Artificial intelligence has witnessed one of the most remarkable growth and development in recent years. New models, projects, and innovations have been developing at a rapid speed. AI which started as an early rule-based system has now grown into a more sophisticated technology with large language models (LLMs) and other features. According to experts, the LLMs like GTP-4 were one of the greatest achievements of artificial intelligence so far. These models helped to ease the work by generating human-like texts with accuracy and precision, translating languages and passages, producing creative content, and clearing doubts put forth by the users. However, the LLMs were fed and trained with large data with the help of human intervention. With the latest news coming up, Mark Zuckerberg’s Meta is planning to introduce the latest and most innovative AI models. Researchers consider this as a massive leap in the development of AI as Meta is planning to develop AI models that do not require human interventions.
On Friday, Facebook owner Meta published an online post on their official website suggesting their latest plan to introduce a new set of AI models. According to the available information, these concepts came from Meta’s research division, Fundamental AI Research (FAIR). Currently, it plans to release two models that include the “Self-Taught Evaluator”, which will improve and enhance its functions and performance with less human involvement, and the other model which will mix both text and speech.
The latest announcements were made based on a research paper published by Meta last August. According to this research paper, Meta introduces how its AI models rely completely on the “chain of thought” technique as already used by OpenAI in its recent version of 0 models. This technique is developed in such a way that the model thinks before responding. So, it is said that the technique will make more reliable judgments on the model’s responses and enhance its working with the users. Experts comment that this technique will also help to break down complex concepts and ideas into simpler ones, which will help enhance the accuracy of responses even in complex ideas like science, coding, and mathematics.
Meta officials claim that the new technologies will help them achieve their company’s goal of achieving advanced machine intelligence with the least human intervention. Some of the suggested updates included in the new models of Meta AI include updated Segment Anything Model 2 doe images and videos, Layer Skip, Meta Spirit LM, OMat24, MEXMA, SALSA, and Self Taught Evaluator.
Self Taught Evaluator
Self Taught Evaluator is one of the latest and most unique techniques introduced by Meta. This AI model claims to be capable of evaluating the works of other AI models with the help of synthetic data. According to the company’s claims, this model will help train new reward models with synthetic data and without relying too much on human interventions. This model is considered a self-improvement scheme of Meta AI as it will provide outputs capable of training, producing, and evaluating LLMs.
So, to put it in simple words, the ‘Self Taught Evaluator’ creates its own data to train new models with human labeling. The different outcomes produced by this model will be used to produce more improved outcomes. Meta also claims that the outcomes produced by this model will be more accurate than the models that rely on human-produced data.
Meta Spirit LM
Meta Spirit LM is another important model introduced by Meta. Unlike the LLM models that convert speech to text and vice versa in a relatively robotic manner, which lacks natural expressiveness, Meta Spirit LM plans to work both on text and speech in a more natural and expressive way. It is also introduced as the first open-source model of Meta.
Also Read: Motorola Finally Competes With Samsung & Google On AI Innovation
