As the United Arab Emirates (UAE) works to lead the Gulf’s generative AI movement, a top-notch Arabic AI tool has been launched by a group with ties to Abu Dhabi’s authorities.
As the United Arab Emirates (UAE) works to lead the Gulf’s generative AI movement, a top-notch Arabic AI tool has been launched by a group with ties to Abu Dhabi’s authorities.
The Mohammed bin Zayed University of Artificial Intelligence (MBZUAI), California-based Cerebras, and G42, the UAE’s technology holding company, collaborated to develop the open-source, multilingual Jais model, which is intended for the more than 400 million Arabic speakers who live throughout the world.
In the midst of a global rush to secure supply for AI expansion, this release coincides with the UAE and Saudi Arabia making significant purchases of Nvidia processors essential for AI software.
The Falcon model, which the UAE previously created with 300+ Nvidia chips, is open-source. One of the largest transactions of its sort with a prospective Nvidia rival was struck this year by Cerebras, which paid $100 million to equip G42 with nine supercomputers.
Despite Arabic being one of the most widely spoken languages in the world, Andrew Jackson of G42’s Inception observed that the majority of large language models (LLMs) concentrate on English. Why shouldn’t the Arabic-speaking community have an LLM, he questioned.
Jackson asserted that although existing sophisticated LLMs such as OpenAI’s ChatGPT, Google’s PaLM, and Meta’s LLaMA can comprehend and produce Arabic text, the Arabic component in these models is much diluted.
According to its creators, Jais performs better in Arabic accuracy than Falcon and other open-source models like LLaMA. Additionally, unlike most US-centric models, Jais is created to have a more true knowledge of the region’s culture and setting, according to Professor Timothy Baldwin, acting provost of MBZUAI.
Baldwin further stated that steps had been taken to guarantee Jais respected religious and cultural sensibilities. It underwent rigorous vetting to eliminate any damaging, delicate, offending, or inappropriate information that did not reflect the ideals of the groups engaged in its production.
Jais, named after the highest mountain in the United Arab Emirates, was trained on a section of Cerebras’ Condor Galaxy 1 AI supercomputer for 21 days. As launch partners for the technology, G42 has worked with other Abu Dhabi organizations including the Abu Dhabi National Oil Company, Mubadala, and Etihad Airways.
The model’s training proved difficult since there is less high-quality Arabic language web data than there is for English. Jais responds to this by utilizing both the widely accepted current standard Arabic and the many spoken dialects of the Middle East, which are taken from media, social media, and code.
Baldwin came to the conclusion that the Arabic AI Tool, Jais, outperforms current models on a variety of tasks while being obviously superior in Arabic and competitively equivalent or even slightly better in English.