Jakub Pachocki, Chief Scientist at OpenAI, discussed the rapid progress of AI and the accompanying importance of ensuring safety in a post on the company's blog.

According to Pachocki, since signs emerged in mid-2023 that the training of inference models could be scaled, AI has begun to exert a significant influence on fields such as scientific research, computer operation, and cybersecurity. He expressed the view that if the current speed of progress is maintained, it could lead to "Recursive Self-Improvement (RSI)," where AI accelerates its own development.

In response to these advancements, Pachocki warned that "extremely cautious handling is required." In particular, he emphasized the importance of "alignment" research—controlling AI to act in accordance with human values—and "Chain-of-Thought monitoring" to oversee the AI's thinking processes. He acknowledged that as models become more sophisticated, the difficulty of monitoring increases, and suggested the necessity of adjusting the pace of development until confidence in safety can be established.

Source: