zalo llm.jpg

On December 16, at AI Summit 2023, Zalo provided information about the LLM in Vietnamese language it is developing. The solution aims to build AI apps generating compatible documents and better serve Vietnamese users.

Zalo CTO Nguyen Minh Thu said the institution set to work on the LLM earlier this year. Some small apps of generative AI technology, such as composing poets and building avatar, have been integrated into the app and welcomed by millions of users.

At Zalo AI Summit 2023, a Vietnamese developer for the first time demonstrated the LLM which was a version under the development. Zalo’s LLM has the ability to communicate and answer the questions raised by users on different fields in Vietnamese language.

To challenge the capabilities of the new language model, the model was integrated into Kiki to compete in knowledge via Kahoot game. Also participating in the game were previously released solutions such as GPT-3.5, GPT-4, Llama22, and QWen, a recently released Vietnamese LLM as well as real players. 

The exam question set included 20 questions related to many different fields, posed by Tran Manh Hiep, administrator of Tinh Te Forum, who did not discuss the questions with Zalo in advance.

The process of answering questions by LLMs was shown on a screen, so that people can keep watching directly and ensure the fairness of the competition.

During the process, Zalo’s LLM performed relatively well when answering popular questions about general knowledge. Kiki and GPT-4 were the two tools which could give right answers to the question related to ‘Wish you were here’, a famous song of Pink Floyd.

“I did not think that the AI chatbot would give an exact answer to the question, because this was a topic that I accidentally thought of and it relates to my hobby,” Hiep said.

Kiki and GPT-4 also could give true answers to questions related to domestic issues, including one about Bau Da liquor, while other LLMs met difficulties.

Meanwhile, all six AIs at the competition could not overcome the challenge when answering a question about the maximum speed of vehicles in urban areas, and a question about the animal designation of 2024. These were the parts that helped real competitors get ahead in Kahoot game.

As such, all LLMs could not defeat real competitors with the set of questions in Vietnamese language. However, GPT-4 and Kiki were the two tools with good performance. Zalo AI’s LLM was not inferior to the LLMs from Meta and Alibaba.

Trong Dat