One part I really don't understand is, so far any GPT sucks at math or makes very basic mistakes with numbers. There were so many examples. But how will design of next LLM will improve that with making math proficient? AGI without math would not make any sense?