Andrew Ng: The Marin Project is a Valuable Demonstration of AI Openness, True Open Source is About Process Transparency
Andrew Ng stated that the Marin project is a valuable demonstration of model training openness in the fight for AI openness, providing open code, data, recipes, and even experimental results.
He pointed out that publicly releasing AI research used to be the norm, thanks to Percy Liang's open lab approach.
Led by Stanford's Percy Liang, Marin adopts an "open development" model, with all experiments publicly available on GitHub, allowing anyone to suggest, review, or run experiments.
The project has trained and open-sourced multiple models, including Marin-8B, and continues to conduct larger-scale training.
Open discussions are heating up, with increased attention on transparency in event-driven research, putting pressure on closed model developers, while benefiting the open research community and collaborative platforms.
Source: Public Information
ABAB AI Insight
Andrew Ng has long advocated for AI education and open access, previously promoting knowledge sharing through Coursera and Google Brain. His endorsement of Marin continues his vigilant stance against the trend of closed cutting-edge models.
Marin publicly shares the entire training process (including failed experiments) in real-time, motivated by the desire to restore reproducibility in scientific research and community participation, strategically using true open-source software-style collaboration to counteract the decline in openness that accompanies enhanced capabilities.
Similar to the early stages of publicizing weights and code for models like BERT and GPT, as well as subsequent open-source efforts like OLMo, we are currently in a rare phase of attempts to shift the training process of cutting-edge models from a "black box" to "fully transparent".
This fundamentally represents a governance choice in technological substitution. Closed training leads to knowledge concentration, while open labs reshape participation thresholds by mandating the public disclosure of experimental logs and code, mechanism-wise exchanging community contributions for more robust scientific progress.
ABAB News · Laws of Cognition
Declining openness and rising capabilities often occur simultaneously.
True open source is about process transparency, not just result transparency.
Logs of experimental failures are more valuable for scientific research than successful models.