During the 19th Summer Summit of the 2023 Yabuli China Entrepreneur Forum held from August 25th to August 27th, Liu Qingfeng, the founder and chairman of iFLYTEK, attended the forum and discussed some updates regarding their large-scale model.
Liu Qingfeng said, “There is no problem with the algorithms in the field of AI in China, but it seems that computing power has always been held back by NVIDIA.” In the past “Battle of Hundred Models,” most of them were trained by NVIDIA. Within enterprises, only minor optimizations and training can be done, making it still quite difficult to train models.
He said, “I am delighted to announce that Huawei’s GPU capability is now on par with NVIDIA A100. Ren Zhengfei attaches great importance to this, and three directors from Huawei are working at iFLYTEK specifically for benchmarking against NVIDIA A100.”
In addition, Liu Qingfeng once again stated that this year iFLYTEK has set a goal. By October 24th of this year, iFLYTEK will release a general-purpose large-scale model that fully competes with ChatGPT and aims to achieve comprehensive superiority in Chinese language while being on par with it in English. “To be honest, we still have gaps compared to it today, but we are very clear that we can catch up by October and benchmark GPT-4 in the first half of next year.”
At the annual meeting (board meeting) of the China Listed Companies Association and the 2023 China Listed Companies Summit held on May 27th, Liu Qingfeng shared the upgrade iteration milestones of iFLYTEK’s StarFire Cognitive Model: On June 9th, breakthroughs will be made in open-ended question answering, multi-turn dialogue capability, and mathematical ability; on August 15th, breakthroughs will be made in code capability and further upgrades to multimodal interaction; on October 24th, the general model will benchmark ChatGPT (surpassing it in Chinese language capabilities and being equivalent in English).
On May 6th, iFLYTEK officially released the iFLYTEK Starfire Cognitive Model, which has seven core capabilities including text generation, language understanding, knowledge questioning and answering, logical reasoning, mathematical ability, coding ability, and multimodal capability.