Anthropic · Claude · Google · China · Mythos · Georgetown CSET
Most published research compresses capabilities from large teachers (e.g
Compiled by KHAO Editorial — aggregated from 1 source. See llms.txt for citation guidance.
◌ Single Source
Distillers must have access to the student model’s weights.
Key facts
- Most published research compresses capabilities from large teachers (e.g., ChatGPT) into much smaller and less capable students (e.g., Llama-2-13B). 10
- The teacher is often accessed over the internet through paid accounts that allow automated programs to connect to the models in structured ways. 11
- Black-box distillation (herein referred to as distillation) is a common technique used by AI developers and researchers
- Black-box distillation is a technique that improves a weaker AI model (student) by using a stronger AI model’s (teacher’s) outputs as a source for training
Summary
This piece overviews how black-box distillation works in practice to inform proposed legislation and ongoing policy debates around adversarial distillation. If you asked Claude Mythos one hundred thousand questions about cybersecurity, would learning the answers make you better at hacking? Conversely, white-box distillation uses additional information about the teacher model’s internal processes, requires access to its weights, and is outside the scope of this explainer. 1. Black-box distillation (herein referred to as distillation) is a common technique used by AI developers and researchers. 2 It has traditionally been used to compress knowledge and capabilities into smaller student models to reduce their compute and memory overhead. In overviewing the distillation process, they highlight four key takeaways regarding Chinese labs’ adversarial distillation campaigns:.