Meta’s unbelievable AI breakthrough

Meta’s Fundamental AI Research (FAIR) team has released an AI tool called the ‘Self-Taught Evaluator’ that can check the accuracy of other AI models, without the need for human intervention, possibly paving the way for less human involvement in the AI development process. 

They first revealed the Self-Taught Evaluator as a concept, in a paper published in August, in which it detailed that it would use the same "chain of thought" method that OpenAI uses for its most recent o1 model, which allows it to break complex questions down into smaller, logical steps, enabling it to “think” before it responds with more considered, reliable and accurate answers. 

The tool generates different responses from AI models and uses another AI system to assess the accuracy and improve the responses to complex questions in areas like Math, science, and coding. 

But the incredible bit is that Meta trained the model on AI-generated data only, removing the need for human intervention, and reported that it performed better than models that rely on human-labelled data (like GPT-4 does, for example).

This breakthrough could pave the way for self-improving models and reduce reliance on expensive, inefficient human processes, that require human data experts to correctly label data and verify responses manually, leading to inevitable human error and inaccuracy.