OpenAI has recently introduced CriticGPT, a new AI tool based on GPT 4, to spot errors in code produced by ChatGPT. This latest innovation is a part of an effort to improve AI systems to give accurate answers. According to the company, CriticGPT improves code review outcomes by 60 per cent as compared to those who did not. OpenAI will soon integrate CriticGPT into its Reinforcement Learning from Human Feedback (RLHF) labelling pipeline. This step will provide AI trainers with better tools to evaluate AI outputs, no matter how complex.The company shared its research paper that states that the aim of CriticGPT is to help human AI reviewers in the task of reviewing code generated by ChatGPT. CriticGPT was built using the GPT-4 LLM, and showed encouraging competency when analysing code and identifying errors, enabling its human colleagues to spot AI “hallucinations” that they may not notice by themselves.OpenAI explained that for the training process, human developers were asked to edit code written by ChatGPT, introducing a range of errors and providing sample feedback. This taught CriticGPT how to identify the most common, and also some not-so-common coding errors.
After the training was done, OpenAI stated that the results were impressive. However, one of the challenges that CriticGPT needs to overcome is that some AI hallucinations are the result of errors spread across multiple different code strings, which makes it much more difficult for CriticGPT to identify the source of the problem.
This tool is one of many new tools that are being developed to improve large language models and make their generative AI even more capable.
The new tech might also help the upcoming AI models developed by OpenAI. The company has recently shared that they are working on their new and more powerful AI model. OpenAI CTO Mira Murati recently shed light on the upcoming AI model and its intelligence.
Murati says that GPT-3 had toddler-level intelligence, GPT-4 had the intelligence of a high-schooler, and the next generation of the model will have the intelligence of someone with a PhD for specific tasks. She also revealed that the next generation of GPT can be expected in “a year and a half”. She even said that when you have a conversation with the next-gen chatbot, in a lot of things you might feel like it’s smarter than you.
To make its new generative models smarter, OpenAI has partnered up with Time Magazine. The artificial intelligence research company has joined hands with Time magazine to enhance and train ChatGPT. On Thursday, the companies announced a multi-year content deal that will allow OpenAI to access current and achieved articles from over 100 years of Time’s past work. The new partnership furthers Time’s commitment to expanding global access to accurate and trusted information.
