The rise of distillation has accelerated a "paradigm shift" towards a new open-source order, where transparency and accessibility are believed to drive innovation faster than closed-door research. As one expert joked, "everybody's model is open source. They just don't know it yet... because it's so easy to distill them".
DeepSeek itself open-sourced its V3 and R1 models, providing a blueprint for anyone to download, use, and modify for free. University and Hugging Face models followed suit, putting pressure on OpenAI's once-secure closed-source strategy, which now looks "more like a target".
This shift has been acknowledged even by OpenAI's CEO, Sam Altman, who notably stated on Reddit that OpenAI has been "on the wrong side of history" regarding its open-source strategy and needs to "figure out a different open source strategy". This is a remarkable admission from a leader who previously championed the closed-source approach for safety, competitive dynamics, and monetisation.
The benefits of this open-source momentum are significant:
- Cost Reduction: Distilled models can be created cheaply and are often given away for free by universities and startups. This "keeps the proprietary players in check from a pricing and performance standpoint".
- Developer Empowerment: Developers can run open-source models on their own infrastructure, reducing the pricing power of proprietary providers. For example, DeepSeek R1 costs $2.19 per million tokens, while OpenAI's comparable O1 costs $60. This drastic cost difference means the "script has now flipped" for AI application makers, giving them a significant advantage.
- Increased Innovation & Use Cases: Making AI more efficient leads to "a dramatic increase in more use cases". It's a win for developers, those building at the AI application layer, and the long-term AI ecosystem due to the "deflationary effect of the cost of these models".
- Commoditisation of Models: AI model builders who were on top are becoming "more and more commoditized". The market is transitioning from a few dominant model builders to "many, many more" comparable models at much lower training costs.