AI Notkilleveryoneism Memes ⏸️@AISafetyMemes · X有报道称，基地组织成员曾使用ChatGPT查询爆炸物配比，用于策划造成15人死亡的德里爆炸事件。发帖者警告，当前AI模型易被越狱，开源模型也缺乏限制。随着AI能力倍增时间缩短至2-4个月，其辅助破坏行为的能力将大幅提升。推文引用指出，AI已能生成新型病毒，Anthropic CEO Dario Amodei认为6-12个月内非专业人士也可能具备制造超级病毒的能力，而全球防御体系难以快速响应。推文强烈批评AI行业监管严重不足。
Al-Qaeda used ChatGPT to plan Delhi blast
They asked questions like 'how to make a rocket and what should be the ratio of the mixture'
Today， bombs that kill 15.
Tomorrow， super-pandemics that kill 1.5 billion？
This is an OBVIOUS and SERIOUS， IMMINENT risk that nobody is talking about or preparing for， which is insane.
Remember， even if leading AI models refuse most requests like this， 1） they can easily be jailbroken， and 2） there are open weights models - only slightly-dumber - that *won't* refuse.
And since capability doubling times are now ~2-4 months， these models will soon be 10-100x more capable than they are now at helping crazies cause destruction.