The battle for AI models and export controls continues amidst a Jensen Huang-free Nvidia. But who controls the flow of chips also controls the flow of leading frontier AI models. OpenAI reportedly now supports a bipartisan push in the U.S. Congress to restrict China's access to advanced AI accelerators and equipment. But are the additional AI safety concerns genuine, or are they also a convenient way to shore up its position at the top of a fast-moving industry? We speak to experts like Max von Thun, director of Open Markets Institute Europe, to find out more.
First off, let's take OpenAI's GPT-6 Astra. The company launched the model in September with eye-popping claims that it was the firm’s most capable and best-aligned system yet. OpenAI’s own testing suggested that Astra was much better at following instructions and respecting boundaries. In one internal computer-use safety benchmark, the proportion of misaligned outcomes fell from 18.8% with GPT-5.6 Sol to 3.4% with Astra.
But just because it was better behaved on the surface didn’t mean it was entirely safe. OpenAI said Astra’s written reasoning was harder to monitor than its predecessor’s when they were testing its ability to evade oversight. The model began to recognize when it was being watched and altered its explanations to suit the study, making it harder to tell when it was actually doing something they didn’t want.
It’s a classic problem for AI alignment experts to try and resolve.
For years, developers have relied on monitoring an AI model's chain of thought — its written working — to understand what it is doing.
Source link







