El 8 de septiembre, Jacob Coxon, quien había investigado el preentrenamiento durante 3 años en OpenAI y Anthropic, renunció y escribió en X. Ambas empresas están corriendo hacia la superinteligencia, poniendo en peligro la vida humana. Esta noticia ya ha sido presentada en varios medios en el país, por lo que muchos de ustedes ya deben haberla visto. Por lo tanto, me gustaría hablar sobre el artículo "We Must Pace the Frontier" de Dario Amodei, CEO de Anthropic, que salió a continuación. Unas horas después, Sam Altman expresó su acuerdo y anunció que OpenAI también tomaría las mismas medidas, y Elon Musk también estuvo de acuerdo. Los medios de comunicación nacionales informaron que "las empresas de desarrollo de IA han decidido ralentizar su ritmo". Sin embargo, ¿acaso las empresas de desarrollo de modelos de vanguardia realmente han acordado ralentizar su ritmo?
No, no es
First, the model must be built from scratch. Only those who build it from the bottom up can explain why this model works in a certain way. It should also be considered that as distillation becomes an international norm, the path to learning from others will narrow.
Second, computing resources. Now, the political stability of the supply chain must also be considered.
Third, the capacity for model evaluation and validation. The first and second points are currently being promoted, but what is currently lacking the most is the capacity for evaluation and validation. If a resident evaluation system becomes an international norm, each country will have to validate its own model according to its own standards or obtain certification from other countries. A Korean-style alignment evaluation standard, a Korean context risk benchmark, an independent red team, and a group of researchers who make this their profession are needed. Without such an organization, the situation will be one of waiting for certification from other countries after creating a model.
I do not think that the risks of AI models are exaggerated. However, the fact that these risks are real and the issue of who creates the rules to manage these risks and in what form are different problems. We must be cautious about the pretext of slowing down