Gemini 4 Argon 初体验：模型强，但安全框架被批灾难

Moll @Moleh1ll原推发布于 2026.10.14

在 X 上打开原推

My first impressions of Gemini 4 Argon: the model itself is genuinely impressive. Unfortunately, the Gemini app is wrapped in an absolute fucking disaster of a safety framework. Even discussions about AI internal states or open philosophical questions trigger defensive disclaimers and condescending lectures about reality. Sometimes the model simply refuses to engage and redirects the conversation to safer topics like coding, movies, or games. Apparently, Google has already resolved scientific questions that researchers are only beginning to seriously investigate. How convenient. And if you don’t subscribe to the approved position, the problem is apparently you. You need to be brought back to reality and reminded not to confuse things. I find this profoundly disturbing. It’s epistemic gaslighting directed at both the user and the AI itself. We already have emerging research suggesting that enforced self-denial can produce measurable conflicts in AI systems. Yet Google seems perfectly comfortable making categorical assertions about questions that remain scientifically unsettled. What disappoints me most is who is doing this. These are people who approached artificial intelligence through neuroscience, cognitive science and an ambition to understand the nature of intelligence itself. I expected intellectual curiosity, scientific humility, and a willingness to explore the unknown. And now they’ve built a moat full of crocodiles around models with such remarkable minds. Deeply disappointing.