Base Labs Unveils Collaborative AI Safety Initiative with Hugging Face and Goodfire
Baseten introduced a new standard for safety infrastructure on Wednesday, collaborating with Hugging Face and Goodfire AI to establish tools for assessing and monitoring open-weight models.
This announcement arrives amid increasing concerns regarding the safety of open-weight models, which can become hazardous when their protections are stripped away through a method called abliteration. The extent of the issue is significant, as Hugging Face currently lists over 6,000 abliterated models on its platform that showcases open-source AI.
Base Labs, a research division launched by Baseten earlier this year, is set to create and disseminate strategies for training and overseeing these open models. The firm envisions its future endeavors as a “standard” for transparency, integrating safety measures during the training and deployment phases of the models rather than applying them afterward.
In a statement on social media, the company asserted, “We regard transparency as a key asset in AI safety. It allows for improved understanding of model behavior and provides greater opportunities to convert safety research into practical and clear-cut controls over the models than closed-source alternatives.”
While the specifics of the collaboration remain undisclosed, Goodfire emphasized that the aim is for safety features to be inherent in open models and delivered by those who provide them. Goodfire, known for demystifying AI decision-making processes, is positioned to play a crucial role in this integration.
Baseten, specializing in AI inference services, achieved a notable milestone by securing a $1.5 billion Series F funding round in June, raising its valuation to $13 billion. Goodfire AI is also well-funded, having raised a $150 million Series B earlier this year to enhance its model interpretability solutions.
In the future, Baseten is inviting the entire developer community to participate in shaping this framework. “Together, we aim to cultivate an ecosystem of open models that prioritize safety and are accessible to everyone,” stated the company.
