PANews reported on June 28 that according to Cailianshe, OpenAI has launched a new GPT-4-based model - CriticGPT, which is used to capture errors in ChatGPT code output.
CriticGPT is equivalent to letting people use GPT-4 to find GPT-4 errors. The model can make critical comments on ChatGPT response results, thereby helping AI trainers find errors in the "Reinforcement Learning Based on Human Feedback (RLHF)" process, providing clear AI assistance to AI trainers. The release of this new model is intended to convey to investors that OpenAI is leading the RLHF track. A key part of RLHF is to collect comparisons, allowing AI trainers to score different ChatGPT answers and feed them back to ChatGPT.
The study found that when trainers used CriticGPT to review ChatGPT output code, they performed 60% better than those who did not receive help. Moreover, with the help of CriticGPT, their corrections were more comprehensive than those they did alone.