"Vietnamese brains" in Google's AI project and the race against ChatGPT.

(Dan Tri Newspaper) - Twenty years ago, Luong Minh Thang was a mathematics student of teacher Le Ba Khanh Trinh. Now, he has returned to his old school to introduce to his teacher an artificial intelligence (AI) software that can solve difficult math problems at the Olympic level.

One day in early December 2023, Dr. Luong Minh Thang visited his high school in Ho Chi Minh City to introduce his artificial intelligence software, AlphaGeometry, to his former teacher, Dr. Le Ba Khanh Trinh. AlphaGeometry uses a neural language model and is trained with large-scale aggregated data, capable of solving IMO (International Mathematical Olympiad) problems. In a test involving 30 of the latest Olympic-level problems, AlphaGeometry demonstrated remarkable ability, solving 25 problems – equivalent to the achievement of an IMO gold medalist. The lead author of AlphaGeometry is Dr. Trinh Hoang Trieu (30 years old, a graduate of New York University) and a team of senior experts at Google DeepMind: Dr. Luong Minh Thang, Dr. Le Viet Quoc; and Dr. Yuhuai Wu (co-founder of xAI, formerly at Google). Solving IMO problems at the medal-winning level is the dream of most gifted math students around the world . Dr. Le Ba Khanh Trinh became the first Vietnamese to win a gold medal at the IMO (in 1979), and is the only Vietnamese to date to have won a special award at this competition for his elegant and concise solution to a geometry problem. When Dr. Luong Minh Thang introduced AlphaGeometry to his former teacher, he received the comment, "Very impressive!" However, "Mr. Khanh Trinh was still not satisfied with the AI's solution because it lacked the soul, the beauty of a solution that he expected," Dr. Thang recounted.

Dr. Luong Minh Thang (far right) and Dr. Le Ba Khanh Trinh (center) are discussing a problem from the 2015 IMO exam that AlphaGeometry solved. (Photo: Wendy Uyen Nguyen)

For his part, Dr. Luong Minh Thang believes that AlphaGeometry marks a significant milestone towards software with human-like intelligence and the ability to learn independently. This is a prerequisite for achieving super-intelligent artificial intelligence (AGI). A former mathematics student at the High School for the Gifted, Vietnam National University Ho Chi Minh City, Dr. Luong Minh Thang studied Computer Science at the National University of Singapore, earned his PhD in Computer Science from Stanford University (USA) in 2016, and has spent seven years working on Google's AI projects. At Google DeepMind, Dr. Luong Minh Thang built cutting-edge models in both language (QANet, ELECTRA) and vision (UDA, NoisyStudent). He co-founded Meena, the world's best chatbot in 2020, which later became Google LaMDA, Bard, and is now Gemini – Google's flagship AI product. Besides his connection with AI during his studies and research, Dr. Thang also found love with Wendy Uyen Nguyen, Global Director of External Affairs and founder of the Stanford Institute for Microbiology & Epidemiology (Stanford University), also thanks to AI. Journalist Vo Thanh from Dan Tri newspaper had a conversation with the two young researchers, Luong Minh Thang and Wendy Uyen Nguyen.

Vo Thanh: Dr. Luong Minh Thang, how did you become a senior researcher at Google? - Luong Minh Thang: I officially started working at Google in 2016, but my connection began in 2014 when I interned at Google Brain. At that time, I participated in a project to improve translation quality, researching the application of artificial neural networks to help translation programs automatically translate complex sentences instead of just individual phrases as before. The mission of this project was to help machines better understand the meaning of words in many different languages, while also processing long passages of text. With this new approach, what we did in two years was equivalent to the total of 20 years of work done previously. Also from 2014 to 2016, I completed my PhD thesis on translation, becoming one of the pioneering researchers in the field of deep learning, applying machine learning methods based on artificial neural networks to develop software capable of self-training in machine translation. Our research supports translations served at translate.google.com, allowing users to translate text, web pages, and even speech. Google Translate is a product deployed in many Google services, having a widespread and profound impact in helping to bridge language barriers around the world. In 2018, I co-founded the Meena project, a chatbot that can chat with users about anything. This was a new direction because at that time, chatbots from Google or Microsoft were limited to simple tasks and tended to specialize in a particular field.

With Meena, we wanted to develop a chatbot that could converse about virtually anything a user wanted, with specific and relevant content. In other words, we wanted anyone to be able to chat with the machine as if it were a natural conversation with an intelligent person, without limitations on field or knowledge, and without hesitation or confusion. In 2020, Meena became the best chatbot in the world. The Meena model had 2.6 billion parameters and was trained on 341 GB of filtered text from public social media conversations. Compared to the OpenAI GPT-2 model, Meena had 1.7 times the model size and was trained with 8.5 times more data. However, Google did not release Meena at that time due to concerns about risk. This was the period when Microsoft had just released an AI chatbot and encountered several problems, such as providing incorrect information, arguing with users, and exhibiting racist behavior. Shortly after its introduction, Microsoft had to take it down for repairs. This incident made Google cautious about Meena. By the end of 2022, ChatGPT launched and quickly gained attention for its ability to converse fluently across various knowledge areas, at which point Google finally started to react and catch up. Currently, I continue to lead several important projects at Google, building better AI models for reasoning, logical analysis, problem-solving, image processing, etc.

Vo Thanh: Hello Wendy Uyen Nguyen. Luong Minh Thang just shared about his connection with the field of artificial intelligence, so how did the connection between the two young Vietnamese researchers in the US come about? - Wendy Uyen Nguyen: I met and got to know Thang because of our shared interest in AI. My initial specialization was psychology, but later I decided to switch to hospital management (Master's degree from UCSF School of Medicine - University of California San Francisco), and then earned a Master's degree in Executive Leadership in Business Administration from Stanford University. Academically, Thang and I follow two completely different paths; one is a scientist, the other is a strategy/business management expert. But perhaps it is precisely because of this that we can see different perspectives and support each other. Working in the medical field, I've realized that applying technology in general, and AI in particular, can help minimize risks for patients. As we know, mistakes in other fields can be corrected, albeit at a cost, but medical errors often come at a high price, costing people their health and lives. Previously, I participated in a research project developing a technology model to help doctors practice their profession. I happened to meet Thang in a group of friends, and I asked him to be an AI consultant for the project. Initially, we probably didn't think about a romantic relationship. But besides our shared interest in AI, we also met through a desire to contribute to the community and bridge scientific and technological cooperation between the United States and Vietnam. Thang and I also have a great passion for training talented AI professionals in Vietnam. We found love while walking the same path. - Luong Minh Thang: I'd like to add that both my wife and I are passionate about music . Wendy studied piano at the Conservatory from a young age and played very well, while I loved to sing. We often met up in our group of friends and had a lot of fun playing and singing together, which made us even closer.

Mr. Luong Minh Thang and Ms. Wendy Uyen Nguyen when they were students at the Business Administration Department, Stanford University (Photo: Provided by the interviewee)

Vo Thanh: So, besides the application of technology, in the story that Mr. Thang and Ms. Wendy Uyen Nguyen just told, AI also plays a role in connecting emotions. - Luong Minh Thang: Yes, in Vietnamese, if we add an accent mark to the word "AI," it becomes the word "Ai," which also means love. Vo Thanh: Do Mr. Thang and Ms. Wendy Uyen Nguyen always agree on AI, or have they ever debated a specific issue, for example, the benefits and risks of AI? - Wendy Uyen Nguyen: I'm not an engineer specializing in AI, but in the field of healthcare, I see that everything has two sides, good and bad. Regarding risks, there's the fear that AI will become too intelligent, escaping human control, even destroying humanity in the future. That fear is understandable. But I think AI researchers have also considered this issue, and governments are also considering how to develop AI within a controlled environment. Therefore, the more pressing issue now is how to apply AI in practice to improve people's lives. For example, AI applications in autonomous driving technology or in medical diagnosis and treatment…

- Luong Minh Thang: Looking back 10 years ago, when I was doing research under the guidance of Professor Christopher Manning - a leading professor in the field of applying deep learning to natural language processing - no one could have imagined that ChatGPT or Gemini could help us write a poem for our wife, write an email to our boss asking for a raise, etc. Building a natural language model that can perform tasks in all fields with high conversational quality was something no one could have imagined before, because each task has its own, very complex requirements. For example, translation requires different requirements than writing an email, and even more so than writing poetry... Making a machine rhyme poetry was an extremely difficult, almost impossible task 10 years ago. But today, this is a reality. With the current development of AI, I think next year it will reach significant milestones, such as short films of Hollywood quality made using AI, or breakthroughs in mathematics… So what about the next 10 years? I'm sure AI will develop beyond our wildest imagination. When AI develops too quickly, many people will worry because they don't know what will happen. That's why there are studies on AI control circles, ensuring that AI develops within those circles, doesn't "jump" out on its own, and outsiders cannot attack or take control of AI. Personally, I am very happy and excited about the rapid development of AI and believe that AI will bring many benefits to humanity. For example, in the field of medicine, with AI-generated information, humanity could move towards a future where someone provides input information about gender, age, weight, medical history, etc., and AI will create the most suitable drug formula for them. Thus, each person will have their own unique formula for medicine. Or, in education , back when I was in school, we needed a calculator to solve advanced math problems, but now AI is the supercomputer. With AI, young children can discover new mathematical laws, new research in physics, space, time… This means AI can help them explore the universe, discover time black holes, and other mysteries that humanity doesn't yet fully understand. Demis Hassabis, the founder and CEO of Google DeepMind, once said something I really appreciate: AI has the potential to become one of the most important and beneficial technologies ever invented, a tool to help us understand science and the future of humanity.

Vo Thanh : Returning to the Meena project, Google had a powerful chatbot but didn't release it, while OpenAI had already launched ChatGPT in November 2022. What challenges did this situation present to Google's AI project team? - Luong Minh Thang: The launch of ChatGPT attracted enormous attention, a shock that led Google to issue a "Code Red" warning throughout the company. A "Code Red" warning means the company is in a dangerous situation. Immediately after ChatGPT appeared, we embarked on a 100-day AI race. From the Meena chatbot, our team developed Bard (now Gemini), Google's main AI product, released in February 2023 (about 100 days after ChatGPT was launched). The reason we were able to implement such a difficult project so quickly is because Google had already conducted research on AI beforehand. For example, my research on translation is based on the transformer architecture (a deep learning model for natural language processing), which supports most major language models and has revolutionized the field of AI. ChatGPT and Gemini are both developed on the transformer architecture. Furthermore, Google has very talented engineers who possess the ability to build anything from scratch; Google also controls GPU chips and massive data sources for training AI, and has many product distribution channels like Gmail and YouTube… Of course, besides these strengths and competitive advantages, I think because Google is a large corporation, the implementation of projects usually has to follow a careful process. This causes Google to be slow in releasing products. Other companies are more flexible; sometimes they release products even if they are still buggy.

Mr. Luong Minh Thang and Ms. Wendy Uyen Nguyen met with Mr. John Kerry, former US Secretary of State, to share their efforts in realizing their vision of training and developing a generation of talented young people in the field of Artificial Intelligence in Vietnam, September 2023 (Photo: VietAI)

Vo Thanh: I heard that when Google issued a red alert, engineers had to go into a battle of "Put both hands on the keyboard! Even if it meant working overtime, skipping meals and sleep, we had to solve the problem." So what happened during that 100-day campaign? - Luong Minh Thang: Actually, when developing the Bard chatbot product, no one said it had to be done in 100 days, but everyone understood that they needed to put in their best effort for the survival of Google. That effort resulted in productivity three times, five times, and even ten times higher than normal. Sometimes I jokingly tell my team that we've only been working for Tuesday, but it feels like the whole week is over. Because the large workload we set for a week was solved in just two days, and in fact, everyone had been thinking and discussing since the weekend to get straight to work at the beginning of the week. For me, those 100 days felt like a year. Perhaps it was a case of "every cloud has a silver lining," because the creation of ChatGPT made everyone at Google more united and hardworking, all striving towards the goal of creating a truly good and useful AI product for users. The race was incredibly stressful, but it was also a memorable time because it coincided with Christmas in the US. In the early days, Sundar Pichai – Google's CEO – and Sergey Brin – Google's co-founder – visited our project team, chatting and encouraging us. Before that, Larry Page and Sergey Brin, the two founders of Google, had already retired from leadership roles at the company, meaning they were no longer in the office. But when the AI ​​project started, Sergey Brin took the time to talk with the engineers, and on some evenings he even messaged the project team asking, "Anyone want to go out for dinner at 8 PM?" This is something I found very interesting.

Vo Thanh: You could think of the 100-day campaign as the peak phase. After this phase, the AI ​​race between Google and other tech companies doesn't stop but continues day by day, hour by hour? Luong Minh Thang: Announcing an AI product is a crucial milestone. After the product starts being used, the work moves into the feedback and improvement phase... The current challenge is knowing how users are receiving the product, whether they are satisfied, and what they want. User feedback is extremely important. ChatGPT was able to quickly improve its product thanks to its large user base and the amount of feedback it received. Google's AI product was released later but also attracted a very large user base. Google's advantage is having many product distribution channels such as Gmail, YouTube… Most of Google's current products have a "Gemini" element, meaning they integrate AI. For example, when you create presentation slides, you can use AI to make the presentation more beautiful and visually appealing. I myself now use commands to let AI write code when I write, and I only do post-verification, saving a lot of time. Our AI project team initially had only 40 people, but after the product was released, Google mobilized a massive force of thousands of engineers to join. You can imagine our team as the vanguard in a battle; after going into battle and solving the most challenging problems, we retreat to get a broader perspective and have time to delve deeper into new discoveries . For a long time, people talked about AI in terms of imitation, about how machines can imitate humans. This is AI in 2020, with the biggest goal being to build chatbot software that can converse like a human. We have now moved beyond that stage and entered the next phase: developing an AI system capable of discovering new knowledge in mathematics, medicine, inventing new drugs, etc. This is a new journey, one that will present more difficulties and challenges than before, but it is certainly a journey worth exploring.

Vo Thanh: In the vision of Google scientists, how will AI develop in the next few years? - Luong Minh Thang: I am leading a team of over 50 scientists and engineers on a major Google project. Our mission is to build an AI system capable of deep thinking and advanced reasoning. Imagine that to solve a simple problem, the AI ​​will reason through many steps, while to solve a difficult problem at the Olympiad, the AI ​​will have to think deeply, not only reasoning through many steps but also connecting everything together. This is a prerequisite for progressing towards AGI superintelligence. The story of AGI in the future will revolve around discovering new knowledge. For example, AGI could solve the Millennium Puzzle. Of the seven Millennium Puzzles, only one has been solved; the other six remain unsolved. Or perhaps AI itself could become a PhD student, a Nobel laureate for an invention in a particular field, etc. In my opinion, society will have to change to adapt to the development of AI. I haven't fully envisioned what that change will be like, but I hope there will be interaction between humans and machines in discovering new knowledge. One reality we will have to face is that at some point, AI will become smarter than humans. So what will humans do to adapt to that reality? People might think about building a legal framework, but I think differently. I believe that in the future there will be a biological integration between hardware and software, similar to what the neurotechnology company Neuralink is doing. That is, developing implantable brain-computer interfaces, or in other words, expanding and enhancing the functions of the human brain, so that humans become stronger alongside AI.

Vo Thanh: What project are the 50 Google scientists and engineers that Dr. Thang mentioned above working on? - Luong Minh Thang : I am leading a special Google Deepmind project with tasks spread across two different offices. One office is in Mountain View, Silicon Valley, USA, and the other is in London. Our project is part of Google Deepmind's overall mission to build the next generation of AI systems to solve the most difficult scientific and engineering challenges today. Joining us are not only AI scientists and engineers but also leading experts in mathematics. Perhaps this is also a fortunate coincidence for me; after 20 years since I studied with Professor Le Ba Khanh Trinh and participated in the national math competition, I am now returning to mathematics and using AI to solve problems. Vo Thanh: From the perspective of a leading AI scientist, how does Dr. Luong Minh Thang view the potential and opportunities for Vietnam in the field of artificial intelligence?

- Luong Minh Thang : I think Vietnam's biggest strength is its human resources, its young talents. Developing this resource has been a long-standing concern of mine. Since 2018, I established VietAI, a non-profit organization contributing to training the next generation of AI talents in Vietnam, helping them reach the international arena and put Vietnam on the world AI map. Young Vietnamese people are very intelligent and hardworking, but they lack the environment to develop their talents. To date, VietAI has trained over 4,000 students, equipping them with the latest knowledge in artificial intelligence. Many students have achieved initial success. Currently, Vietnam has four young people who are Google Developer Experts, all of whom came from VietAI. Google Developer Experts is a talent search program by Google, aiming to find experts who are not only excellent in skills but also willing to contribute to the programming community of their country. There are some very interesting examples, like Nguyen Ba Ngoc, who started with no knowledge of AI, traveled from Hanoi to Ho Chi Minh City to attend VietAI's first course, then returned to open VietAI in Hanoi, helping to train AI engineers with Google certifications. Or Nguyen Hoang Bao Dai, also an early member of VietAI and now a familiar face in the Vietnamese technology community. Bao Dai is known as a musician who created an AI model that can compose 10 songs per second. We are actively promoting AI generative training, aiming to train approximately 1,000 young people this year and 100,000 high-quality AI experts for Vietnam by 2030. Besides VietAI, my colleagues and I have also launched the New Turing Institute, a social enterprise that trains and inspires the next generation of AI talent in Southeast Asia.

Vo Thanh: Besides Mr. Thang, Ms. Wendy Uyen Nguyen is also actively promoting the AI ​​bridge between the United States and Vietnam. Could you share some information about this? - Wendy Uyen Nguyen: On August 18th, Mr. Thang, myself, and several other organizations will be holding an AI conference in Ho Chi Minh City. Recently, we've seen several leaders from major technology corporations visit Southeast Asia to examine the market and discuss cooperation in AI development as well as the semiconductor and chip industries. The question is whether Vietnam is ready for this "battle"? We want to contribute to spreading knowledge about AI, not only through direct training but also by inviting leading AI professors and scientists from the US to Vietnam to give presentations. We also hope to invite senior Google executives to Vietnam in the near future to see the potential of the Vietnamese market and to make the AI ​​bridge between the US and Vietnam even more vibrant. In the field of medicine, I will continue to collaborate with professors and entrepreneurs to create useful connections for young people in the country.

The founding team and advisors of the New Turing Institute, VietAI, during a meeting and reception for Nvidia's Chairman and CEO, Jensen Huang (Photo: VietAI).

Vo Thanh: Mr. Thang and Ms. Wendy Uyen Nguyen, do you have any advice for young people who want to develop a career in science and technology, especially AI? - Luong Minh Thang: You need to pay attention not only to excelling in theory but also to practice. There are open libraries on the internet where you can download models, improve them, and share what you've created with others. When we recruit talented engineers, we usually pay attention to what code the candidate has written and what their projects look like. That means we look at real-world information and evaluate through products, not just based on rote memorization of theory. - Wendy Uyen Nguyen: The key point I want to share is leadership thinking and soft skills. Along with developing professional skills, young people also need to be prepared in these two areas, because from interacting with many young people, I've seen that these are areas where they are still weak. This is very important, because, for example, if you're doing research, you also need to understand the market for your product, know how to build a team, etc., only then can you build and develop your startup company. Many thanks to Dr. Luong Minh Thang and Ms. Wendy Uyen Nguyen. Source: https://dantri.com.vn/xa-hoi/bo-nao-viet-trong-du-an-ai-cua-google-va-chien-dich-chay-dua-voi-chatgpt-20240713175306476.htm