The above sharing was given by Professor Vu Ha Van, Scientific Director of VinBigdata, at a discussion with reporters on the sidelines of the recent ViGPT launch event.

ViGPT needs contributions from technical people and the community

Professor Vu Ha Van said that for large companies like Google, when developing large languages, they will choose English or French as the main language. Although there is also Vietnamese, the search or lookup results will be relatively slow compared to other languages. To some extent, the answers of these large language models to questions from Vietnamese will not be complete and accurate.

Therefore, VinBigdata hopes that over time, ViGPT will surpass them in accuracy on questions directly related to culture, history, geography, etc., information that is specific to the Vietnamese people. This is what the creators of the Vietnamese language model want and aim for in the future when asking questions about Vietnamese people, this will be a better source of comparison than foreign ones.

Going deeper, the Director of Science of VinBigdata analyzed, for example, a question in a "sensitive" political period about the history of Truong Sa and Hoang Sa, it is very difficult for us to ensure that the answer from Google or OpenAI does not carry the political bias of the founders or behind these companies. Here we have other options in Vietnam, it would be better if we think about that issue.

“Our purpose in building a large language model for Vietnamese people is to bring the best answers to Vietnamese people, we cannot know their purpose,” Professor Vu Ha Van shared.

Admitting that there are many things that ViGPT currently cannot do as well as ChatGPT or Google Bard, because the investment rate of these businesses and the time they spend to implement them are thousands of times higher. However, Professor Vu Ha Van said that for some questions that are biased towards Vietnam such as "Whose flag is embroidered with six golden characters?", ViGPT will answer that it is Tran Quoc Toan's, while the other versions may be wrong. In the future, with in-depth questions like this, ViGPT will do better if there is feedback from domestic users.

“If users only criticize, or think that this big language model is still stupid when my 10-year-old child knows the questions that he doesn’t, or asks trick questions to prove that we are smarter than AI. We are smarter than AI, but it is not for any purpose, here we do not make the product better but make the people who make the product sadder. Therefore, VinBigdata needs the common contribution of technical people and the community, we need the companionship of Vietnamese people in perfecting the product so that it is not just a simple service tool, but also the pride of Vietnamese people”, Professor Vu Ha Van emphasized.

Ready to support and accompany the Vietnamese language model

Speaking with VietNamNet , representatives of startups working on AI in Vietnam said they are ready to support and accompany VinBigdata's Vietnamese language model.

Mr. Dinh Tran Tuan Linh, CTO of Unikon Joint Stock Company, the owner of the Aicontent.vn platform, said that currently, not many countries in Asia have made efforts to successfully train their own large language models, leading the way are China, Korea, Japan... Therefore, ViGPT is an important signal for the Vietnamese people's efforts to invest in core technology. According to Mr. Dinh Tran Tuan Linh, any journey of a thousand miles must start with the first steps, as a pioneer in AI application, Unikon is willing to participate in contributing, testing, giving feedback and even using ViGPT in some suitable scale projects.

Meanwhile, Mr. Dang Huu Son, Co-founder of Lovinbot, said that VinBigdata's listening to the community and experts' comments is a very good thing to develop a large language model specifically for Vietnamese people. As a technician, Mr. Dang Huu Son also gave feedback to VinBigdata's technical team after using the product.

According to Mr. Dang Huu Son, a newly launched product cannot be completed immediately, but it also cannot receive full support from the community right away, because Vietnamese people have long thought that Vietnam can not do that technology, so it still needs time. At the same time, VinBigdata needs to have specific instructions on how the community can support and accompany it better.

Mr. Dang Huu Loc, founder of the Mindmaid platform, also shared that currently there are very few countries in the world that have built a native language model. Even rich countries with strong information technology such as India, or countries with higher GDP than Vietnam such as Indonesia, the Middle East... cannot do it just because they want to, because it also depends on the characteristics of the language. Therefore, from a broader perspective, Vietnam has a strategic advantage in building a native language model, and this will be a strategic advantage for Vietnamese people to compete globally.

According to Mr. Dang Huu Loc, any effort to build a large Vietnamese language model is valuable, and needs to be commented on in a specific way to make the model more perfect every day, instead of using some current shortcomings to deny all the efforts of domestic technology units. Vietnamese people should also widely disseminate the importance of large language technology in the AI ​​era and discuss more about how to apply it to create value for themselves and Vietnamese businesses, instead of comparing the large Vietnamese language model with the best large language models in the world today. Because large language is a general AI technology, it may not be good at this problem, but it is suitable for other specific problems. In particular, the large Vietnamese language model will have a better advantage in problems related to understanding and generating Vietnamese.

Community ViGPT will be provided free of charge to non-profit organizations . Community ViGPT will be provided free of charge by VinBigdata to non-profit organizations. However, organizations using this version will have to pay for infrastructure costs such as cloud and other resources when deploying.