“I hope that AI becomes a partner to people rather than something that replaces or eliminates humans. I believe that in ten years AI will be more reliable, allowing us to trust it with many tasks. If robots become common, that is a good thing because they save time on labor like washing dishes. Today, language models still hallucinate, so we double-check their work. In the future, I hope we can rely on AI without constant verification, coexisting with it and becoming far more productive together.” - Kelvin Chan, AI researcher at Google
“One year ago, I did not expect image editing or image generation to become this good. There is always something new in this field, which is why I stay excited working in AI at Google. We do not know where the limit is, and that uncertainty drives me every day. Ironically, I have no artistic sense at all, yet I work on images. When I take photos for friends, they usually retake them because I cannot frame good shots. That became a motivation for me to work on image editing and generation, because now I can take a random photo and ask AI to adjust the angle or make it more artistic. It is genuinely useful, and it saves me from my friends’ sarcasm.” - Kelvin Chan, AI researcher at Google
“Google encourages us to use the AI tools we build because using them is the fastest way to understand what people need and what can be improved. When we build the tools and then use them ourselves, we learn how to refine them and create better models for the public. This feedback loop makes the work more effective and is what makes this an exciting moment to be working at the frontier of AI.” - Kelvin Chan, AI researcher at Google
Kelvin Chan, an AI researcher at Google, joins Jeremy Au to unpack his unconventional path from mathematics in Hong Kong to applied AI research across Singapore and the United States. They explore how AI research differs from traditional academic work, why iteration and results often matter more than theory, and how scale has transformed research culture from small experiments to highly collaborative, compute-heavy systems. The conversation covers the rapid evolution of image and video models including Google’s Nano Banana model, the push toward world modeling and embodied AI, and how AI tools are reshaping daily productivity for engineers. Kelvin also reflects on choosing AI in 2018 before it was mainstream, and why he believes the long-term future lies in AI as a trusted partner that augments human work rather than replaces it.
03:18 Image processing redirected Kelvin away from finance: Hands-on work with visual data revealed a stronger pull toward applied problem solving than abstract financial paths.
06:00 AI research prioritizes iteration over proofs: Progress comes from training models, debugging failures, and refining results rather than deriving formal guarantees.
09:16 Nano Banana reflects Google’s applied AI approach: Large-scale models are used to speed up coding, debugging, documentation, and internal productivity.
11:00 Results matter more than explanations in applied AI: Kelvin focuses on whether models work in practice, not on fully understanding internal neural mechanisms.
16:12 Scaling models reshaped research culture: Moving from millions to billions of parameters forced deeper collaboration and reduced solo experimentation.
20:05 World modeling targets physical understanding: Researchers aim to teach AI how gravity, motion, and real-world constraints actually behave.
26:25 Choosing AI before it was mainstream required risk: Kelvin’s decision to pursue AI in 2018 became the most defining and courageous move of his career.
Jeremy Au (01:01) Hey Kelvin, good to see you!
Kelvin Chan (01:03) Yeah. Thank you so much for inviting me.
Jeremy Au (01:04) Yeah. I think it's such an incredible journey from Hong Kong to Singapore for your PhD to America, working with AI and R&D, so lots to learn. Could you introduce yourself?
Kelvin Chan (01:14) Yeah, so long story short, I was born in Hong Kong and then I studied mostly mathematics in my bachelor's and master's degree. And then after my master's degree, I went to Singapore for my PhD degree in computer science. And then during my PhD, I did an internship at Google in the United States. And then after that, I returned to Google as a full-time employee. I have worked in Google since then and until now.
Jeremy Au (01:41) Yeah. Amazing. What was it like growing up in Hong Kong and how did you decide that you're going to work on mathematics?
Kelvin Chan (01:46) Yeah, I didn't have a very comprehensive plan back at the moment. I liked mathematics back in my high school. And many people told me that because Hong Kong is a very financial city with a lot of financial institutions in Hong Kong, I needed to choose what I wanted to study in the university. Many people told me if you want to work in financial institutions, you need to have good mathematical skills. And then I listened to them. Back in Hong Kong, many people think working in finance is the way to make money. I had the same feeling, and then I just thought, okay, then I need to do mathematics. I knew that back in that time, there was a double degree program which is Mathematics and Information Engineering, and therefore I chose that program. I went to CUHK, the Chinese University of Hong Kong. I studied mathematics, and then I remember back in that time I hated coding.
Kelvin Chan (02:34) So I focused more on the mathematics side. I continued working in applied mathematics. In my master's, I tried to use mathematics to solve some image processing problems. This is the whole journey of why I chose mathematics. At some point, I felt like I didn't like finance that much. Through information engineering and my final year project, I touched base on some of the image processing programs. I felt like I wanted to do something like image processing more than finance. Therefore, I did applied mathematics in my master's degree. Growing up in Hong Kong, I think it's a very crowded city and it's a very beautiful city. The pace is extremely fast. I feel very happy growing up in Hong Kong because it's a multicultural city. I learned a lot of Chinese and Western culture. But at the time, I felt like I wanted to explore something new and therefore I considered doing my PhD somewhere else. First, of course, the US was the place I applied.
Kelvin Chan (03:45) This is another coincidence because remember the second degree in my bachelor's, I studied Information Engineering. My final year project supervisor is actually my PhD supervisor. I knew him back in the FYP, and then I did my master's in mathematics and we didn't have a lot of contact back there. When I tried to apply for a PhD, he gave me the chance to follow him as a PhD student. This is why I started doing computer science or AI since my PhD. Before my PhD, I didn't think of doing AI. Everything started at my PhD.
Jeremy Au (04:25) Amazing. And interesting because the PhD was two changes, right? One of course was you starting to look at AI seriously, even though you didn't like coding.
Kelvin Chan (04:34) Yeah.
Jeremy Au (04:35) And then the second part, of course, is moving from Hong Kong to Singapore. So what was it like discovering both of these things during your PhD in Singapore?
Kelvin Chan (04:42) Yeah, I feel like it's a totally different style. Many people say that Hong Kong is similar to Singapore and in some sense, yes, it's a very small and beautiful city. Everything was new to me and I was studying something new. I was very excited to start something new. I knew many good people in my PhD, both my supervisor and my classmates, and everything was very good. I was very happy to start this journey and know all these amazing people.
Jeremy Au (05:14) And what was it like to start exploring AI during a PhD? What were you working on in terms of projects? Obviously it takes a while, then you have to build out your thesis. So how did you explore and focus on what you were going to do for your PhD?
Kelvin Chan (05:27) Yeah, because in my master's I was doing applied mathematics on image processing. Looking at images is intuitive. For example, when you do image processing, you have a very good sense. If you look at the image, it becomes better, it becomes clearer. You have a very immediate feeling of that. That's why I started my PhD. Before my start of PhD, I needed to learn some basics. I just watched some YouTubes. There are many tutorials on YouTube, so I started watching them. I remember at the start of my PhD, my supervisor asked me to propose a few topics that I was interested in. Because I had some experience in image processing, I just proposed a few topics. For example, I remember there was style transfer. You give me an image and then try to transfer it to another style, which was a very hot, popular topic back in my PhD times.
Kelvin Chan (06:19) And then another one was about super-resolution, which is to upscale an image from a low-resolution one to a higher-resolution one. Because my supervisor has a lot of experience and had some students working on super-resolution, he suggested I start working on super-resolution. I just followed them. They were the main contributors, and then I just had meetings with them to learn how to do research as an AI researcher. I learned a lot from them. I think it took months for me to learn how to do research because it's very different from doing mathematics research. Many people say you only need to have pens and whiteboards and you can do research; maybe for applied mathematics, you need a laptop. But I think doing AI research is quite different in my sense. So I took some time to get familiar with all the processes. When I started my PhD and I tried to have my own projects, I also learned from a lot of different people. This was a very good journey to learn from others.
Jeremy Au (07:16) So how is AI research different from normal research? For example, I'm quite familiar with biological research, right? So it's very simple: this is a drug, this is a target, this is—the mouse doesn't work. That would be biological research. But you're saying AI research is very different. Obviously, the movie version would be like you working on your laptop, there are a lot of screens, there are a lot of science fiction numbers flying around. But can you explain what AI research looks like?
Kelvin Chan (07:45) So may I talk about what I did in my master's using mathematics first? In my master's, I was doing image processing using mathematics. We spent a lot of time investing in those algorithms. For example, I was also working on something related to super-resolution in my master's but mostly using mathematics. We had a lot of algorithms that derived some properties of the image and then we would try to optimize it based on those mathematics algorithms. At that time, I spent a lot of time doing some proofs to make sure that everything I did was correct. At that time, I realized that I'm not a mathematics person. I cannot handle those proofs. So I started looking into something else and then I saw AI. In my PhD times, I realized there are different branches in AI.
Kelvin Chan (08:33) In my direction, I focus mostly on how to train that model because AI, as many people say, is a black box. Comparatively, we care less about the algorithm inside because we can't really describe the whole network. We have a lot of fine details, for example, on what training data we need to use, and then how we need to tune and adjust the parameters of the network so that it trains better. Also, the architecture design. So I spend less time on the mathematical formula itself. I spend more time interacting with the model. For example, I train the model, and then it doesn't look good. I'll go back and then debug to see if there is a bug or if there is a problem with my design. It's more an interaction because we don't have those mathematical formulas here. If you know that your formula is correct, then everything works in that direction. But in AI, you train the network, but you don't really know whether this should work in this way or not.
Kelvin Chan (09:28) So it's more an interactive process for me. I enjoy that process actually, because I'm not a mathematics person. It is an iterative process because it's like an exploration, right? You don't know whether you'll get to the goal, but when you really get to that result, it gives me some happiness. Everything is unknown and it's like working in a maze until you find the solution. I just enjoy this process.
Jeremy Au (09:55) If you are not a mathematics person, then I don't know what I am, because I'm sure your mathematics is way better than mine. I also agree with you because when we are talking about mathematics, it's almost like A to B, B to C, C to D, right? And then all the formulas are there. I think for AI, a lot of people are curious about it because it feels like you put in A, and then the magic happens at D and nobody knows what happened at B or C. That's a quite interesting challenge for a lot of folks. Do you have any ways of how you try to explain it or how you think about it?
Kelvin Chan (10:27) I think in the AI research community, many people are trying to figure out what's happening within the neural network and how we explain this. I would say myself, I'm not that type of person. That's why I'm saying I'm not a mathematics person, because I care about the results more than why it happens like this. I enjoy the result that the images look clearer.
Jeremy Au (10:49) Yeah.
Kelvin Chan (10:50) I had the feeling that I don't even know how I work myself. But it works, right? So I enjoy the weight of getting the result instead of knowing what happened. Maybe I'm not that very pure researcher, but then I'm more in applied ways. I don't care if I don't understand it. I'm happy as long as I know how to use it. So maybe I'm more like a practical person. I don't have that mindset.
Jeremy Au (11:16) Yeah. I love the phrase "I don't know how I work myself." Yeah, no, that's interesting, right? Because you're right. We think we are our brains, we think we're moving everything, and I get a glass of water—I get a glass of water, but I don't really know what's happening in between.
Kelvin Chan (11:29) And yeah, I embrace the unknown, because I know that there are many things that we don't know. Maybe ultimately we need to understand the mechanism to improve the AI models, but at this stage, I just focus on how to use AI to make it better. That's what I'm enjoying right now.
Jeremy Au (11:45) Fantastic. And was it just after things in Singapore that you moved to America? You were doing an internship and working at Google. Can you talk to me about how you found that opportunity?
Kelvin Chan (11:53) Yeah, again, I didn't plan for any of this. My original plan was that I would graduate from my PhD and then maybe I'd find a job in Singapore. But just in the last year of my PhD, I got a chance to apply for the Google PhD Fellowship. Fortunately, I got the fellowship. Along with this fellowship, there is a chance to do an internship at Google. I just contacted some people at Google and a team which was working on super-resolution back in that time, and they were interested in me. They were based in Seattle. I just contacted them and then they interviewed me and they felt I was okay. I joined them in the summer of 2022. This was the first time I went to the US. The first time I arrived in Seattle, I felt like this seemed to be the place I wanted to stay because it's very different from Hong Kong.
Kelvin Chan (12:47) Seattle is flat comparatively. There are not a lot of tall buildings and it's not very crowded. Again, I wanted to try something different because this is totally different from Hong Kong and Singapore. I did my internship there and I was happy in this team. I returned to the same team after I graduated from my PhD. I started as a full-time research scientist at Google in January of 2023. So it's a very different place. But I'm happy that I joined Google in my last year.
Jeremy Au (13:22) Oh, amazing. And I think Google has been doing so much incredible work in AI and everything. Now putting aside obviously all of the research and the models, how do you use AI in your daily workflow? I see you have an Apple Watch. How do you use AI in your personal life?
Kelvin Chan (13:38) Yeah, I think things changed quite a lot too. For example, such as the work nature—in the first year of my Google life, we worked on smaller models. We trained models with millions of parameters. Recently, we trained models with billions of parameters. The working culture is different because when you train small models, you don't need to use a lot of compute. You can freely explore a lot of different things within your team's resources. In recent days, when you train larger models, it's not about a small team. It's a collaborative effort, so everything becomes much greater. You need a lot of collaboration with different people. I think this is a culture change.
Kelvin Chan (14:24) Of course, coding is a very different thing right now. Back in the day, we didn't have a lot of coding agents. In 2023, maybe ChatGPT or Gemini was not that popular at that moment. It was not that sophisticated. So we wrote code mainly by ourselves. We debugged mainly by ourselves. But then, in this year, 2025, whenever you have a bug, you just ask Gemini, "Oh, I got this bug. Can you solve it for me?" Maybe it's not a hundred percent correct, but they give you some suggestions—maybe this is not correct, there is not correct. I'll just go to those places and then I'll try to debug it.
Kelvin Chan (14:59) And then the very straightforward use of AI is to help us save some time in coding. Another AI usage, for example—Nano Banana, the recent model is from us. I can use it to prepare slides for us. Because we need to make a lot of documents to talk about what we are going to do, and then we need to present it to our teammates and everyone else. In the past, we needed to spend a lot of time thinking about how to write that English and also generate some fancy images or pipeline graphs to present our ideas. But right now we have a lot of good tools in Google, either internally or externally. You can just type your work and they generate nice graphics for me, or you just write some very rough English and then they can polish it for me. We save a lot of time doing this kind of presentation so we can spend more time developing our algorithm and focus on our coding R&D.
Kelvin Chan (15:59) I'm not that kind of presentation-type person, so I don't know how to make fancy graphics or nice English. So it really helped me a lot because I can focus more on the coding and the development instead of drawing images. I think this is a really good use. Google encourages us to do this because we build a lot of AI tools, and the first thing we need to do is use them. When we use them, we know what people need and what we can improve. I think this is a very good cycle because we build the AI tools and then we use them, we know how to further improve them, and then we can do better models for the public. This is a very exciting moment for us to work at Google, which is at the frontier of the AI community.
Jeremy Au (16:41) Amazing. I'm just curious because some people think that AI is going to accelerate. Some people think AI will continue being where it is going. And some people think that we're plateauing and all the gains have been made. I'm just curious, where do you feel AI is?
Kelvin Chan (16:56) I think even if we are at a kind of plateau right now, we are far from the limit. We already use it for many productivity tasks. For example, we use it for coding, use it for email, use it to generate fancy images for presentations. But I think we can do much more than that. For example, there is a trend about world models right now. We try to understand the physical world and then there is another direction of embodied AI. We try to build robots using AI. I think we are just at the start of the entire development. We have a very good language model right now; it can have an interactive conversation with you. But for world models and for embodied AI, we don't have any very sophisticated solutions right now. So I think we are at the start and then it will improve in the future.
Kelvin Chan (18:02) But yeah, maybe there are some limitations or bottlenecks at the moment, for example, compute. Because models become larger and larger right now, you need a lot of GPUs or TPUs to train your models. So it becomes harder and harder for us to scale up the models. I think this could be one bottleneck, but I believe that we could have a way to solve it. I am not sure how, but I believe that we will get over this. So I'm quite optimistic about this. I think eventually, maybe what we have right now—for example, the transformer architecture—eventually may not be the best way. So maybe we have better solutions in the future, but I believe that we'll find a way to further improve our technology and make AI more useful to the public. So I'm the optimistic guy here.
Jeremy Au (18:40) So I like the fact that you're optimistic about more applications. Actually, I'm interested in you saying "understanding the physical world." Can you explain to me what that means? What the research direction of that will look like in general?
Kelvin Chan (18:51) Yeah. For example, in the old days, you may have seen some memes about Will Smith eating spaghetti. Two years ago, everything was just very corrupted. It was not looking good. But in just two years' time, when Veo 3 or Sora came out, everything became much better. One improvement is that the quality is, of course, better. Another thing that is improving is understanding the world. Because many people think that when you can generate videos with good physical properties, that means you understand the physical world. Many people are going in that direction. We want to have something that understands the physical world. That means we are not just using it for fun. For example, when you want to generate a video of a person walking on Mars, it's different from walking on Earth, right? Because the gravity is different. That means we want to have something that understands the world because whenever you have robotics or embodied AI, you need something that really understands the world to make it useful in the real world.
Kelvin Chan (19:57) So I think many people are developing something called a world model to try to understand and incorporate some physical properties or allow interaction between the user and the model to make it more practical in the world. Because in the past, the quality was not good and then we generated some weird stuff. Because the model didn't really understand what was happening in the world. Hopefully, in the future, we'll have a better model that understands how the real world works—for example, understanding those laws of physics. Then we can really make a practical AI model.
Jeremy Au (20:34) Yeah. I think it's quite interesting because one of the ways that I watch some AI videos and the way I can tell it's AI is because the weight or the momentum of those objects doesn't feel correct. Like they're holding a cup, but the cup looks a bit too light or moves a little bit too fast if they're holding water in it.
Kelvin Chan (20:48) Yeah.
Jeremy Au (20:49) So there are interesting changes. How long do you think before it's effectively indistinguishable from reality? Images seem solved. They say videos two years ago—Will Smith eating spaghetti—were very obviously fake. One year ago, I think people were using the number of fingers for images to see if it was fake. Now I think they seem to have mostly solved it. So when do you think it'll become indistinguishable from your perspective?
Kelvin Chan (21:13) Yeah. It's really hard to say, but maybe just within a year. As you said, in the past, we just looked at hands to see how many fingers were in the hand. Then you'd know whether it was correct or not. But in one year's time, Google's Imagen can now do it. You already can't distinguish whether an image is real or is a generated image. And people think, okay, for images, maybe editing is a more difficult task. Many models out there cannot do image editing well. But then when Nano Banana Pro comes out, this task becomes much better again. No one expected this, actually. Back in that time, we didn't expect the model would be that good. Maybe we needed some more time to iterate, but suddenly magical things happen and everything becomes much better.
Kelvin Chan (22:00) Some people think it may take longer, but I think maybe one or two years—maybe faster, because I think there are a lot of efforts in this video generation community. There are a lot of open-source models and people can explore and try them out. So I think the progress will be much faster. Of course, compute is one bottleneck because video models are much more expensive than image models. It takes much more resources to train. For example, small labs may not have enough compute to explore a lot in this direction, but I believe that those large companies, such as Google, have a lot and they are putting a lot of effort in this direction. So I believe the progress will be very fast and hopefully, in one year we'll have something that is much better than what we have now.
Jeremy Au (22:48) I love the prediction there because what you're saying is: one year before it gets really good, and probably two years before all videos are indistinguishable. So it's going to be quite interesting.
Kelvin Chan (22:56) Yeah. One good thing is you really don't know what will happen tomorrow because, as I said, I work on images; one year ago, I didn't expect image editing or image generation could be this good. A good thing is that you always have something new in this community, and that's why I still feel excited working at Google and working in AI because we don't know where the limit is. That's what's driving me to work every day.
Jeremy Au (23:20) Do you take photographs as a hobby or something?
Kelvin Chan (23:23) Yeah, ironically, I am a person with zero artistic sense. Yeah. But I'm working with images. For example, when I take an image for my friend, they always take another one because I don't know how to take pictures. But maybe this is the motivation for me to work on image editing and image generation, because I don't need to do it myself now. I can just take a random picture and then I can ask AI to adjust the angle and make it more artistic. I really think this is very useful and maybe this is the extra motivation for me because I don't need to suffer from that sarcasm from my friends. They always think I have zero artistic sense. Now I can just take a random picture, process it with Nano Banana, and just pass it back to them.
Jeremy Au (24:08) Yeah. I think looking toward the far future—if you look at say 10 years out—because right now you're saying one year it will get really good for video, and in two years it'll be indistinguishable. But we look 10 years out, in 2035. Where do you think AI is going to be? Do you think you'll see humanoid robots? What do you think AI will be doing from your perspective?
Kelvin Chan (24:28) I hope that AI will become a partner of people instead of what many people say—that AI will replace humans or terminate humans. I hope this is not the case. I believe that maybe in 10 years' time, AI will become more reliable and we can rely on AI to do a lot of stuff. For example, if robots become very common, then it's a good thing because we can save a lot of time on labor, like just washing my dishes. And I hope that AI will become more reliable because, for example, when we interact with the language model, sometimes they'll still hallucinate. So I hope that in the future we can totally rely on AI. We don't need to double-check everything for AI.
Kelvin Chan (25:11) An analogy is that it's a collaborator. For example, if you are working with someone else on a project, when you ask him or her to do something, you won't expect that you need to double-check whether he or she's doing it correctly. For example, you want your partner to implement code—you won't double-check every line of code, right? So I hope that in the future, AI will become this kind of reliable partner. You ask AI to do something, then you just trust it. You don't need to revise everything for the AI. So I hope that they will coexist and then we can be much more productive with AI and we can do what really needs human flexible creativity. I hope we can split the work and then AI can handle most of the stuff for us, and then we can focus on what's more important for us to further improve.
Jeremy Au (26:02) Yeah. And my last question I have for you is: Could you share a personal story about a time that you've been brave?
Kelvin Chan (26:09) I think the bravest moment for me was to start doing AI. Because when I finished my master's in about 2018, AI was becoming more popular, but it was not as common as it is nowadays. I struggled a bit actually whether I should find a job in Hong Kong, doing finance in Hong Kong, or if I should just try something new in a new place and try some new directions. I'm glad that I chose to do something new. Looking back, I think I'm a kind of conservative person and I didn't expect that things could change that much. So maybe this was the bravest moment for me—deciding to do AI, deciding to study for a PhD in Singapore, and then moving to the United States. Everything would be very different for me right now if I did not start doing my PhD in AI at that moment. Hopefully, I will have better moments in the future, but at this time, that moment would be my best.
Jeremy Au (27:11) I think it's very true that it's a very brave moment because, in 2018, AI was a bit of a winter still.
Kelvin Chan (27:18) It was getting better, but people still didn't fully understand what AI could do. For example, they had image recognition and segmentation, but we didn't have much generative AI. Some thought it was a bubble, who knows. But eventually, we proved that AI is not a bubble, yet at least. I'm glad that I chose that at that moment.
Jeremy Au (27:44) It's very brave because in 2018, I can imagine if you told someone you were a Hong Kong math student who had been thinking about finance, everyone's going to be like, "Is he going to become a quant?" or high-frequency trader? That's how people think you're going to end up with your career. So it's quite interesting for you to build your own career and your own path.
Kelvin Chan (28:01) Yeah. I didn't plan for this, but everything is like it was planned. I hated coding and then I did math because I wanted to be a quant, but then everything changed. I did computer science and information engineering in my bachelor's, and then I met my PhD supervisor in my FYP, and then this became the first opportunity for me to start AI. It's very magical. I didn't plan for it. But then suddenly everything assembled together and it became who I am right now.
Jeremy Au (28:32) On that note, thank you so much for sharing. Let me summarize the two big takeaways. First of all, thanks so much for sharing about your early career and personal journey of discovery. I love what you said—that you're not a mathematics guy, even though you study a lot of mathematics. And then you said, "I'm not an AI guy as well," but you became an AI guy. We love that discovery process where you discovered Hong Kong, Singapore, your PhD, and of course the AI research piece of it.
Jeremy Au (28:56) Secondly, thanks so much for sharing about the difference between normal research and AI research. I love what you said about your training—there's still a lot of research being done on trying to understand what actually happens in between the instruction and the output. And lastly, thanks so much for sharing the fact that you're optimistic that AI is going to continue growing and improving. There may be a bit of a plateau in the short term, but there are lots of applications. I thought it was interesting you talk about some of the research directions, like being better able to simulate the real world as well as some of the challenges around the usage of compute and so forth. So thank you so much for sharing.
Kelvin Chan (29:35) Yeah. Thank you for inviting me.
