Tim Soret: "If you want to understand - how this is possible - why LLMs are NOT stochastic parrots overfitted on training data - why they actually approximate the underlying mechanics of the data here is the most clear explanation. This is why large language models can understand the world