“Macrocosmos plans to train models larger than 100 billion parameters within the next 12 to 18 months.”