Focusing on “Researchers’ Big Picture,” this series explores the future societies they envision through their work. How does their research intersect with our lives? Their visions might provide valuable insight into the path ahead for all of us.
Generative AI is quickly becoming a staple in our daily lives—whether for research or just trying to find the right words. Its speed of evolution is remarkable everyday. We asked Mr. Asano about a future where the relationship between human and generative AI is even closer.
Interviewing:
- Name: Hikaru Asano
- Affiliation: PhD Student Graduate School of Arts and Sciences, The University of Tokyo
From Hardware to the Information
▶︎Your research focuses on AI, a field at the forefront of technology today. What motivated you to enter this field?
I originally entered a National Institute of Technology (KOSEN) to participate in the “Robocon” robot contest. As a mechanical engineering student, I studied metal processing and was in charge of structural design for the team.
However, when I began preparing to transfer to the University of Tokyo, I became interested in AI, which was a hot topic at the time. I decided to specialize in AI rather than robot design or control programs. That was likely the starting point of my career as an AI researcher.
▶︎There are many types of generative AI today, such as ChatGPT. What specific aspects of AI are you interested in and researching?
AI is a broad term. It refers to technology that allows computers to perform intellectual tasks, such as understanding language, reasoning, and problem-solving—things that only humans could do before.While many people today imagine “Generative AI” when they hear the term “AI”, Generative AI is just one of the kinds which generate content like text, images, videos, and music. Within Generative AI, there are Large Language Models (LLMs) specialized in text generation and ChatGPT and Gemini fall into this category. My reseach focuses on these LLMs. The hierarchy looks like the diagram below.
▶︎I see how they are categorized now. Among all these different types of AI, what made you choose LLMs as your research topic?
Humans input information like language, think about it, and then output their thoughts back into words. LLMs follow a similar process by generating verbal responses to questions. I focused on this field of LLM because I find the similarity between human and AI thinking processes to be very exciting.
Will AI Start Improving Itself?
▶︎The idea that LLMs function like humans is very interesting. I’d like to ask you two questions : what are you working on right now, and what are your interests and vision for the future of your research?I have two questions for you.
I am currently researching ways to overcome the challenges within LLMs. As part of this, I am interested in a new relationship between humans and AI. Specifically, I am interested in “personalized LLMs” that adapt to the unique thinking and needs of each individual user.
▶︎I see. First, let’s start with your current work. I am also amazed by how quickly LLMs are improving, but what are the challenges that need to be solved?
One of the issue is the skyrocketing cost of development. LLMs have evolved by accumulating knowledge from vast amounts of human-written text and learning from human evaluations of which answers are better. In the early stages, LLMs could only handle tasks that a human could do in about three seconds and only required simple data. However, as they become more advanced, models require expert-level examples for training and its need for specialized data is driving up costs rapidly.
https://metr.org
▶︎That’s a perspective we don’t often hear—the cost of creating training data was rising alongside the performance of LLMs. It makes sense that your research is focused on addressing this very bottleneck.
I am focusing on the idea of “self-evolution,” where LLMs generate high-quality training data themselves without human intervention. My research explores whether this self-evolution is possible in areas where LLMs currently struggle. This method involves assigning different roles to multiple AIs—for example, one acts as the “solver” and the other as the “evaluator.” By working together, they can improve their overall performance. While we have confirmed that this approach increases accuracy in specific areas, the barrier to autonomous improvement remains high.
It is similar to how difficult it would be for a person with no knowledge of chemistry to raise their test score from 60 to 80 through self-study alone. The key to overcoming this barrier is a method called “Weakly Supervised Learning.”
▶︎Could you tell us what kind of learning method that is?
Previously, the data used to train LLMs had to be perfect and error-free. Weakly Supervised Learning, however, is a method that allows AI to learn the correct answers statistically, even from incomplete data that contains ‘noise’—for example, a collection of ‘dog’ images where 20% are actually cats.This means that even without a perfect textbook, the AI gains the ability to infer the right answer from flawed materials. When we apply this to LLMs with different roles, even if the AI initially makes mistakes 40% of the time, it can gradually turn those experiences into valuable learning resources and achieve step-by-step self-evolution.
The New Path for AI and Us
▶︎As AI evolves and becomes more advanced, it will certainly be convenient to leave everything to AI. On the other hand, many people are concerned that it might take away our jobs.
In the past, AlphaGo defeated professional players because the game of Go has a clear metric for winning. AI excels in areas with defined goals, such as competitive games or academic writing.
However, creative writing is not about winning or losing. If we compare Ryunosuke Akutagawa and Soseki Natsume, we cannot say who is better because it is a matter of personal preference. I also believe that planning a research project is difficult for AI. Furthermore, areas that depend heavily on human preferences and emotions, as well as tasks requiring schedule adjustments and management, are things only humans can do. I consider these to be essential domains that will remain in human hands in the future.
A Future Built on Networked Personal LLMs
▶︎This means we can coexist with AI by finding the right balance between the areas humans should lead and those we can delegate to technology.
Currently, large-scale LLMs are the mainstream, but they often return an overwhelming amount of general information at once. This has become a major bottleneck; the burden on users to review and refine these answers is only increasing.
To lighten this load, we don’t need a single giant model, but rather a group of smaller models tailored to each individual. I believe AI research will shift from the current trend of being “larger and more universal” toward being “more personal and diverse.” My hypothesis is that this shift will open up an entirely new field of research.
▶︎So this is that new direction for the human-AI relationship you’ve been fascinated by. Could you tell us more about this hypothesis?
Based on this hypothesis, I am conducting research on a personalized LLM that mirrors the preferences and thought patterns of each user—much like a “Virtual Hikaru Asano.”
For example, “Virtual Hikaru Asano” whose decision-making is similar to me can handle negotiations with business partners on my behalf. However, since it would act strictly as my proxy, it would not make final decisions on its own if it senses a conflict with my values; instead, it would refer the matter back to me. This would ensure that no unintended decisions are made, leaving me in control of the final judgment.
If this is realized, the AI could handle multiple tasks of mine simultaneously. Since I would only need to make decisions on the specific cases the AI refers back to me, I believe we could be liberated from the mentally draining workload of daily task management.
▶︎This could help people do much more on their own. However, I am not used to big business, so I might not fully understand it yet. Are there any simpler examples from daily life?
For someone who’s too busy to keep up with everything, the AI can act as a personal shopper—It can find and buy things you want at the perfect time when you are too busy. It can also look for items you might like and tell you when is the best time to buy them.
It could also help you utilize an app like Mercari, like price negotiations, or listing items you don’t need anymore but just staying at home, or manage the transaction/messages until the sale is finished.
This creates a chance to sell items that were just sitting in your room. It is a convenient way to clean up your space without much effort.
▶︎It is like having a “clone” of yourself, which is great. Besides doing boring tasks for you, it seems like it would be helpful in many other ways.
When you look for a book at a bookstore, you might find another interesting one by chance. This kind of “serendipity” often happens in real-world. Online shops have recommendation features, but they often show things you already bought or things you don’t want right now.
A personalized LLM could be different. It understands your interests and how you think, so it can suggest things that are unexpected but perfect for you. This helps you find great books you didn’t plan to buy. It also helps the bookstore sell more books, so it is a win-win situation. Most importantly, you can enjoy the excitement of looking around, just like in a real bookstore or a flea market.
▶︎I feel that online shopping has fewer chance encounters compared to real stores. The excitement of “happening to find something” is definitely an important point.
What’s even more exciting is when these personalized LLMs become networked. It will allow us to engage with a vast range of diverse values that we might never encounter in the physical world. I believe this will be a powerful tool to help broaden our perspectives and lead us toward true fulfillment.
This isn’t just for consumers; it applies to researchers as well. Even when you’re working on a project alone, your personalized LLM could bring in fresh perspectives from the LLMs of other researchers. This opens up the possibility of finding unexpected breakthroughs and sparking entirely new ideas and discoveries.
By connecting personalized LLMs like this, I believe we can create an environment that respects each person’s thoughts, needs, and opinions. Eventually, this will spread to companies and the entire society. When the happiness of each individual connects and grows, it will lead to the benefit of society as a whole.
▶︎It’s been nearly a decade since I watched Ex Machina, a film that explores the complex relationship between humans and AI. Ever since then, I’ve felt a deep sense of awe—and perhaps a little fear—toward the evolution of AI. Talking with him today has brought those feelings into even sharper focus, but at the same time, I’ve become genuinely fascinated by the potential of highly personalized LLMs.
While we might need to implement something like an ‘ethical circuit,’ I can see a future where virtual LLMs truly grasp our intentions and support us. It feels like a world where everyone can find more peace of mind, leading to a much more harmonious society. I’m truly excited to see where your work takes us. Thank you so much for your time today.
Value Exchange Engineering PR, Kawanaka