@yunnanchen.bsky.social and I have a fun ODI Global paper that will be coming out this spring looking at cofinancing networks for “green” Belt and Road investments.
We used DeepSeek to classify 18,000 AidData Chinese loan records according to impact on energy transition
Paper will be out this spring. imho, using LLMs allowed us to uncover policy-relevant findings about who cofinances green projects (policy-driven lenders, not profit maximizing lenders) that would not be feasible to uncover (in terms of $ + time) using tradition tools.
To me the key question will be building workable validation best practices.
For our project, the API cost for running 18k projects over 15 hours came to a grand total of less than $2. My back of the envelope math on our cost using human coders would be $20k+
Comments
We used DeepSeek to classify 18,000 AidData Chinese loan records according to impact on energy transition
https://bsky.app/profile/tealemery.bsky.social/post/3lfl6yuua522r
We tested competing models against each other and against a substantial sample of human-coded outputs.
We will be making our GitHub repo public so that hopefully others won’t need to reinvent the wheel.
For our project, the API cost for running 18k projects over 15 hours came to a grand total of less than $2. My back of the envelope math on our cost using human coders would be $20k+
https://bsky.app/profile/tealemery.bsky.social/post/3lfl7nsjhic2r