer sheet — you can't reverse-engineer the teacher's entire knowledge base from a few answers. Chinese models like DeepSeek have risen through solid foundational engineering in data pre-training and reinforcement learning, not shortcuts.

Academics divide model distillation into two types: "soft distillation," which relies on probability distributions, and "hard distillation," which only uses text answers. Soft distillation is a standard post-training technique and can't easily be done via API calls. The current debate centers on hard distillation. What big US companies are really guarding against is API abuse that violates terms of service — rivals using jailbreaks and crafted prompts to force models to output hidden reasoning drafts, verification steps, and self-correction processes. These detailed step-by-step outputs are still just text, not the underlying probability distributions, but they can save competitors hundreds of millions of dollars in blind RL exploration costs. Big companies can't fully prevent jailbreaking because outputting detailed reasoning chains is essential for high intelligence in reasoning models; blocking them to prevent theft would severely degrade performance.

The rush to frame API misuse and terms-of-service violations as "attacks" is driven by a gap in copyright law. Under current law, AI-generated text isn't copyrightable, so leading companies can't stop rivals from legitimately using that text as training data. That leaves them in a commercial bind: they bear the high cost of RL exploration but can't stop competitors from legally using the data to catch up. So they tend to frame competitive practices as "distillation attacks" in public relations and policy, seeking moral sympathy and legal protection to defend their first-mover advantage.