“Vipul Ved Prakash predicts that AI models will achieve 90% performance on the SWE-bench benchmark "pretty soon."”