MUFG has invested in and begun collaborating with Sakana AI, a generative AI startup. Why did this traditionally conservative financial institution choose to partner with an AI startup? Morito Emi, Head of the Digital Strategy Division, Ryosuke Okuta, Project Manager at Sakana AI, and Nobutake Suzuki, CEO of MUFG Innovation Partners (MUIP), discuss the challenges and prospects of using generative AI (GenAI) in the financial sector.
The potential of collaboration between organizations with complementary strengths
Sakana AI was founded in Japan in 2023, and garnered attention not only for its technology but also for the diverse backgrounds of its founding members. CEO David Ha was a former Goldman Sachs trader who later worked on machine learning research at Google. CTO Llion Jones, a former Google employee, played a key role in the development of the deep learning model ”Transformer,” widely used in GenAI technologies today. COO Ren Ito previously worked at the Ministry of Foreign Affairs and Mercari before leading the business division of Sakana AI.
A notable aspect of the collaboration between these two organizations with different cultures is their mutual understanding of each other’s strengths. Okuta appreciates MUFG’s approach to work and the careful use of language, which reflects the corporate culture developed over the company’s long history.
Common challenges often seen in collaborations between large companies and startups, such as differences in speed or excessive caution, are notably absent. Okuta adds, “The characteristics of a company are deeply ingrained over time, so it’s better to leverage those characteristics rather than change them.”
In MUFG’s case, “They communicate in a very flat and honest way,” shared Okuta referring to the transparent communication between the teams. This high level of transparency has become a key factor in the success of their collaboration despite the cultural differences between the organizations.
However, as practical implementation begins, new challenges are likely to emerge. One particularly important issue going forward will be building relationships with employees who may resist using AI. This will be a key focus in the coming stages of the collaboration.
Practical challenges of implementing GenAI and tech solutions
One of the biggest challenges in implementing GenAI is evaluating the return on investment (ROI). “When securing a budget, it’s typically necessary to clearly explain the expected return,” says Emi. With AI projects, however, it is difficult to make predictions about both feasibility and effectiveness.
Nevertheless, investing in new technologies is unavoidable. To address this, MUFG has implemented special measures for its budget approval related to GenAI.
We’ve secured a considerable budget for highly important, long-term initiatives. (Mr.Emi)
This is not an unconditional investment because there is an opportunity cost. In exchange for special treatment in the budget, other budgets are being cut. Maintaining discipline while moving forward in an agile manner requires striking a delicate balance.
Another major challenge in implementing GenAI is handling unstructured data. In response to this challenge, Okuta presents new possibilities.
Traditional machine learning required massive datasets for training. Recently, methods have been developed to add new information to foundational models that already possess a common sense understanding of the world. (Mr.Okuta)
This approach significantly reduces the amount of data needed compared to traditional methods.
However, this challenge has not been completely resolved. Okuta points out that “the most difficult part is digitizing paper data left by humans, such as handwritten notes from phone conversations.”
In addition, the issue of hallucination, where the AI generates fabricated information, presents technical challenges in both knowledge and reasoning. Okuta explains,
There is a lot of ambiguous information in the world. The sources from which large language models (LLMs) learn can include this ambiguity.
As the AI performs more inferences, something that was initially just a possibility can be mistaken for a fact, or information provided earlier can be lost.
In response to these issues, Sakana AI proposes a validation system involving multiple AI agents.
One agent’s answer is checked by another agent, and the results are integrated to verify the veracity. (Mr.Okuta)
“In the case of humans, a fourth checker might not take the verification as seriously, but AI will definitely do so,” he adds, highlighting the benefits of this approach. The ability of AI agents to reliably perform tasks equivalent to human double or triple checks is particularly valuable for industries like finance, where precision is critical.
Sakana AI is exploring a combination of technical solutions and practical approaches in pursuit of a different approach from traditional machine learning suggesting a different AI utilization in financial institutions.
Exploring the sustainability of GenAI development
Concerns surrounding the sustainability of GenAI development are growing, particularly regarding computational resources and energy consumption. A Goldman Sachs report noted that GenAI development requires enormous GPU resources and electricity. In fact, in the United States, new nuclear power plants are being built at twice the pace of previous years due to the demand generated by AI.
In response to this situation, Okuda said,
The fact that this industry is using so much energy, money, and resources is simply due to the huge expectations from society.
However, he acknowledges that this situation is not sustainable in the long term.
Sakana AI is taking a different approach in tackling this challenge.
We are deliberately not building massive models using enormous amounts of electricity at a global level. (Mr.Okuta)
He questions whether creating a Japanese language foundational model in Japan from scratch is truly helpful. This reflects skepticism about spending time and resources on developing huge models when new techniques and technologies are emerging rapidly to displace existing models.
Instead, Sakana AI focuses on developing technologies to efficiently combine existing models.
We are developing a technology that combines the latest models to create high-performance models with new characteristics. This can be achieved using a single GPU within a day’s computing time. (Mr.Okuta)
One of the reasons MUFG invested in Sakana AI is its efficient approach. MUIP’s Suzuki likens Sakana AI to a DJ. They arrange and play various music, creating something everyone can dance to. Instead of creating music from scratch, they take the best parts and create something new,” he says.
This approach aims to balance sustainability in GenAI development with business goals. It requires no massive investments and, with high technical capabilities, can differentiate itself. Sakana AI’s strategy, which focuses on efficiency and the art of combination, may become a model for future AI development.
A new vision for the industrial structure in the AGI era
What will the financial market look like in the era of AGI (Artificial General Intelligence)? Emi offers an intriguing outlook: “If everyone starts using the same AI to perform their activities, foreign exchange and stock markets will no longer work.” This is because transactions would not be able to take place between entities that all operate based on the same logic.
Therefore, he envisions a world where people each have their own AI that understands their data and risk appetite, and they compete with it. He predicts a dramatic increase in the efficiency and productivity of financial markets.
Moreover, customer relationships could change significantly. He predicts that customers will begin conducting financial transactions using AI agents. In such a scenario, financial institutions will naturally respond through AI as intermediaries. This could change the way current UI/UX design is conceptualized. “Instead of designing interfaces for humans, we need to refine APIs designed for AI,” he says, suggesting a fundamental shift in the way financial services will be designed in the future.
On the other hand, when looking at the entire industrial landscape, the most significant changes are expected to occur in the IT industry. “The programming industry is going to undergo a major transformation,” points out Okuta. There have already been reports in the U.S. of a drastic decline in demand for programmers.
The speed and performance of coding have drastically improved with the evolution of GenAI. “Programming itself may no longer be needed. The ability to articulate what you want to create might still be required, but the actual coding and debugging work will increasingly be replaced by computers,” Okuta predicts.
Such changes will also impact the structure of business itself. “There is already a trend in the U.S. where entrepreneurs are using AI to handle everything, from programming to running websites, all by themselves,” he observes. Okuta notes that a lot of the work that was previously outsourced to employees in startups is already being replaced by GenAI, including ChatGPT.
However, this does not necessarily indicate a decrease in the value of human workers. In fact, the automation of intellectual processes requires a combination of human skills and GenAI. As GenAI continues to evolve, humans will be required to shift toward clearly articulating what they want to create or what they want to achieve and collaborating with AI.
This shift represents a major turning point in the kind of talent required across industries. In particular, the characteristics of GenAI that dramatically enhance the productivity of small organizations and individuals hold the potential to bring about fundamental changes to traditional organizational structures and work styles.
What emerged from the session was that the transformation brought about by GenAI will go beyond mere business efficiency and automation, to more fundamental changes. It holds the potential to fundamentally alter traditional industrial structures, such as the creation of new markets through AI-to-AI interactions, the redefinition of the roles of humans and AI, and the liberation from corporate size and organizational structure.
The technological evolution of GenAI shows no signs of stopping. However, the success or failure of its utilization ultimately may depend on human imagination and execution. As seen in the collaboration between Sakana AI and MUFG, organizations with different cultures and backgrounds can create new value by leveraging each other's strengths. This process could hold the key to sustainable innovation in the GenAI era. Indeed, such efforts are beginning to take shape in Japan’s financial sector.
