DeepSeek, Moonshot used OpenAI's Astra to train AI models: Report
What's the story
Chinese artificial intelligence (AI) developers DeepSeek and Moonshot AI have been using OpenAI's powerful GPT-6 Astra model to train their latest models, The Information reported. The practice, known as distillation, has become a contentious issue among leading AI companies amid the escalating US-China technology competition. They argue that this method allows competitors to replicate sophisticated capabilities more quickly and at a lower cost.
Model development
OpenAI traced distillation campaign to Moonshot AI
The distillation of Astra has reportedly been instrumental in the development of Moonshot AI's Kimi K3.1 Pro model, which was launched last week.
DeepSeek's upgraded V4.1-Flash 1017 model, launched on October 17, also leveraged this technique.
In a blog post dated September 30, OpenAI revealed that it had traced a major part of a distillation campaign launched in July to people connected with Moonshot AI.
Attack prevention
Anthropic implements measures to protect Claude AI models
Earlier this month, Anthropic announced that new measures to prevent rivals from replicating its Claude AI models have been successful.
The data harvested in so-called distillation attacks has been reduced by about 90% to 99%.
This development highlights the ongoing battle between AI companies as they try to protect their advanced capabilities from being quickly and cheaply reproduced by competitors.