Switching from imitative training of LLMs to self-competitive reinforcement learning allows them to surpass human level in general language-domain tasks.