AI Researcher Rejects Latent Spaces For Neural Symbolic World Models