This Waterfox post responds to Mozilla’s “AI-first” browser vision with a more skeptical framing: the browser is uniquely sensitive software, and giving a general-purpose model deep access to your browsing context is a very different bet than shipping a narrow, constrained ML feature.
The author draws a line between single-purpose, locally-run tools (like translation models) that have clear, auditable scope, and large language models, which are treated as hard-to-audit “black boxes” whose behavior and data handling are difficult for users to verify. The takeaway isn’t “no ML ever,” but “be precise about what kind,” and ship it in a way that preserves user trust: tight permissions, explicit boundaries, and a clear story for privacy and control.