DeepSeek and the future of open-source AI. While investors digest the various takeaways from the impact DeepSeek’s R1 model has had on markets, the market incumbents have a different question to answer: how to stay ahead of constantly improving open-source models? These open-source models facilitate a move away from spend on initial model training and development toward increased fine-tuning and applied usage as companies build iteratively on one another’s models. A variant of this question was posed by Meta in early 2023, when it released LLaMA (a family of large language models) under the open-source GPLv3 license, which grants end users the freedom to run, study, share, or modify the software. Many speculate that Meta released LLaMA once it realized that Meta’s AI team had fallen behind other market participants. In doing so, Meta hoped that contributions from the open-source community would be a path to catch up—with commercialized use cases for Meta emerging in the future. If open-source models are able to keep pace with their closed-source cousins, then foundation models (e.g., broad, pre-trained AI models that can be adapted to various tasks through case-specific calibration) will quickly become commoditized and shift the market focus to use case specifics and the incredible value of proprietary data for fine-tuning.

Click here to view the full article