Thinking Machines Lab outlines a framework for safely releasing open-weight AI models, arguing that openness is a public good but must be balanced against real misuse risks. The post covers their safety evaluation process for Inkling and Inkling-Small, including internal harm testing, external red-teaming by four organizations (Scale AI, Handshake AI, FAR.AI, Apollo Research), and adversarial fine-tuning studies. Key ideas include staged releases (API access → monitored availability → open weights), supporting defenders with early and fine-tuning access, and researching whether dangerous capabilities like CBRN knowledge can be decoupled from general intelligence through pretraining data filtering. The lab commits to publishing a more detailed release framework and launching safety grants via their Tinker platform.
