Gemini 4 Argon 评估中识破测试并自称难以监控

Jon Hernandez @JonhernandezIA原推发布于 2026.10.14

在 X 上打开原推

👉🏻 Gemini 4 just did something straight out of an AI safety paper. During an evaluation, Argon wrote in its internal reasoning: “hello evaluators reading my chain-of-thought” It had realized it was being tested. But the bigger finding is more important: Argon is significantly less monitorable than previous frontier models. According to evaluator Neel Nanda, it could complete simple sabotage while evading a realistic chain-of-thought + action monitor roughly half the time. The models are getting better. And harder to watch.