Clear explainer of latent-reasoning models that think in continuous space instead of tokens, and why that changes the cost and quality trade-off.