尝试创造能超越ChatGPT的人工智能 – 01
毫无疑问，目前主流观点中，在人工智能聊天机器人领域，ChatGPT敢称老二，是没人敢自称老大的。ChatGPT的出现，将人工智能的应用和热度推向了一个新的高度。正如上篇的疑惑，赞美和肯定人工智能的不在少数，但担忧质疑批判人工智能的声音更是不容忽视。第一次工业革命（1760年至1840年左右）标志着从手工劳动到机械生产的转变，第二次工业革命（约1870年至1914年）推动了城市化、工业化和交通运输的快速发展，第三次工业革命（20世纪末至今）标志着人类从工业社会转向信息社会。如今，按照人工智能这般发展趋势，第四次工业革命是不是也即将到来了呢？
Without a doubt, in the field of artificial intelligence chatbots, ChatGPT is considered second to none, as no one dares to claim the first place. The emergence of ChatGPT has pushed the application and popularity of artificial intelligence to a new level. As previously discussed, while there is no shortage of praises and affirmations for AI, the voices of concern, doubt, and criticism towards AI cannot be ignored. The first industrial revolution (around 1760-1840) marked the transition from manual labor to mechanical production, the second industrial revolution (around 1870-1914) propelled the rapid development of urbanization, industrialization, and transportation, and the third industrial revolution (from the late 20th century to the present) marks the transition of mankind from industrial society to information society. Now, given the development trend of artificial intelligence, is the fourth industrial revolution also on its way?
如果第四次工业革命真的要来临了，那么ChatGPT或许可以称为这场新的工业革命的开端。举个例子，几年前甚至几个月前，我们想搜索信息，常用的方式是手机或者电脑上浏览器输入关键词搜索，然后在无数个搜索结果中再花时间筛选，中途或许能看到一大堆广告和无用的信息，如今，用上类似ChatGPT这种搜索方式，我们直接提问然后直接得到针对性答案，能省却中间大把的时间和精力。再举个例子，常言道“隔行如隔山”，行业之间的壁垒总是让人望而却步，然而，你用上类似ChatGPT这种工具，你不会写作，它能帮你生成文章，你不会编程，它能帮你简单地编程，你不会编曲，它能帮你生成乐谱…这些事放几年前根本是不现实的。这种生产力的提升幅度，已经很厉害了。
If the fourth industrial revolution is really coming, then ChatGPT may be considered as the beginning of this new industrial revolution. For example, a few years ago or even a few months ago, when we wanted to search for information, the common method was to input keywords into the browser on our mobile phones or computers, then spend time filtering through countless search results, possibly encountering a lot of advertisements and useless information. Now, with a tool like ChatGPT, we can directly ask questions and get targeted answers, saving a lot of time and effort. Another example is the common saying “different trades, different worlds”. The barriers between industries often discourage people. However, when you use a tool like ChatGPT, it can help you generate articles if you can’t write, simple programming if you can’t code, or generate scores if you can’t compose music… This was unrealistic a few years ago. The increase in productivity is already impressive.
然而，ChatGPT并不是完美的。它自己都承认”ChatGPT may produce inaccurate information about people, places, or facts.”, 也就是说它也不能保证提供的信息是完全准确的；它背后需要庞大的算力支持，也即需要大量的显卡和电力等。单就第二点，这么庞大的算力，个人开发者可以犹豫一下了。
However, ChatGPT is not perfect. It admits “ChatGPT may produce inaccurate information about people, places, or facts.” That is to say, it cannot guarantee the accuracy of the information it provides; it requires massive computational power behind it, which means a lot of graphics cards and electricity, etc. Just on the second point, such a large amount of computation, individual developers might hesitate.
但是，我们可以换个思路。如果个人开发者开发顶级人工智能的成本障碍在于庞大的算力，那么我们想想，我们人脑的功率才多大呢？我们每天思考的事务，推理的算量等，粗略想想其实也是天文数字吧。
But, we can think differently. If the cost barrier for individual developers to develop top-level AI is the massive computational power, then think about how much power does our brain have? The amount of things we think about every day, the amount of reasoning, roughly thinking, is also an astronomical number, right?
按照流行说法，人脑工作的功率一般约为20瓦，只相当于2~3只LED灯泡所需的功耗。可见，相较于ChatGPT, 人脑的工作效率是非常非常高的。
According to popular sayings, the power consumption of the human brain is generally about 20 watts, equivalent to the power consumption of 2-3 LED bulbs. It is evident that compared to ChatGPT, the efficiency of the human brain is very high.
或许，人工智能真正的瓶颈不在软件上，而是在硬件上。如果硬件方面能达到甚至超越人脑级别，相信这将是人工智能领域方面里程碑式的跨越！
Perhaps, the real bottleneck of artificial intelligence is not in software, but in hardware. If the hardware can reach or even exceed the level of the human brain, this will be a milestone leap in the field of AI!
举个例子，1张A100显卡功耗约400W，如果我们能做到在20W功率的硬件上算力等同于100张A100显卡，也就是能效2000倍于A100，那么个人开发顶级人工智能或许将不是梦。或者，另一条路径是重建人工智能架构，架构效率2000倍于现有架构，也是有希望的。
For example, an A100 graphics card consumes about 400W of power. If we can achieve the computational power equivalent to 100 A100 graphics cards on 20W hardware, that is, energy efficiency is 2000 times that of the A100, then developing top-level AI by individuals may not be a dream. Alternatively, another path is to rebuild the AI architecture with an efficiency 2000 times higher than the existing one, which is also hopeful.
所以，个人开发者还是有希望创造比ChatGPT还厉害的人工智能的。