Meta, the parent company of Facebook, has introduced a new artificial intelligence model called the “Self-Taught Evaluator,” aimed at significantly reducing human input in AI training and evaluation.
According to a Reuters report, the development which was revealed on Friday, highlights Meta’s commitment to advancing autonomous AI systems.
Unlike traditional models that rely on human-generated data for training, the Self-Taught Evaluator was trained entirely using AI-generated data.
This groundbreaking approach eliminates the need for human intervention during the training phase, offering a potential path toward fully autonomous AI agents capable of self-learning and correction.
According to Meta researchers, the AI tool can learn from its own mistakes, positioning it as a major step forward in the AI field. Jason Weston, one of Meta’s lead researchers on the project, emphasized the significance of this model.
“We hope, as AI becomes more and more super-human, that it will get better and better at checking its work so that it will actually be better than the average human.
“The idea of being self-taught and able to self-evaluate is basically crucial to the idea of getting to this sort of super-human level of AI,” Weston said.
Meta explained in a publication last month that the Self-Taught Evaluator is based on a “chain of thought” technique.
This technique involves breaking down complex problems into smaller, logical steps, which improves accuracy in fields like coding, mathematics, and science.
The Self-Taught Evaluator is expected to reduce the dependency on a process known as Reinforcement Learning from Human Feedback (RLHF), which requires human experts to label data and verify responses.