Leaving the US to return home, Nguyen Hoang Quan and his colleagues at VILM developed a free artificial intelligence (AI) system for Vietnamese people to use in customer care fields with more than hundreds of thousands of downloads per month.
In June 2023, Nguyen Hoang Quan, 25 years old, and his colleagues Pham Nhut Huy, 23 years old, an artificial intelligence engineer at ZaloAI, and Dao Minh Dung, 24 years old, a PhD student at the University of Cork, Ireland, co-founded the non-profit organization VILM, with the desire to help Vietnamese people experience the most advanced AI technology in the most optimal way.
After nearly 6 months of research and application, the team successfully developed three free AI models including OpenHermes, VinaLlama and Vistral. These are the foundational studies for developing AI systems that can understand and act according to user intentions (Large Action Model). These models are applied in technology fields such as machine control, robots to help support people with disabilities better without caregivers, or help programmers fix errors, virtual assistants to take care of customers or ask questions for free.
OpenHermes reaches 85,000 downloads per month, reaching the top 10 most downloaded language models on HuggingFace (the world's largest AI model sharing site). Screenshot
The OpenHermes model is a large language model that supports English like ChatGPT, but with superior scores. They allow users to download the model to their personal computers to use without the internet. Notably, the amount of training data of OpenHermes is only 1/100 of the training data of ChatGPT from OpenAI. Currently, this application receives more than 50,000 downloads per month. OpenHermes-2.5 and OpenHermes-2.5-Vision are being used by more than 40 startups in Silicon Valley (USA),
VinaLlama and Vistral are two language models focused on serving the Vietnamese market, aiming to help domestic users experience the most advanced AI technology more easily.
Hoang Quan spent 7 years studying in the US and worked at OpenAI as a research engineer for the ChatGPT artificial intelligence model, despite not having graduated from university. In 2022, he worked as a data engineer for Microsoft and OpenAI's Bing Chat product, earning thousands of dollars. By 2023, facing a wave of technology layoffs in the US, Quan realized that the post-graduation job market was very bleak, but seeing opportunities in Vietnam, he decided to return home.
Nguyen Hoang Quan. Photo: NVCC
At VILM, Quan is the chief engineer responsible for researching data improvement techniques as well as AI training. While Nhut Huy takes on the role of technical research in AI training and Minh Dung proposes new methods in theoretical research.
Quan explained that current large language models such as ChatGPT (Large Language Model) can only provide text output, while humans have many ways to communicate and acquire knowledge. That is why the team aims to create a system that can operate flexibly between different types of input and output (can receive and output data such as language, images, videos, sounds), not just stopping at the language level.
To achieve the goal of creating Large Action Models, the team had to overcome two problems: security and speed. Current AI applications mostly use user data and send it to the servers of companies like OpenAI for processing, which raises security concerns. The team focused on creating AI models that were small and fast enough to be processed directly on mobile devices, while balancing performance and speed to avoid affecting the user experience.
The experimental team used data generated from AI to train the AI itself, instead of going down the path of using data from real sources. Initially, they had difficulty finding computational resources (computers to train AI), but later convinced large companies and labs around the world to sponsor.
Quan said that the main purpose of making these products is to help people access AI applications quickly and with quality not inferior to ChatGPT or Bing Chat, and to make research and creation of AI models in the future simpler. Instead of using ChatGPT, which is limited in Vietnamese language and culture, Vietnamese businesses can download VinaLlama in Vietnamese.
VinaLlama language model easily solves a problem in Vietnamese, in the picture is a demo of VinaLlama product in solving math problems. Screenshot.
Mr. Dang Hai Loc, Founder of the AI Chatbot building platform Mindmaid, said that from the perspective of an AI application developer, he realized that cost and data privacy are the two issues that businesses are most concerned about when deploying AI applications. The most satisfactory solution to this problem is open-source LLM models, which can run on the enterprise's infrastructure and can learn (fine-tune) more of the enterprise's own data. Therefore, Vietnamese open-source LLM models such as VinaLlama, Vistral... are very valuable in promoting AI applications in Vietnam.
"These open source models also enable more programmers and technology enthusiasts to access the AI Engineer field with just a macbook instead of having to invest in expensive GPU (graphics card) infrastructure. This will also promote the AI Engineer force in Vietnam, a role that is in high demand in the near future," said Mr. Loc.
According to Quan, Vietnamese people have a very good foundation in scientific theory, are good at AI, and ChatGPT also has human resources participating in research, but they have more difficulty in quickly catching up with the ever-changing wave of technology. "What Vietnamese people need is experience in making products for end users to truly understand the problems they face in order to properly orient their research," Quan said about the reason for researching free AI models to support Vietnamese people in technology. He said he has collaborated with many international groups and is always ready to cooperate with research groups in Vietnam.
Nhu Quynh
Source link
Comment (0)