Preeti
OpenAI’s success with ChatGPT was attributed to human trainers guiding the artificial intelligence model on what constitutes good and bad outputs. They pioneered the use of reinforcement learning with human feedback to fine-tune the AI model, making chatbots more reliable and preventing misbehavior. However, this technique has limitations, such as inconsistent human feedback and difficulty in rating complex outputs. OpenAI developed a new model, CriticGPT, to assist human trainers in assessing code, which proved to catch bugs missed by humans. The company plans to extend this approach beyond code in the future to improve AI models and tools like ChatGPT.
The new technique aims to improve large language models and ensure AI behaves acceptably as it becomes more capable. Anthropic, a rival to OpenAI, also announced advancements in its chatbot, Claude, through improved training and data. Both companies are exploring new ways to inspect AI models to prevent unwanted behavior. OpenAI is training its next major AI model with a focus on trustworthy and aligned output. The company is serious about ensuring responsible AI development, especially after disbanding a team dedicated to assessing long-term AI risks.
The concept of using AI models to train more powerful ones has been discussed for years and is seen as a natural development in AI research. The effectiveness and applicability of this approach remain to be fully understood, but it could lead to significant advancements in individual capabilities and more effective feedback in the long run.
Source link
Source link: https://www.wired.com/story/openai-rlhf-ai-training/
GIPHY App Key not set. Please check settings