How many of the learnings over the past two years went into the prompt that you gave the model? Is it possible that you wouldn't have been able to prompt it so well without those learnings?
Some, but actually surprisingly little. The thing is that my hobby projects are usually extremely "prior art". It's like "make a renderer (just like one that 1000 people have made before)". It's not novel. I just needed a renderer and none was suitable. But the AI knows exactly what to do with almost no context and almost no decisions.