According to Reuters, Facebook’s parent company, Meta, on Friday showcased a collection of cutting-edge AI models created by its research section.
One of the most notable tools is the “Self-Taught Evaluator,” which could lessen the requirement for human intervention in the creation of AI. This advancement could lead to more intelligent and independent digital agents since it is a significant step towards building AI systems that can learn from their own errors.
Together with the Self-Taught Evaluator, Meta also made available datasets to aid in the discovery of novel inorganic materials, upgrades to its image-identification Segment Anything model, and a tool for speeding up answer generation times in large language models (LLMs).
The Self-Taught Evaluator, which was first presented in a research paper in August, employs the same “chain of thought” method as OpenAI’s most recent models. To improve accuracy in domains like physics, coding, and mathematics, this method divides difficult jobs into manageable chunks.
Importantly, Meta’s researchers did not require human involvement during the training phase because they trained the evaluator only on AI-generated data.
Researchers at Meta explained that the capacity of AI to accurately assess other AI models creates new opportunities for self-improving autonomous AI systems. This might result in the creation of digital assistants that can carry out a variety of duties without the need for human assistance.
