OpenAI seems to have introduced CriticGPT, a new AI model based on GPT-4. Reportedly, OpenAI, the maker of ChatGPT, has designed a new AI model that will help to identify users’ errors in code produced by ChatGPT.
Reportedly, the new AI-model CriticGPT which is in trials, improved code review outcomes by 60 percent when used compared to those who did not.
Decoding CriticGPT
According to sources, OpenAI plans to include CriticGPT into OpenAI’s Reinforcement Learning from Human Feedback (RLHF) labeling pipeline. It looks like OpenAI aims to provide AI trainers with better tools to evaluate complex AI outputs.
The GPT-4 models that power ChatGPT are expected to be helpful and interactive through RLHF. This process involves AI trainers comparing different responses and rating their quality. It is believed that as ChatGPT’s reasoning gets improved, its mistakes become subtler, making it harder for trainers to identify inaccuracies.
According to the research paper titled, ‘LLM Critics Help Catch LLM Bugs,’ CriticGPT showed decent competency when analysing code and identifying errors that help humans to spot hallucinations that they may not notice on their own. The paper also highlighted that the researchers trained CriticGPT on a dataset of code samples with bugs. The bugs are expected to have been inserted on purpose so it could recognise and flag coding errors.
What’s next
According to reports, during the experiments of testing CriticGPT, teams using CriticGPT produced more comprehensive critiques and identified fewer false positives compared to those working alone. “A second trainer preferred the critiques from the Human+CriticGPT team over those from an unassisted reviewer more than 60 percent of the time, as reported by,” ‘LLM Critics Help Catch LLM Bugs.’
However, critics argued that CriticGPT has limitations. It looks like CriticGPT was trained on short ChatGPT answers. So, it needs further development to handle longer, more complex tasks. Also ‘ChatGPT hallucinating’ which is when an AI model generates incorrect information but presents it as if it were a fact, still remains a loophole in CriticGPT.
Also, trainers occasionally make labeling mistakes and in this case the focus on single-point errors needs to expand to address errors spread across multiple parts of an answer. This key limitation can be associated with RLHF. It is believed that these advanced models can become so knowledgeable that human trainers might struggle to add meaningful feedback in CriticGPT.
Follow FE Tech Bytes on Twitter, Instagram, LinkedIn, Facebook
