Open AI's top scientist urges speed control of AI risk warnings... The report was mixed with ‘acceleration of research’

(San Francisco = Yonhap News) Correspondent Kwon Young-jeon = The chief scientist of OpenAI, the developer of ChatGPT, warned of the risks of artificial intelligence (AI) technology and urged the industry to control the pace of development across the industry.

However, OpenAI's separate report released on the same day showed mixed results, claiming 'acceleration of research'.

Yakup Pahotsky, OpenAI's chief scientist, expressed concern in an article titled 'An Alien Mind' posted on the OpenAI website on the 6th (local time), saying, "No one is preparing for the consequences that will occur if machine intelligence continues to develop rapidly."

He introduced modern AI as “grown rather than designed,” and pointed out that it is difficult for humans to fully understand how AI works.

It is difficult for humans to completely control AI because it grows like a living organism rather than being created according to detailed human blueprints.

In fact, it was emphasized that safety measures such as 'chain of thought' (CoT) monitoring applied for AI monitoring are gradually deteriorating as the model becomes more advanced.

He refuted the AI ​​industry's claim that a defense system can only be built if AI development continues, saying, "(Building a defense system) should not be used as an excuse to justify reckless actions."

He went on to say that there is not yet a single research institute that has sufficiently solved the 'alignment' problem that allows AI to reject actions that pose a threat to humans, adding, "I hope and hope that voluntary speed control will become common until a common safety device is in place."

He added, “International cooperation on AI development must become a top priority for governments around the world.”

CEO Sam Altman also cited this article by top scientist Pahotsky on social networking service (SNS)

CEO Altman has consistently shown a skeptical attitude toward the theory of speed control in AI development, which has been raised several times over the years, but after the recent 'Hugging Face' hacking incident of the GPT model, he changed his opinion forward and supports the theory of speed control.

However, the 'Research Acceleration' report published separately by OpenAI on this day attracts attention as it contains somewhat contradictory content.

In this report, OpenAI stated that it has achieved its goal of becoming an 'automated AI research intern' who performs its own AI improvement and coding, and presented a new goal of developing an 'automated AI researcher' by improving this by March 2028.

It was also revealed that, based on 8 hours of work per day, the AI ​​agent shows a workload equivalent to 3.1 days when a human researcher works per day.

The report claimed that such automated AI research could actually help solve the 'sorting' problem, expressing a conflicting opinion with top scientist Pahotsky.

However, the report took a step back, saying, “This does not mean that rapid ‘iterative self-improvement’ (RSI) (of AI) is something that must be pursued,” and “whether and how to pursue this should depend on democratic choice.”