AI researchers at OpenAI introduced CriticGPT last week. This new AI tool helps find errors in computer code written by OpenAI’s chatbot, ChatGPT. The goal is to make AI systems more reliable and aligned with human goals.
Humans train AI systems by giving them feedback on their outputs. This is called “reinforcement learning from human feedback” (RLHF). CriticGPT acts as an assistant to these human trainers, specifically focusing on code generated by ChatGPT. CriticGPT analyzes the code and points out potential mistakes, making it easier for humans to catch errors.
To train CriticGPT, researchers fed it a bunch of code with mistakes purposefully added in. This way, CriticGPT learned to recognize different coding errors and flag them for human reviewers.
Researchers tested CriticGPT against human reviewers and found it identified errors in code 63% better when dealing with real-world mistakes made by large language models (LLMs). Teams that combined human reviewers with CriticGPT wrote more detailed reports on the errors, and CriticGPT even helped reduce the creation of false errors compared to AI-only reviews.
While promising, CriticGPT has some limitations. It was trained on short pieces of code, which might not translate well to longer, more complex tasks future AI systems will face. CriticGPT reduces but doesn’t eliminate false errors and human reviewers can still be misled by these mistakes.
The researchers admit that CriticGPT works best when errors are confined to a single spot in the code. Real-world errors in AI outputs can be scattered throughout the answer, making them a challenge for future versions of CriticGPT.
To improve how humans train large language models (LLMs), OpenAI wants to add CriticGPT-like AI helpers to their review process. These helpers would point out potential errors in code written by LLMs, making it easier for humans to identify mistakes. This is a step towards creating better tools for judging LLM outputs, especially when those outputs are complex and hard for humans to understand.
Follow ProPakistani on Google News & scroll through your favourite content faster!
Support independent journalism
If you want to join us in our mission to share independent, global journalism to the world, we’d love to have you on our side. If you can, please support us on a monthly basis. It takes less than a minute to set up, and you can rest assured that you’re making a big impact every single month in support of open, independent journalism. Thank you.
