Researchers show how ChatGPT 3.5 performs better than ChatGPT 4.0 in several tasks, often math related, highlighting the issue of "drift" in improving AI models (Josh Zumbrun/Wall Street Journal)

Josh Zumbrun / Wall Street Journal:
Researchers show how ChatGPT 3.5 performs better than ChatGPT 4.0 in several tasks, often math related, highlighting the issue of “drift” in improving AI models  —  AI chatbots have stoked fears that they could spin out of control, but they also suffer from a type of deterioration called ‘drift’