Comment by jeswin

At thousands of tokens per second, LLMs (harnesses) can start to do a broader tree search of possibilities even in inefficient token space. This unlocks capabilities outside programming.