Tencent seeks collaboration with other major AI developers to improve tech for vulnerable users



Specialised data sets can make AI services more helpful to vulnerable users who have become progressively reliant on them for emotional support and health assistance, according to Lu Shiyu, a senior researcher at Tencent Research Institute (TRI).

Since 2024, Lu’s team at TRI – the public strategy research unit of Shenzhen-based Tencent – has been developing specialised data sets with vulnerable users in mind. Data sets, such as the entire internet, equip large language models (LLMs) with general knowledge in the pre-training phase, before being fine-tuned in the post-training phase before the models are deployed.

Last year, Lu’s team and researchers at the University of Science and Technology Beijing tested leading US and Chinese-developed LLMs, including Tencent’s own Hunyuan model. They found that all the models needed to improve their handling of topics such as sex education and other subjects relevant to China’s 69 million left-behind children – kids, often in rural areas, whose parents migrate to cities for work.

The team had also worked with Chinese non-profit organisations that help vulnerable groups to develop an “elderly data set”, derived from thousands of example question-and-answer forms contributed by elderly respondents.

“To make an impact, the next step is to work with major AI model developers that have many users,” Lu said.

  • Related Posts

    China’s Lenovo warns of ‘prolonged’ memory crunch, looks to AI for rebound

    Lenovo Group, the world’s largest personal computer (PC) maker, warned of the “prolonged impact” of a global memory crunch after reporting a 21 per cent profit drop in the December…

    Continue reading
    Chinese premier tours rare earths production hub amid rising US tech rivalry

    China’s premier has vowed the country will consolidate its advantages in traditional industries like rare earths while also striving to accelerate innovation in frontier technologies such as artificial intelligence, as…

    Continue reading

    Leave a Reply

    Your email address will not be published. Required fields are marked *