Why would the deepmind cofounder go for 1 model and not a MOE architecture like gpt4? 1.3 billion dollars is not enough to get you to gpt4?