CriticGPT: new tool can help fix ChatGPT errors

OpenAI trained a new model, the CriticGPTbased on GPT-4 for find errors in responses generated by ChatGPT.

According to the company, when people enlist the help of CriticGPT to review codes generated by ChatGPT, they are able to overcome potential problems without additional help 60% of the time

The series of GPT-4 models that power ChatGPT are corrected through what is called “reinforcement learning from human feedback” (RLHF), which means that the platform improves its performance as users, too. called artificial intelligence trainers, they classify the answers given by ChatGPT and indicate possible errors.

As ChatGPT improves, its errors will also become more subtle and specialized, which can make it difficult for AI trainers to notice inaccuracies when they occur. And that’s where CriticGPT can help.

The next step, according to the company, is to incorporate models similar to CriticGPT into the RLHF process, so that human feedback also has the help of AI feedback.

While CriticGPT’s suggestions aren’t always correct either, they can help AI trainers spot issues that would otherwise go unnoticed. Furthermore, just like the GPT-4 model itself, CriticGPT will also improve as more users use the tool and provide feedback.

According to OpenAI, tests between the two models revealed that the notes made by CriticGPT were considered better than those made by ChatGPT itself 63% of the time. Furthermore, the new tool also demonstrated fewer hallucinations — moments in which the AI ​​invents something that is not true and presents it as if it were.

Even so, CriticGPT still has some limitations and can only help up to a certain point of specialization. If a task or answer is extremely complex, even an expert with the help of a model may not be able to evaluate it correctly.