“The first version of Fractal Brain's language model was trained using approximately 0.1% of the electricity that OpenAI used for GPT-1 and GPT-2.”