OpenAI has launched a new tool called CriticGPT, a new artificial intelligence (AI) large language model (LLM) designed to assist humans in spotting mistakes in code written by ChatGPT.
In an effort to improve the accuracy and usefulness of its GPT models, OpenAI uses a technique called Reinforcement Learning from Human Feedback (RLHF), where humans review and critique output from the AI models. CriticGPT is designed to aid the AI trainers in the RLHF process.
According to OpenAI, “When people get help from CriticGPT to review ChatGPT code they outperform those without help 60% of the time,” in an experiment that showed a human AI trainer preferred critiques from a team of human-plus-CriticGPT over a human without the AI tool 60% of the time.
The major benefit of pairing a human and an AI to provide feedback in the RLHF process is that CriticGPT can write more comprehensive and thorough critiques than a human operating alone, and humans can step in when the model is reporting hallucinated errors, or when CriticGPT is reporting too many unhelpful “nitpicks”.
How was CriticGPT trained?
CriticGPT was also trained using OpenAI’s RLHF methodologies. Trainers took code written by ChatGPT and then inserted errors into the code. They then wrote example feedback that a model could provide if it had found the inserted error.
The trainer then compared output from the new model, looking at multiple critiques of the modified segment of code and seeing when the model had correctly identified and explained the error inserted.
What are the limitations of CriticGPT?
The training data so far for CriticGPT has focused on only short answers written by ChatGPT and further research and development will be required for it to address longer and more complex outputs. Additionally, it is not free from the dreaded AI hallucinations that plague LLMs.
Further work will also need to be carried out before the model can identify and critique dispersed errors – it currently can only handle discrete errors that appear in one place.
What’s next for CriticGPT?
OpenAI intends to integrate CriticGPT into its RLHF pipeline and then go on to scale the work further and increase its utility.
The post OpenAI launches CriticGPT to catch ChatGPT errors appeared first on ReadWrite.
