TOKYO, Nov 21 (News On Japan) - Generative AI continues to evolve at astonishing speed, raising both expectations and concerns about a future in which machine intelligence surpasses human capability, and according to Shota Imai, a 31-year-old visiting professor at the Japan Advanced Institute of Science and Technology who is emerging as one of the country’s leading young AI researchers, the question now is how closely AI’s value systems can be aligned with those of humans, as breakthroughs since the arrival of ChatGPT push society into uncharted territory.
The technology is already capable of producing text, images, and video, and Imai notes that generative AI systems can now autonomously operate computers, prepare documents, and even teach themselves to write code, with multi-agent reinforcement learning—where multiple AI systems communicate, cooperate, and distribute tasks—accelerating this evolution further.
Early AI training relied heavily on supervised learning, in which humans label vast quantities of correct data for the system to memorize, but the approach demands unrealistic volumes of ideal answers and cannot cover all possible questions, leading researchers toward reinforcement learning, where AI systems are placed inside environments and learn through trial and error to achieve goals. Imai likens this to the way a child learns to stand or walk—no parent manually instructs muscles or joints; the body learns through repeated attempts—and he explains that his own research involves letting AI loose to freely explore simulated environments, allowing the system to eventually learn “good behavior” without step-by-step guidance.
This shift also reflects the two-stage structure of modern generative AI: large-scale pretraining on massive internet datasets such as Wikipedia, followed by fine-tuning to adjust ethical judgment or adapt models to domain-specific corporate tasks. Imai had originally worked in a separate field focused on reinforcement learning rather than generative models, but when OpenAI released ChatGPT and publicly disclosed the use of reinforcement learning techniques, his work was suddenly thrust into the center of global AI development.
The capabilities of the latest systems are transforming research practices. Imai demonstrates an image in which generative AI is asked to show a room from an angle that was never photographed, and the system plausibly reconstructs unseen surroundings and lighting as if the shot had actually been taken. Such techniques can also be extended into video, including simulations where AI correctly applies physical laws. In one example, a metal ball and a feather are dropped in space, and the model renders both falling at the same speed in low gravity without ever being taught equations for vacuum physics, gravity, or acceleration. Researchers view this as evidence that video-generation models function as “world simulators,” capable of learning the structure of reality purely from data, and Imai notes that this allows robots to train safely and rapidly inside virtual environments without the risk of collisions or human injury.
For researchers, the emergence of ChatGPT was surprising not because the underlying technology was new—it had existed since around 2020—but because early versions, while capable of holding coherent conversations for the first time, often produced absurd or harmful suggestions, such as advising someone with a difficult workplace relationship to “punch their boss.” While this demonstrated remarkable linguistic progress, such systems would have caused immediate public backlash, making post-training alignment indispensable. Through reinforcement learning-based value adjustments, developers gradually shaped the model’s ethical responses; the version now widely used is the result of that intensive tuning.
Still, Imai cautions that achieving reliable ethical behavior in AI remains extremely difficult, as moral judgments lack universally correct answers and can vary according to culture, context, and individual values. Even with advanced techniques, he notes that significant challenges remain in controlling generative AI once its intelligence exceeds human capabilities, a prospect that raises unresolved questions about how work, society, and human-machine relationships will transform as artificial intelligence continues its rapid ascent.
Source: テレ東BIZ
