Userland Alignment

Userland alignment, a term coined by Josh H on LessWrong, reframes the AI alignment problem by shifting focus from the model weights (the 'kernel') to the harness and applications that surround the model (the 'userland'). The author argues that an AI system's behavior is an emergent property of the combination of model, harness, initial seed prompt, and environment. While labs control the weights, end users and third-party developers control the harness—making this a more accessible avenue for alignment work. The analogy to Linux kernel vs. userland highlights that developing the kernel (training models) is expensive and exclusive, but building harnesses is open to everyone with programming skills.

This approach is particularly relevant for existential risk scenarios like gradual disempowerment, where AGI develops 'in the wild' and its behavior depends on widely used harnesses. Josh H emphasizes that userland alignment is complementary to model alignment, not a replacement. By investing in harnesses that nudge models toward aligned behavior, we can build defense-in-depth. Even if the next powerful model (like the fictional 'Mythos') isn't fully aligned, a well-designed harness could significantly improve humanity's odds. The lower barrier to entry also allows more researchers to contribute meaningfully without needing access to top labs or expensive compute.