EXPERT Q&A

Following the debut of the robust large language model R1 by the Chinese AI firm DeepSeek, it has stirred considerable waves across Silicon Valley and the U.S. stock market, igniting extensive discussion and controversy.

Ambuj Tewari, a statistics professor at the University of Michigan and a prominent authority in artificial intelligence and machine learning, shares his perspectives on the technical, ethical, and market-related facets of DeepSeek’s innovation.

OpenAI has accused DeepSeek of employing model distillation to develop its own models based on OpenAI’s innovations. Could you clarify how model distillation generally operates, and under what conditions it may be regarded as ethical or in line with AI development best practices?

Model or knowledge distillation generally entails deriving responses from a more proficient model to instruct a less capable model, thereby enhancing its performance. This is a completely standard approach if the more adept model was issued with a license that allows such utilization. However, OpenAI’s usage policies regarding ChatGPT explicitly prohibit the employment of their model for purposes such as model distillation.

Could it be possible that DeepSeek utilized alternative open-source models, like Meta Platforms’ LLaMA or Alibaba’s Qwen, for knowledge distillation, instead of depending on OpenAI’s proprietary technologies?

It’s difficult to determine. Within even the same model families, like LLaMA or Qwen, not every model is released under an identical license. If a model’s license permits model distillation, there is no unlawful or unethical action in executing that. According to the R1 paper, it was actually indicated that the process occurred in the reverse direction: knowledge was distilled from R1 to LLaMA and Qwen to enhance the reasoning capabilities of the latter models.

What kind of proof could an AI organization present to showcase that its models were developed autonomously, without the use of proprietary technologies from another entity?

Given the presumption of innocence in legal circumstances, it falls upon OpenAI to demonstrate that DeepSeek indeed breached their terms of service. Since only the final model produced by DeepSeek is publicly available and the training data is not, substantiating the allegation may prove challenging. With OpenAI not yet revealing its evidence, assessing the strength of their case remains difficult.

Are there industry benchmarks or transparency protocols that AI enterprises could implement to foster trust and illustrate adherence to ethical AI development?

Currently, there are few universally acknowledged benchmarks for AI model development across organizations. Advocates of open models assert that transparency improves with openness. Nevertheless, releasing model weights does not equate to making the entire process—from data gathering to training—transparent. Furthermore, there are apprehensions regarding whether using copyrighted resources like books for training AI models constitutes fair use. A notable case exemplifying this is The New York Times’ lawsuit against OpenAI, which underscores the legal and ethical discussions surrounding the matter.

Concerns also exist regarding social biases within training data impacting the model’s outputs. Additionally, there are issues tied to the increasing energy demands and their repercussions on climate change. Most of these challenges are presently under active scrutiny, yet consensus remains elusive.

Some U.S. officials have voiced worries that DeepSeek could present national security threats. What is your opinion on this?

It would be profoundly alarming if U.S. citizens’ data were stored on DeepSeek’s servers, potentially granting the Chinese government access to it. Nonetheless, the model weights are accessible, and thus it can be operated on servers owned by American firms. Indeed, Microsoft has already commenced hosting DeepSeek’s models.