I first ran each post title through gpt-3 to classify as either likely or unlikely to be insightful to founders. Then, if it is likely to be insightful I pass the entire post content including up to 100 comments into gpt 4.

This comes down to around 13-14 million tokens used for the 15k posts, which for the latest gpt-4 model is ~$150.

Im sure this could be done cheaper with something like Llama 2 but tbh I am fine with this for now since I was aiming for speed with the initial list to gauge interest :)