跪拜 Guibai
← All articles
Frontend · AI Programming · Artificial Intelligence

Redis Creator: Chinese AI Models Aren't Just Distilling US APIs

By 恋猫de小郭 ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

The distillation narrative doubles as a marketing claim from US labs and a policy talking point about export controls. Understanding the actual limits of API-based imitation clarifies that compute access, not data leakage, is the primary constraint on non-US model development.

Summary

The claim that Chinese LLMs owe their strength to distilling US models via API access is technically incoherent, according to antirez. True soft distillation requires access to a teacher model's internal probability distributions and reasoning traces — data that commercial APIs never expose. What API-based hard distillation can do is tune output style and fill narrow knowledge gaps, but it cannot replicate the general reasoning power that comes from trillion-token pretraining and massive compute investment.

Even with full white-box access to open-weight frontier models, labs in Europe and elsewhere still struggle to match their performance, which undercuts the idea that distillation is a shortcut. The real bottleneck is a compute deficit, not a copying playbook. The debate itself is muddied by terminology: what many people casually call "distillation" is actually response imitation via SFT, a far weaker signal than learning from a teacher's full probability distribution.

In practice, hard distillation is widely used and can be done intelligently — filtering synthetic data, using strong models as verifiers or reward models, and building data flywheels — but that is engineering sophistication, not simple theft. No amount of API-generated text alone produces a DeepSeek-class model.

Takeaways
Soft distillation requires access to a teacher model's full logits and internal representations, which commercial APIs do not provide.
Hard distillation using only API-generated text outputs can improve style and format adherence but cannot create frontier-level general capabilities.
Even with full access to open-weight frontier models, many well-resourced labs still fail to replicate their performance.
The primary gap between Chinese and US models is a compute deficit driven by export restrictions, not a reliance on distillation.
Apple's Foundation Models used a distillation-based refinement pipeline with Gemini outputs, a form of hard distillation.
Smart distillation treats strong models as tools in a data flywheel — filtering, verifying, and scoring — rather than as dumb data printers.
Conclusions

The conflation of soft and hard distillation under one term makes the public debate nearly useless; one side argues about copying probability distributions while the other argues about copying text outputs, and they talk past each other.

Antirez's argument is strongest when applied to pretraining-scale capability, but the line blurs in post-training, where API-generated data is already used commercially by companies like Apple.

The claim that distillation cannot produce frontier models is empirically true, but the counter-claim that it is useless is overstated — hard distillation is a standard post-training tool, just not a substitute for pretraining.

Yao Shunyu's distinction between crude and smart distillation reframes the ethics question: the problem isn't using a teacher model, it's using it without a strategy, which signals a team with no research direction.

Concepts & terms
Soft Distillation (White-box)
A training method where a student model learns to mimic a teacher's full probability distribution (logits) using KL divergence, capturing 'dark knowledge' about relative token likelihoods rather than just the final output text.
Hard Distillation (Black-box / Response Distillation)
A training method where a student model is fine-tuned only on the discrete text outputs generated by a teacher model, using standard cross-entropy loss. It transfers style and format but little of the teacher's internal reasoning.
Dark Knowledge
The information contained in a teacher model's soft probability distribution — why it assigns 70% probability to one token and 20% to another — which helps a student model generalize better than training on hard labels alone.
Compute Deficit
A structural disadvantage in available GPU/TPU computing power, often caused by export controls, that limits a lab's ability to run large-scale pretraining runs regardless of algorithmic sophistication.
From the discussion

A translation dispute over the article's title dominated the thread. Several comments argued the original Chinese rendering was correct, citing the English negative-transfer structure 'not because' and the surrounding context. Beyond the grammar debate, the core argument surfaced: Chinese AI progress stems from independent investment and engineering, not API distillation, with computing-power restrictions as the real differentiator. A smaller, cynical strain dismissed the entire framing as a recycled Western narrative of Chinese copying.

A translation of the title as 'not because' is grammatically correct, reflecting an English negative-transfer structure where 'not' modifies the reason clause, not the main verb.
White-box distillation of closed-source US models is practically impossible, and black-box distillation cannot replicate the deep foundational capabilities required for frontier models.
The performance gap between Chinese and US AI models is primarily driven by restricted access to computing power, not by distillation techniques.
The term 'distillation' is loosely used in public discourse, conflating soft distillation (transferring knowledge) and hard distillation (training on outputs), which obscures the real technical picture.
The 'China copy' accusation is a decades-old rhetorical pattern, and the current AI distillation claim is just its latest iteration.
Featured comments
小怼子 5 likes

Summary: Redis author antirez refutes the claim that 'Chinese models are strong because of API distillation from US models.' He argues that white-box distillation of closed-source projects is basically impossible, black-box distillation has limited effect, and it cannot create strong foundational capabilities. The gap between Chinese and US models stems more from restrictions on computing power access. He also points out that the concept of 'distillation' is abused, divided into soft distillation and hard distillation, which people often don't distinguish in daily talk. He believes the progress of Chinese models comes from their own investment; pure hard distillation cannot produce a powerful model.

redbuck 2 likes

In the AI era, it's just a rehash of 'China copy.' For over a decade, white people have been using this same rhetoric.

See top comments, translated →
Source: juejin.cn ↗ Google Translate ↗ Backup ↗