The Japanese government offices' budget requests for the 2024 fiscal year were announced a few weeks ago (source). While the fact that it is the biggest requested budget ever is noteworthy, the requests also give us a picture of the direction the government is heading in. Given that the current prime minister recently met with Sam Altman, CEO of OpenAI, (source), it comes as no surprise that the government wants to accelerate the AI sector. However, the titles of the projects that the government offices are requesting budget for suggest that they are more heavily focused on generative AI than on AI in general.
The emergence of Midjourney, and the subsequent release of Stable Diffusion last year, that brought the masses the capacity to generate high-quality images, sparked considerable excitement surrounding generative AI. However, most of the discussions back then were more about safety and the dangers of deepfakes than the promise of a new technology. Despite that, there were many attempts at creating a localized Japanese version of generative AI.
When ChatGPT came into light, it generated a lot more buzz than image generation. News coverage was non-stop regarding conversational AI in the beginning of 2023, and there was kind of a race towards general application in the private sector. The government was also eager to get on the AI hype train, with some talks about adopting ChatGPT in administrative tasks (source).
Problems with Generative AI
As a neuro-symbolic researcher, my opinions are biased (I believe in logic more than in magic). However, I believe that the current generative AI technology is not yet ready for real-world applications. Despite the significant investment of time and resources, there has been little tangible innovation in the field. This suggests that there is a significant gap between the current state of the technology and its potential for real-world use.
Most technological breakthroughs follow a pattern where the fundamentals are difficult to figure out, but once they are understood, the details are relatively easy to work through. The current generation of generative AI, however, is completely backwards. The fundamentals of conversational AI, such as transformers and natural language processing technology, have been understood for a long time. The new breakthrough in this field is the scale of the data and the model. However, the details, such as hallucination, seem incredibly difficult to solve. Many prominent AI figures have concluded that conversational AI is just stochastic parrots, meaning that they cannot make any statement that is grounded in facts.
The fundamentals of image generation technology are rooted in generative adversarial networks (GANs), which have been around for a few years. The new breakthrough, however, is the ability to use language as instructions, which ties in to the breakthrough from language models. However, the details, such as the number of fingers and the ability to place objects in specific positions, still seem impossibly difficult. Even as problems like fingers are being worked out, there will always be some details of an object that is inconsistent with what is in the real world.
Many people believe that the solution is to combine these technologies with logic and reasoning (which is what I'm currently working on). With proper reasoning ability, AI models should be able to produce results that are consistent with our expectations. Unfortunately, no one yet knows how to do this effectively.
Japan's Ego and AI
So why is the Japanese government, a notoriously conservative organization (source 1, source 2), rushing to adopt ChatGPT, a new and emerging generative AI technology with many known problems?
One possible explanation is the shadow of Japan's formerly world-leading semiconductor industry. The semiconductor industry's advancement in Japan, along with other factors, led to the country's booming economy of the late 1980s and early 1990s, a period in which many Japanese seniors take great pride on.
Since then, however, Japan has lost its lead in semiconductors, failed to take the lead in the information and technology boom of the 2000s, and has fallen behind in electric vehicle technology. Its current leads, such as high-speed railways and hybrid engine technology, are not seen as globally relevant or as exciting as others. This has led to a general sentiment that the failure to produce relevant world leading technology being the main reason of the current stagnant economic.
With many Japanese who lived through the "great Japan" period of the early 1990s now leading important branches of government, it is no wonder that they feel a strong urge to restore their country's former glory. Add to that the threat of rising nations like China and India, there is also a strong political will to get ahead of everyone else.
This, I believe, is the driving force behind Japan's current strong motivation to invest further in generative AI.
Wrong Investments May Prove Fatal
Even though there is a strong sentiment that Japan is not doing well, I still think that it is incredibly wealthy and prosperous. Yes, there are obviously issues like an aging society and decades of plateauing growth. However, the investments made when Japan's economy was booming were largely appropriate and have continued to pay off.
Of course, there is nothing wrong with wanting to make huge investments in new areas when you are doing well. However, there is a risk of throwing away what you already have in pursuit of something that might not turn out as good. Even though the Japanese government has not yet announced an all-out bet on generative AI, there is still potential for harm if this bet turns out to be bad. One can look at SoftBank Capital to see examples of bad Japanese investments.
As a researcher and technologist, I hope that the current hype of generative AI will soon pass. The recent years of hype trains that have led to huge capital investments (like cryptocurrency, for example) have turned out to be a huge waste of resources. Even the current trend of SDGs, while trying to do something good for the planet, is not as obviously good as it seems.
If I could have a say in any of this, I would hope that resources would be directed more towards the present issues (shrinking population, weakening of rural communities) instead of attempting to chase something that is uncertain and in the indefinite future. Unfortunately, my status as a foreigner and as a neuro-symbolic researcher, means that my best course of action, perhaps, is to come up with a truly logic grounded AI technology.