These days, artificial intelligence is everywhere. The majority of large tech companies have embraced artificial intelligence, either by creating their own models or incorporating third-party models into their services or new features. Meta, one of the industry’s big players, has announced a new “Self-Taught Evaluator” AI model that aims to autonomously evaluate and train other AI models.
One of the main problems with developing AI models is the related costs. In the current state of the industry, large investments are needed to catch up and be competitive. Furthermore, developers use a technique known as “Reinforcement Learning from Human Feedback” (RLAIF) during the training process. As its name suggests, RLAIF requires human participation, which can slow down the process. That said, Meta’s new Self-Taught Evaluator aims to eliminate such a requirement.
Meta’s Self-Taught Evaluator AI model can train and evaluate other AIs without human intervention
RLAIF makes use of human experts to ensure that the AI in development gives solid and reliable answers. After all, it doesn’t matter how powerful an AI is if it has a high error rate. Human parties must also ensure that the data used for the AI training process is factually correct. This not only increases development times but also associated costs.
However, Meta’s Self-Taught Evaluator model is capable of evaluating and training other AI models. To achieve this, Self-Taught Evaluator uses the “chain of thought” technique that OpenAI implemented in the o1 models. This technique is based on addressing complex problems by dividing them into smaller logical steps. This results in more precise answers in advanced areas such as science, coding, and mathematics.
In fact, Meta developed the Self-Taught Evaluator model itself with the chain of thought technique. They used data purely generated by AI to train it. “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,” said Jason Weston, one of the researchers involved.
Teasing a potential future full of autonomous AIs
AI capable of learning and evaluating by itself without human intervention sounds like a futuristic concept straight out of a science fiction movie. However, recent developments suggest that we might not be too far away from something like this. AI experts suggest that the implementation of such models in various fields could largely eliminate human intervention.
