Sparks of AGI prompts on GPT2XL and its variant, RLLMv3 — LessWrong