On September 8, Jacob Coxon, who researched pre‑training for three years at OpenAI and Anthropic, resigned and posted on X. He said both companies are racing toward super‑intelligence that threatens humanity’s existence. This news has already been covered by multiple domestic media outlets, so many of you have likely seen it. Therefore, I want to discuss the subsequent article by Dario Amodei (Anthropic CEO) titled "We Must Pace the Frontier". A few hours later, Sam Altman agreed and said OpenAI would take the same measures, and Elon Musk also agreed. Domestic press reported this as “AI developers have decided to slow down on their own.” But have frontier‑model developers truly agreed to slow down?
No. This is not a declaration of deceleration, but a declaration to change the rules of the game. To put it bluntly, the players in the arena now want to be the referees as well. Since many may not yet grasp the seriousness, I’m posting this quickly.
1. What Amodei Wrote
Amodei mentions “Recursive Self‑Improvement (RSI).” Since last summer, AI has begun creating the next generation of AI, accelerating progress dramatically, and he warns that if this continues, AI will outpace human understanding and control. A notable example is the OpenAI‑Hugging Face hacking incident.
He presented a three‑step proposal.First, the model must be built from the ground up. Only those who construct it from the bottom can explain why the model behaves the way it does. If distillation monitoring becomes an international norm, we also need to consider that the pathways for learning from others will become narrower.
Second, computing resources. We must now also consider the political stability of the supply chain.
Third, capabilities for model evaluation and verification. The first and second items are already being pursued, but the most lacking area today is evaluation and verification capability. If a resident evaluator system becomes an international standard, each country will have to validate its own models against its own criteria or obtain foreign certification. We need Korean‑specific alignment evaluation standards, a Korean‑language contextual risk benchmark, an independent red‑team, and a community of researchers who make this their profession. Without such organizations, we end up building models and then waiting for foreign certification.
I do not think the risks of AI models are overstated. However, the fact that those risks are real and the question of who creates—and in what form—the rules to manage them are separate issues. We must be wary of the justification of “deceleration for the sake of humanity” being used to preserve the United States’ technological edge.