Обратная разработка GPT-6 Cyber

Оригинал

Reverse Engineering GPT-6 Cyber

OpenAI Just Shipped GPT-6 Cyber. Here's How It Actually Works, Reverse-Engineered From the System Card. OpenAI's last two cyber models mostly proved that frontier models could find bugs. GPT-6 Cyber is about whether you can trust one to stay inside the lines while it does. One caveat before we start. Nobody outside OpenAI has the weights, the parameter count, the architecture, or the training data, so "reverse engineering" here means reconstructing the system from what OpenAI published: the sy