On October 31st, according to Smart Episode, ex-Ali, byte model managerYang HongxiaAfter leaving bytes, a new company was created in July 2024 InfiX.ai and yesterday in hong kong, the latest developments were revealed。
Yang Hongxia suggested that large model pre-training should not be a calculator competition for a small number of giants, but should involve small and medium-sized enterprises, research institutions and even individuals through " decentrization " 。
YANG HONGXIA LED THE DEVELOPMENT OF A LARGE M6 MODEL AT THE ARIDA MO COMPOUND, AND THEN CONTINUED TO WORK DEEP IN THE DIRECTION OF THE LARGE BYTE。
She noted that, while the existing "centreization" model could provide a technological breakthrough, there were limitations in the application of landing sites, especially in the context of data-sensitive and localized deployments, where post-training was difficult to fill the knowledge gap at the pre-training stage, leading to a high incidence of hallucinations。
Core technical pathways for InfiX.ai include:
InfiR2 FP8: A maximum training rate of 22% and a reduction in visible storage consumption of 14% with little loss of performance
(a) Model integration technology InfiFusion: by integrating "expert models" in different areas, avoiding duplication of training and increasing knowledge coverage
InfiMed: Demonstrated the reasoning capacity beyond the same model in complex medical missions such as cancer
Multi-Intelligent: Automatically decomposed complex tasks and reduced development costs。
Yang Hongxia emphasized that in the future each enterprise would have its own large field model and that cross-cutting, cross-regional integration of knowledge through model integration would form the basis of globalization. She thinks it's universal artificial intelligenceAGIIt should not be limited to top-level institutions, but will evolve into “all-people collaboration”。