Who needs humans to train AI? Meta's new AI can train other AIs without human feedback
Meta is now announcing an AI model that basically teaches itself and others without human feedback.
Artificial Intelligence is everywhere nowadays and its presence is also bound to grow as we move forward. Many tech companies have welcomed AI and some have also created their own models instead of taking advantage of other companies' models.
Meta is one of the big players in the tech world, and has now announced a new AI model called "Self-Taught Evaluator". The aim of it is to autonomously evaluate and train other AI models.
Training AI is quite a costly endeavor, and it's a pretty tough race right now with all the competition. Developers use a technique called "Reinforcement Learning from Human Feedback", otherwise referred to as RLAIF, to train AI models. As you may presume, this technique needs human feedback and this can lead to a slower process. Well, Meta's on this with its new "Self-Taught Evaluator".
Meta AI. | Image Credit - Meta
Human feedback is needed to ensure the AI that's in development provides accurate and reliable answers, and also, checks for errors with processing data. However, the new model is capable of evaluating and training other models. It uses a technique called "chain of thought" that OpenAI used in the o1 models.
The "chain of thought" technique addresses complex problems by breaking them up into smaller logical steps. This makes an AI give more precise answers to scientific, coding, or math problems. Meta's new model itself is developed using this technique. This way, the AI trains on AI-generated data... sounds too sci-fi? Well, probably not for long, I reckon!
Jason Weston, one of the researchers involved with the process, stated that as AI evolves and becomes "more super-human", it will get better at checking its own work and it will even become better than the average human at checking its own work.
FCC OKs Cingular\'s purchase of AT&T Wireless
Things that are NOT allowed: