OpenAI has launched a brand new instrument known as CriticGPT, a brand new synthetic intelligence (AI) giant language mannequin (LLM) designed to help people in recognizing errors in code written by ChatGPT.
In an effort to enhance the accuracy and usefulness of its GPT fashions, OpenAI makes use of a method known as Reinforcement Studying from Human Suggestions (RLHF), the place people assessment and critique output from the AI fashions. CriticGPT is designed to help the AI trainers within the RLHF course of.
According to OpenAI, “When individuals get assist from CriticGPT to assessment ChatGPT code they outperform these with out assist 60% of the time,” in an experiment that confirmed a human AI coach most popular critiques from a group of human-plus-CriticGPT over a human with out the AI instrument 60% of the time.
The most important advantage of pairing a human and an AI to offer suggestions within the RLHF course of is that CriticGPT can write extra complete and thorough critiques than a human working alone, and people can step in when the mannequin is reporting hallucinated errors, or when CriticGPT is reporting too many unhelpful “nitpicks”.
How was CriticGPT skilled?
CriticGPT was additionally skilled utilizing OpenAI’s RLHF methodologies. Trainers took code written by ChatGPT after which inserted errors into the code. They then wrote instance suggestions {that a} mannequin might present if it had discovered the inserted error.
The coach then in contrast output from the brand new mannequin, taking a look at a number of critiques of the modified section of code and seeing when the mannequin had accurately recognized and defined the error inserted.
What are the constraints of CriticGPT?
The coaching information thus far for CriticGPT has targeted on solely brief solutions written by ChatGPT and additional analysis and growth can be required for it to deal with longer and extra complicated outputs. Moreover, it’s not free from the dreaded AI hallucinations that plague LLMs.
Additional work can even must be carried out earlier than the mannequin can determine and critique dispersed errors – it at the moment can solely deal with discrete errors that seem in a single place.
What’s subsequent for CriticGPT?
OpenAI intends to combine CriticGPT into its RLHF pipeline after which go on to scale the work additional and improve its utility.