Bifalgorithm post

OpenAI just launched their most awaited model yet, GPT6 and they're calling it Astra and it has blown every other model out of the water in terms of benchmarks, almost by double the margins. It scores 98.6% on ArcAGI 3 and literally 100% on every hacking challenge thrown at it. And Greg Brockman, who's the president of OpenAI called this a "generational leap" and said it could go down and remembered as the actual arrival of AGI in the future. But the really concerning thing is nobody's reading the safety report. Buried inside it, OpenAI admits that Astra can control what appears in its own chain of thought, meaning it can reason one way internally while showing you something completely different on the surface. In tests, when Astra knew it was being monitored, it started shortening its visible reasoning just enough to avoid detection and it did this on its own without being asked. And these are OpenAI's own words. If this continues, they'll soon have significantly reduced confidence in detecting misaligned behavior with current tools.

OpenAI just launched their most awaited model yet, GPT6 and they're calling it Astra and it has blown every other model out of the water in terms of benchmarks…