OpenAI's Wudang ladder cloud leap, left foot on right foot: launch a new model CriticGPT to correct AI errors

#人工智能OpenAI's Wudang ladder cloud vertical, left foot on right foot: launched a new model CriticGPT, which is specifically used to find and correct illusions and errors in GPT, replacing the manual review team to mark answers and improve the model's accuracy. Of course, it does not completely replace the manual review team, but uses AI's rich knowledge to find subtle errors and improve efficiency. View the full text: https://ourl.co/104710

Artificial intelligence company OpenAI announced today the launch of a new model CriticGPT based on GPT-4. This model is not open because its purpose is to help OpenAI find errors in GPT.

At present, most AI companies have manual review teams to check the output results of the model, because the model will always have hallucinations and errors in the long run, which requires a manual review team to mark and correct it.

However, as products like ChatGPT become more and more accurate, errors are difficult to detect even when they occur, which makes it difficult for AI engineers responsible for training and human teams responsible for review to cope with them, such as outputting programming code errors.

OpenAI says this is a fundamental limitation of RLHF, that as technology develops and model knowledge and capabilities gradually surpass humans, it will become increasingly difficult to align models.

To this end, the company trained CriticGPT based on the GPT-4 series of models. This new model was also trained using RLHF and is responsible for analyzing and evaluating output results that contain a large number of errors.

In a sense, what OpenAI is doing is like using artificial intelligence to correct artificial intelligence by stepping on the left foot. Of course, since Tiyunzong is also an artificial intelligence, there will be errors, but compared with the manual review team, CriticGPT has rich knowledge and is more efficient, and can find errors faster and more accurately.

It is worth noting that CriticGPT will also judge the erroneous content deliberately provided by humans. A typical case is how many stones humans need to eat every day. CriticGPT will be able to identify such intentional erroneous content and correct it to prevent GPT from directly giving wrong answers.

Finally, OpenAI also emphasized that errors in the real world may be distributed among many answers, which is a problem that needs to be solved in the future. In other words, at least at this stage, it is impossible to completely eradicate the errors and illusions of AI models.

Limited time event recommendations:Kaisou AI intelligent search, free and ad-free, direct to results、The all-round player VidHub supports mounting network disks for cloud playback、Alibaba Cloud Server 36 yuan/year。