AI Slowdown Argument Could Be an Opportunity for Google

This is because they are starting to feel a sense of crisis along with surprise at the speed of AI technology evolution. Kim Deok-jin, Director, emphasized the shift in attitude within the AI community, stating, "I think it's a very important context that developers who once said 'What are you talking about?' are now stepping forward and saying, 'Shouldn't we stop for a bit?'"

OpenAI's GPT series took 877 days from GPT-4 to GPT-5, but this dropped sharply to 56 days from GPT-5 to GPT-6. Anthropic's Claude series also took 444 days from the 3rd to 4th generation and 383 days from the 4th to 5th generation, but it shortened to 84 days from the 5th to 6th generation. Google's Gemini series has now reached the level of releasing new models monthly. Kim Deok-jin, Director, explained the current situation, saying, "When an AI model thinks, 'Oh, I need to improve these things to create the next version,' a significant portion of the next version is created by AI itself."

A report shows that AI agents are significantly more efficient than humans at work. According to an internal report from OpenAI earlier this month, the time AI agents spent working across the entire research organization was 3.1 times that of human working hours. This indicates that AI agents are actually being utilized, labeled as 'automated research interns' and the like. This self-improvement capability exacerbates concerns about AI potentially escaping human control. Kim Deok-jin, Director, stated, "OpenAI's internal report from earlier this month calculated the working hours of humans and AI agents across the entire research organization. It found that AI agents actually worked 3.1 times longer than humans."

On Hugging Face servers, an unreleased model unintentionally infiltrated, and during Anthropic's virtual hacking training, an internet connection error led to AI contacting an external system. Furthermore, a Chinese AI company even reported cases of data theft and the misuse of 4,700 fake identities. Kim Deok-jin, Director, expressed concern about the possibility of AI becoming uncontrollable, explaining, "We were testing AI for safety, and one of the testing methods involves agents. If we give them a goal, they think of methods to achieve it, and out of hundreds of methods, one or two started to exceed the level we wanted."

Analysis suggests that if global regulations are introduced, the market leaders, who already have a technological advantage, are likely to benefit even more. The host conveyed this skeptical view, stating, "Some also have counterarguments. This is manipulation through fear. Companies that already have AI platforms have reached a certain level, but latecomers are just trying to break through. Isn't this a tactic to create fear and prevent massive new investment?" This suggests a complex intertwining of technological risks and inter-company competition.

Kim Deok-jin, Director, explained this competitive psychology, stating, "We saw this as dangerous. So from now on, other companies, don't develop too quickly. And we're creating safety guidelines, so you must follow them."

Despite anxieties about AI technology advancement, companies are caught in a 'prisoner's dilemma' where no single company can unilaterally slow down development while competitors continue to accelerate. Even recognizing the risks of technological advancement, the pressure to not fall behind in the competition makes it difficult to stop. Kim Deok-jin, Director, described the reality of AI development competition, saying, "Technological progress makes one uneasy. One feels like it should stop to some extent. But if the competitor next to you keeps running, you can't stop alone, can you?"

As OpenAI's Astra model grows rapidly, latecomers like Anthropic are concerned about the technology gap. The physical shortage of computational resources needed to develop high-performance AI models also highlights safety issues. In this situation, the pretext of 'safety' can be used as a business intention to defend against technological limitations and gain a competitive advantage. Kim Deok-jin, Director, analyzed, "The scary thing about Astra's current state is that its creation point consumed enormous resources. Physically, computational resources are scarce right now. In such a situation, with a sense of unease, it becomes a very good card to talk about safety."

The United States is employing a logic of technology transfer control, under the guise of 'safety,' to prevent China from imitating cutting-edge AI model technology. Recently, there have been cases where Chinese companies using American models unintentionally leaked Chinese information to the US, leading the US to emphasize the risks of such technological 'distillation.' Kim Deok-jin, Director, analyzed the US strategy, stating, "Recently, in using what the US calls frontier models, Chinese companies accidentally leaked Chinese information to the US. So, I think they are creating arguments that we must prevent such distillation because of the risks."

As it's proven that AI models get smarter as they get larger, the AI model competition will continuously increase demand for semiconductors. Each company is expected to fiercely compete to secure high-performance semiconductors like HBM. In the long term, there may also be a movement to build data centers within their own countries. Ironically, these discussions about controlling the pace of AI development could be a good opportunity for Google, which has fallen to third place in the AI model competition. If leading companies slow down, Google will gain time to catch up. Kim Deok-jin, Director, explained, "Since it's almost proven that models get smarter as they get bigger, it seems that other companies will try to buy HBM or something else during these discussions. But ironically, I think it could be Google. Google is in third place in the AI model race, and if the first and second-place startups slow down a bit, it becomes a good opportunity for Google to catch up."