Is DeepSeek stealing from OpenAI | Lex Fridman Podcast

Summary

The transcript discusses the ease with which Chinese companies can use hosted model APIs from US-based companies like OpenAI, addressing ethical and legal concerns surrounding the distillation process. It explores the complexities of using OpenAI's models for training, including the grey area of whether it's considered a competitor. The conversation touches on issues of intellectual property, training on publicly available data, and the challenges of industrial espionage in AI. The discussion also highlights the importance of ideas and human resource movements in the tech industry, including the potential for espionage and ethical dilemmas related to AI development.

Takeaways

• 😀 Chinese companies can easily use hosted model APIs from the US, as seen with DeepSeek using OpenAI's API for training their model.
• 😀 Distillation is a common practice in AI, where outputs from a powerful model are used to train a new model, improving efficiency and performance.
• 😀 OpenAI has rules in place that prevent companies from using its model outputs to train a competitor, but the term 'competitor' remains vague and unclear.
• 😀 Ethical and legal debates arise when training models on outputs from other companies' models or the internet without permission, though these activities are common in the industry.
• 😀 AI models may still reference OpenAI, even when not directly trained on OpenAI data, because of internet-scraped content or pre-existing outputs used in training.
• 😀 There is a grey area in terms of the ethics and legality of training models using outputs from OpenAI or other AI sources, with different companies using various methods to avoid breach of terms.
• 😀 AI startups often use outputs from OpenAI's API for initial model training and can avoid being banned, although some companies, like ByteDance, have faced bans in the past.
• 😀 Distillation processes are also common in the AI community, where models like DeepSeek are distilled into Llama models, which are easier to serve and more widely adopted.
• 😀 It is unclear whether training on outputs from OpenAI is truly illegal or unethical, as the concept of 'competitor' is subject to interpretation, and such practices are common in the industry.
• 😀 Japan has a law allowing AI companies to train models on any data, which sidesteps copyright issues, making it a potential solution to avoid legal restrictions in training AI models.

Q & A

How easy is it for Chinese companies to use hosted model APIs from the US?

It is relatively easy for Chinese companies to use hosted model APIs from the US. OpenAI, for example, has publicly stated that DeepSeek, a Chinese company, uses their API, and there are also claims that outputs from OpenAI’s models can be used for further training by others.

What is the distillation process in machine learning models?

Distillation is a process where a smaller model is trained to mimic the behavior of a more powerful model. This allows researchers to leverage a larger, more capable model for training without directly using the original model in a product. It's a common practice in the AI industry, including for training language models.

What are the ethical implications of training AI models on data generated by other models?

The ethics of training AI models on outputs from other models are debated. Some argue it's hypocritical since companies like OpenAI train on publicly available internet data without explicit permission. However, the process of distillation or generating data from one model and using it to train another raises questions about intellectual property and fairness.

What legal issues arise from training models on data from OpenAI or other companies?

The key legal issue revolves around OpenAI's terms of service, which prohibit building a competitor using their model’s outputs. However, terms of service are different from licenses, and there are potential loopholes that allow companies to train models on OpenAI-generated data indirectly, making enforcement complex.

Why do some models claim to be trained by OpenAI even if they are not?

Some models claim to be trained by OpenAI because their outputs may contain traces of OpenAI-generated content. This can happen when training data includes outputs from OpenAI's models that have been copied or reused, making it difficult to fully distinguish between models.

What is the significance of terms of service versus a formal license agreement?

Terms of service are rules that govern the use of a platform or service, while a formal license agreement is a contract specifying how copyrighted content or intellectual property can be used. Violating terms of service can result in account termination, but it is not the same as violating a legally binding license.

Is it illegal or unethical to use data generated by OpenAI's models for further training?

It is not definitively illegal, though it could break contractual terms of service. Many consider it ethically complicated, especially since OpenAI and others have trained on data from the internet without explicit permission. The gray area lies in the use of model outputs versus raw internet data.

How do some companies get away with using OpenAI-generated data for training without facing consequences?

Companies often circumvent direct violations by generating data using OpenAI models and then using that data for training purposes, creating a 'broken link' that avoids legal ramifications. This allows them to use outputs without directly violating OpenAI's terms of service.

What are the broader implications of restricting training on internet data?

Restricting the ability to train on internet data would significantly hinder the development of AI models, as most models today rely on large datasets scraped from the web. While there are concerns from creators and authors about their content being used without permission, some argue that unrestricted training on internet data is crucial for advancing AI.

How do espionage and industrial secrets play into the development of AI models?

Espionage in the AI industry is not uncommon. Companies often try to steal ideas, either through direct theft or by hiring top employees who bring ideas with them. While stealing actual code and data can be difficult, the transfer of ideas, especially through employee movement, is a significant factor in the competitive AI landscape.