Since the birth of generative AI like ChatGPT, it has shown its influence and can process linguistic data remarkably. As a result, the whole world is changing rapidly, and countless Gen AI in languages are born to use the service.
However, all Gen AIs seem to be particularly fluent when working in English, but with other languages, including Thai, they are not as effective as they should be. Many people are waiting for when there will be Gen AI who specializes in the local language of each country. In order to be able to use it in practice.
On March 8, 2024, SCBX and SCB 10X jointly organized the SCBX Unlocking AI: EP6 Unveiling SCB 10X's Typhoon seminar to delve into the background of the Typhoon. Large Language Model optimized for Thai
According to the evaluation of its capabilities, the performance of the Typhoon LLM is similar to GPT-3.5 and can process or analyze Thai text and words 2.62 times more efficiently than GPT-4. With an understanding of the vocabulary and culture of the Thai language.
Kaweewut Temphuwapat, Head of R&D and Innovation Lab from SCBX, and Kasima Tharnpipitchai, Head of Al Strategy from SCB 10X, explained why this LLM was developed and how it can help improve Thailand effectively.
Not only using technology, but also developing technology.
Mr. Kwewut explained that in the past, there have been many organizations that have created many new innovations. These innovations are strengths that help to elevate the organization. But that is what Thailand lacks because it is only interested in adoption rather than development itself.
If you want to see development that leads to survival in the future, Organizations only think about users. no SCBX adheres to this principle and is constantly striving to innovate, and one of them is Generative AI called Typhoon.
SCB is an organization that is already familiar with the use of Gen AI. Currently, Microsoft Co-Pilot is used in many sectors of work, including research, meeting summaries, and many others, which found a pain point that even though Gen AI from abroad is good, it is not proficient in Thai at all. Therefore, it is an opportunity for organizations to develop new AI to solve this problem.
The important thing is that it will be open source for everyone, not just people in SCB, to access and take advantage of this technology.
Gen AI Open Source เพื่อชาติ
Meanwhile, Mr. Kasima said that he was one of the people who co-pushed for the creation of AI storm waves like Typhoon in the first place, and the more he continued to develop, he found that it would be better to let this typhoon blow through Thailand. Not only in SCB's departments.
"Currently, useful technology should be more open source because the AI competition is a global competition. If we are going to compete with foreigners who mainly use English. Closed AI development will not be able to make Thailand compete with anyone else. Except for fighting each other."
"Therefore, we should work together to develop more to form a community and create an ecosystem to work together in a harmonious manner."
Mr. Kwewut added that Thailand has many talented developers, but this does not mean that SCB will be able to work with everyone across the country. Open Source is the best way to help Typhoon develop faster and expand its usage model to achieve higher performance.
Data Access Headaches
To make Gen AI develop. The important thing to have is that a huge amount of data is gradually Feed it to artificial intelligence gradually. But the main problem that Mr. Kasima encountered was where to get the information from so that Typhoon could learn well.
"Searching for information in Thai is not the same as English, which is easier to find, so it takes more investment to find information. In addition, when we know where to find information, we will have to go to the hospital. We must also screen and clean up that information to leave only good and quality information."
Mr. Kasima said that when developing the Typhoon 7B model, it will be taught through data in the form of ONET, TGAT, TPAT, and many other tests. Glossary The context or culture of the Thai language, as well as general knowledge that occurs around the world.
And the result of learning surprised him many times more than he thought!
Typhoon vs ChatGPT, punch to punch, who is more knowledgeable about Thai culture?
In addition to sharing the concept of developing Typhoon at today's event, Mr. Kasima and Mr. Kwewut also tried out this new Gen AI model for the participants to see what the results are after more than 7-8 months of development, and how specific information can be provided in Thai language that is specific to Thai culture.
One of the examples revealed at this event was to fill out a prompt to request a recipe for 'grilled chicken' from Typhoon. What can confirm that the recipe is a Thai recipe is the use of raw materials such as roasted rice.
Meanwhile, If you ask ChatGPT about the recipe for making 'grilled chicken'. Although the information obtained is more systematic, there is nothing to indicate that it is actually a Thai version of the grilled chicken recipe, and it is possible that it may be a more international grilled chicken recipe, which shows that Typhoon understands the unique Thai culture quite well.
For the next plan. SCBX and SCB 10X aim to develop Typhoon to increase the competitiveness of Thailand's AI industry to be more efficient and advanced in the future.
Developers interested in helping Typhoon grow together can sign up for a trial of the initial version of the Instruction-tuned model in the form of an API to develop large-scale Thai models with increased efficiency and advancement soon. https://opentyphoon.ai
