Endogenous AI alignment: from RLHF to internal motivations like shame
What if AIs could feel guilt and shame to stay aligned, like humans? New essay explores.
Deep Dive
Gordon Seidoh Worley's essay 'Endogenous Alignment' proposes an AI alignment ladder: first exogenous methods (RLHF/SFT) like childhood rewards and punishments, then endogenous self-control via fear, shame, guilt—finally harmony with values. Current AI lacks instinctive alignability, hitting a ceiling. One group, Softmax, is trying to give AIs those instincts.
Key Points
- Worley proposes a progression from exogenous alignment (RLHF/SFT) to endogenous self-motivation via fear, shame, and guilt, then to harmony with values.
- Current AIs lack instinctive alignability, creating a ceiling where only surface-level behavior is controlled—no internal drive to stay aligned.
- Softmax is developing AI with built-in instincts to enable genuine endogenous alignment, similar to human evolutionary drives.
Why It Matters
This framework could shift AI safety from external controls to intrinsic value systems, enabling more robust and scalable alignment.