GPT-4's Code Regression: From Prodigy to Buggy
·50 words·1 min
It’s great these authors took time to quantify something a lot of folks had been observing. Authors find that for code generation from GPT-4 now produces directly executable code only 10% of the times in June vs 52% in March! I haven’t observed such huge drop myself though. https://x.com/matei_zaharia/status/1681467961905926144