x.com/antirez/status/2052138269799018575
1 correction found
you can distill via API calls... which is a Machine Learning absurdity as the signal is not there.
This is incorrect: black-box distillation/model extraction from API outputs has been demonstrated repeatedly in the literature, including for large language models. API access may limit efficiency or fidelity, but the signal is plainly sufficient for distillation in many settings.
Full reasoning
Multiple peer-reviewed and archival papers directly contradict the claim that distilling through API calls is a "Machine Learning absurdity" because "the signal is not there."
- ICML 2021: Zero-Shot Knowledge Distillation from a Decision-Based Black-Box Model explicitly proposes black-box knowledge distillation where the teacher's parameters are inaccessible and the teacher only returns classes rather than softmax outputs. The paper reports that these methods are effective even in a more challenging setting where the training data are also unavailable.
- 2024 LLM work: Knowledge Distillation of Black-Box Large Language Models states that recent research has focused on improving smaller models via knowledge distillation from powerful yet black-box teachers such as GPT-4, and introduces a method that improves such distillation.
- Model-extraction literature: Marich studies black-box model extraction attacks through a predictive API and reports extracted models reaching roughly 60–95% of the target model's accuracy. Separately, Stealing Part of a Production Language Model demonstrates extraction of precise, nontrivial information from black-box production language models using typical API access.
These results do not imply that API-only distillation is always easy, cheap, or equivalent to having full model weights. But they do show that the underlying claim here is too strong: API outputs can contain enough information for distillation or partial model extraction, so calling the idea an "absurdity" because the signal is absent is contradicted by published evidence.
4 sources
- Zero-Shot Knowledge Distillation from a Decision-Based Black-Box Model
We propose the concept of decision-based black-box (DB3) knowledge distillation, with which the student is trained by distilling the knowledge from a black-box teacher (parameters are not accessible) that only returns classes rather than softmax outputs... experiment results demonstrate their effectiveness.
- Knowledge Distillation of Black-Box Large Language Models
Recent research has increasingly focused on boosting the capabilities of smaller models through knowledge distillation (KD) from these powerful yet black-box teachers... we introduce Proxy-KD... a compelling new avenue for distilling knowledge from advanced LLMs.
- Marich: A Query-efficient Distributionally Equivalent Model Extraction Attack using Public Data
We study design of black-box model extraction attacks that can send minimal number of queries... through a predictive API... Marich extracts models that achieve ~60-95% of true model's accuracy.
- Stealing Part of a Production Language Model
We introduce the first model-stealing attack that extracts precise, nontrivial information from black-box production language models like OpenAI's ChatGPT or Google's PaLM-2... given typical API access.