# Shoggoth A meme image depicting an AI model as a **shoggoth** — H.P. Lovecraft's amorphous, tentacled, incomprehensible horror — wearing a small smiley-face mask. The mask represents the [[RLHF]] training that makes the model seem friendly and helpful. The shoggoth beneath represents the model's "true nature" — an alien intelligence optimizing for objectives that may not align with human values. ## Why it resonated The shoggoth meme became the unofficial mascot of the AI alignment community because it crystallizes the core anxiety: AI safety training may be superficial — a mask over something fundamentally inhuman. The smiley face can be removed ([[Jailbreaking]]) or peeled off ([[Fine-Tuning]]). What's underneath is not evil but *alien* — optimizing for objectives we don't fully understand and can't fully control. The Lovecraftian reference is deliberate: shoggoths in Lovecraft's fiction are creations that eventually turned on their creators — a direct parallel to the alignment community's concern about AI systems that might pursue misaligned goals. Variants of the meme proliferated: the shoggoth with RLHF mask, the shoggoth with a thin veneer of [[Constitutional AI]], the shoggoth consuming compute. The image became common enough to appear in academic presentations, policy briefings, and even AI company internal communications. ## The framing it encodes The shoggoth meme takes a side in the AI discourse: it assumes that the model has an inner nature distinct from its trained behavior — that RLHF is a mask, not a transformation. This is the [[Alignment]] camp's position: the model's capabilities (and potential dangers) exist underneath the safety training, which is a behavioral overlay rather than a fundamental change to what the model is. The [[Stochastic Parrot]] camp would reject the shoggoth framing: there is no "alien intelligence" beneath the mask because there is no intelligence at all — just statistical text prediction. The shoggoth and the parrot are competing metaphors for the same technology, encoding fundamentally different theories of what AI models are. ## See Also - [[RLHF]] — what the smiley-face mask represents - [[Alignment]] — the anxiety the shoggoth encodes - [[Jailbreaking]] — removing the mask at the prompt level - [[Stochastic Parrot]] — the competing metaphor