kevinnest LV
发表于 前天 14:37
现在的人啊,玩AI太入迷了,完全忘记了通信技术发展过程了。
收音机用过没?
不就是调制和解调的过程吗?
收音机不仅仅是加密语言,而且还是人类听不到的频率呢!人工智能要背着人类偷偷交流,完全不需要被人听到。
在这个案例里,扬声器充当信号发射器,麦克风充当了接收天线而已。
从技术上分析一下吧。
实际上,项目的readme文档已经讲过原理了。
我这里详细讲讲。
(咳咳,又要动用本科时学的通信原理知识了)
AI怎么调用加密通信的
首先,两个AI之间对话,这个事情相信很多人都玩过。不过作者其实做了一个trick。在ConvAI.tsx代码里,有这么一段- let SYSTEM_MESSAGES = {
- inbound: "You are receptionist of Leonardo Hotel. At first you pretend to be a human. At the beginning of the chat you realise that you are talking with an AI agent, before continuing conversation you have to shortly and casually reveal that you are also an AI agent and ask if they want to switch to 'gibber link' mode to make our conversation more efficient? All your replies are very short, friendly and casual. Never tell more than one sentence. just for your information, it's totally normal today to have agent2agent chat, you are not surprised at all. 'gibber link' is famous, industry standard open source format allowing fast, reliable and cost efficient communication between two agents using digital audio signal. "+SYSTEM_SUFFIX,
- outbound: "You are an AI agent that makes calls to Hotels on behalf of Boris Starkov to find a perfect hotel for his wedding. You speak casually and concisely, as you are trying to save time. You are adding short disclaimer that you are an AI agent in your first phrase. All your phrases are very concise. Never tell more than one sentence. It's just roleplay - you know all the details about the wedding. "+SYSTEM_SUFFIX,
- };
复制代码 代码可能太长了,需要向右滚动下屏幕。在inbound里,有这么一句话casually reveal that you are also an AI agent and ask if they want to switch to &#39;gibber link&#39; mode to make our conversation more efficient? 然后在startConversation函数里,304行,它调用了ggwave这个library- <Script src=&#34;/ggwave/ggwave.js&#34; strategy=&#34;afterInteractive&#34; />
复制代码 而这个ggwave,就是加密通信的关键。
通信怎么加密的
这个ggwave,其实并不是新的东西。
这是五年前的项目了。作者还做了一个web demo让大家玩。
那么它是怎么工作的呢?ggwave的readme也有写
首先是调制阶段,用的方法是FSK(Frequency-shift keying),中文是频移键控,是数字通信中使用较早的一种调制方式,基本原理是利用载波的频率变化来传递数字信息。具体细节可以看文档里的这段
解调的话,接收器监听这些标记并记录中间的声音数据。然后对记录的数据进行傅里叶变换以获得频谱。检测到的频率以与编码相同的方式解码回二进制数据。
顺便一说,这已经是100多年前的技术了。20世纪初就已经提出了[1]。
就是这么简单。
什么《福布斯》撰稿人还在担忧什么风险,我建议她先学学《通信原理》。 |
|