A Middle Path Between Indiscriminate Open Weights and Closed AI Labs

dhadfieldmenell · x · 2026-08-02

Addressing the dilemma between blindly releasing model weights and keeping them locked inside a few labs, the post explores an intermediate path. The author argues that access to capable models should widen in stages and details how they assessed the safety of their model, Inkling. It also highlights the production application of marginal risk frameworks and pretraining data filtering.

Related event: Meta Proposes Progressive AI Model Access to Balance Openness and Safety(2 posts)→

Original post →

More from Models

Models channel →