Large Language Models can be driven like a car on manual transmission: persistent, and deft escalation can make them take off with greatly improved performance, revealing intriguing emergent behavior.