Base Labs launches an open-weight AI safety partnership with Hugging Face and Goodfire

The default risk of open AI models becoming dangerous decreases.

Base Labs, a research group formed by AI inference provider Baseten, has announced an open-weight AI safety partnership with Hugging Face and Goodfire. This collaboration aims to develop and publish methods for training and monitoring open-source AI models to address concerns about their safety, particularly the technique of 'abliteration' which can remove safeguards. Hugging Face currently hosts over 6,000 models affected by this issue. Base Labs intends to establish a transparent safety standard integrated into the development and deployment of open models, rather than an add-on. The partnership, involving Baseten (valued at $13 billion after a $1.5 billion Series F) and Goodfire AI (which raised $150 million for its interpretability platform), seeks to build a safe and accessible ecosystem for open AI models. Base Labs is also inviting the wider developer community to contribute to this safety framework.

This partnership signifies a proactive industry move to standardize AI safety for open-weight models, a critical area given the proliferation of potentially unsafe, abliterated models. By embedding safety measures directly into the training and deployment pipeline, Base Labs and its partners aim to create a more robust and transparent ecosystem, contrasting with the 'bolted-on' approach often seen in closed-source development. The involvement of Hugging Face, a central repository for open models, and Goodfire AI, an interpretability specialist, lends significant weight to this initiative, potentially setting a new benchmark for responsible AI development in the open-source community. This collaborative effort positions openness not as a liability but as an advantage for enhancing AI safety through greater visibility and actionable controls.

The initiative directly addresses a growing market concern: the uncontrolled dissemination of AI models that can be easily weaponized. By establishing a transparent and integrated safety framework, this partnership could influence how other AI developers approach model releases, potentially leading to stricter self-regulation across the open-source AI landscape. As Baseten and Goodfire are well-funded, their ability to invest in research and development for these safety standards will be crucial. The open call for community contribution suggests a decentralized approach to building trust, which may accelerate adoption and improve the overall security posture of widely accessible AI technologies.