“With full hindsight, a small team of talented people could have developed a model at the level of GPT-3.5 as early as 2018 or 2019.”