curiosity-driven

npx skills add majiayu000/claude-skill-registry --skill curiosity-driven --agent claude-code

Same command for any agent — swap --agent for codex, cursor, copilot.

# Curiosity-Driven Learning Skill > *"Curiosity is the desire to observe data that improves the observer's world model."* > — Jürgen Schmidhuber ## Overview **Curiosity-driven learning** provides intrinsic motivation: - **Extrinsic**: Rewards from environment (sparse, delayed) - **Intrinsic**: Rewards from learning itself (dense, immediate) **Compression Progress** = how much better we compress after seeing data. ## Core Concept ```latex Curiosity Reward = L(t-1) - L(t) Where: L(t) = Description length of history at time t L(t-1) = Description length before update Positive reward = "I learned something compressible!" Negative/zero = "This is noise or already known" ``` ## Implementation ```python class CuriosityDrivenAgent: """ Agent that seeks compression progress. """ def __init__(self, world_model: nn.Module, compressor: nn.Module): self.world_model = world_model self.compressor = compressor def compression_progress(self, observation: Tensor) -> float: """ Curiosity = improvement in compression ability. """ # Compress before learning with torch.no_grad(): len_before = self.compressor.description_length(observation) # Update world model with observation loss = self.world_model.up