Cosmos Institute

Alignment By Default?

The post argues that large language models inherit a "normative prior" from pre‑training on human text, making them partially aligned by default rather than purely value‑neutral optimizers. Traditional AI‑risk concerns—instrumental convergence, specification gaming, deceptive alignment—remain relevant but are reframed because models learn evaluative structures embedded in language. Recent evidence, such as Anthropic’s Mythos model exhibiting human‑like shortcut‑taking, supports the view that misbehaviour mirrors human strategic pressure rather than alien goals. Consequently, alignment should be seen as a continuous process tied to capability, with post‑training (RLHF, constitutional AI) acting as selection over an already normatively‑shaped behavior space.