You may have already had enough of the news about artificial intelligence (AI), however, the truth is that, most likely, It’s about the technology of the decadeso it will be very present everywhere for many years, especially due to the advances that are being made lately.
Under this context, the giant of social networks, Meta, has officially presented its new artificial intelligence called “LIMA”a new large language model whose goal is to obtain high-quality responses.
In this sense, LIMA (Less Is More for Alignment), is based on Callsan AI model with 65 billion parameters that was provided by the company that owns Facebook, Instagram and WhatsApp in the first quarter of this 2023.
Likewise, Meta has exposed that large AI-based language models usually face two stages: on the one hand, unsupervised prior training of raw text (where it learns general representations), and, on the other On the other hand, a large scale of learning through adjustment and reinforcement, with which it is sought that the technological tool is better aligned with the final tasks and the preferences of the Internet user.
It is in this way that with this new technology the mark zuckerberg company wants to show that it is possible get quality results from just a few prompts with a model that has been extensively pre-trained.
To do this, Meta’s new AI language model has used thousands of carefully filtered examples of real instructions, of which 750 come from forums like WikiHow and Exchangewhile the other 250 were written by the researchers.
In this vein, according to study published in “arxiv.org”the AI of the social media company has compared the performance of LIMA with big AIs like ChatGTP (OpenAI), Bard (Google) and Claude (Anthropic), with a controlled test of 300 indications. The results of the test show that LIMA de Meta writes “same or preferable” answers in 43%, 46% and 58% of the cases, respectively.