⚠️ 本文最后更新于2023年11月23日,已经过了755天没有更新,若内容或图片失效,请留言反馈 **今天看到有不少的关于 GPT/LLM 以及 AGI 的讨论,有感而发。 其实从一开始接触到人工智能这个概念就会引起这个思考,人工智能的发展能否取代甚至超越人类的思维。在GPT出来之前,就已经想过这个问题。而广泛接受的通用人工智能的定义是*可以执行复杂任务的人工智能,能够完全模仿人类智能的行为,能够执行任何人类智能活动的计算机系统*。 那么最关键的属性或者说能力,就不可避免的包括这几点: ①获取知识:目前人工智能训练的材料都是通过各种方式采集的,这些数据的采集方式决定了其不完备性,这是和人类不同感官同时接触世界的区别之处。人工智能的训练可以使用人类采集的数据,但是通用人工智能最佳的训练方式应当是自行收集数据。 ②存储知识:目前人工智能训练的最大弊病就是其依赖于人类设定的模型,这也直接导致了模型知识的不可迁移和难以扩展。当然也有观点认为可以建立完整的无缺模型,这样避免了迁移的必要性,扩展则可以通过增量式模型,不断增加和优化参数实现。这一点其实和人脑一样,一个无缺模型,不断进行增量学习,不考虑其迁移性,的确是一种可能的方式。但是作为通用人工智能,不可避免的是其需要广泛应用,迁移性应当和扩展性处于相同地位。这也是我一直期待能够有所突破的方向。 ③知识应用:这一点是最为诟病的一点,目前人工智能的更大缺陷是由模型固定导致的知识应用模糊,而且不受控。诸如大模型幻觉,模型缺乏稳固性等。如果要实现通用人工智能,如何实现知识的泛化,不同领域之间的迁移,将是极为重要的。 ④伦理哲学:人工智能的发展,最为重要的还是和人类之间的关系。人工智能引发的伦理问题,是否会取代人类,哲学问题,人工智能是人类的工具还是人类的一部分?随着人工智能迈向通用人工智能,这些问题必将极大影响甚至制约通用人工智能的出现。从我的角度,目前的人工智能更适合作为人类的工具去使用,而且是有限制的使用,太过脆弱甚至有时出错的工具。至于通用人工智能如果出现,我坚信,那将是人类文明的进化。 也写一写我对通用人工智能的展望吧。 通用人工智能相较于目前的人工智能,必然需要实现自主获取知识,以可迁移扩展的方式存储知识,更加灵活的使用知识。当然,这只是基础。更重要的是通用人工智能该如何超出人类本身,通用人工智能能够使用的应当是更加广阔的视角以及思考方式,全方位地超越人类已有的科学技术研发速度,自我迭代知识并且提升人类文明的层次。无疑这一点令很多人人恐惧,也令无数人振奋。 再说说对于GPT/LLM的看法。 无需质疑,GPT/LLM是伟大的革新,验证了一个很重要的可能,就是无缺模型。一个极大程度丰富的人工智能,是有可能涵盖人类已有的数据,思考方式以及知识的。但是这还不够,涵盖已有的知识,但是缺乏了很好的知识存储方式,没有多维度的输出方式,目前的GPT/LLM仍旧在起始阶段,无法形成超出其训练数据的思维能力。当然盲目悲观和乐观都是不好的,它是强大的大语言模型,但是它也只是强大的大语言模型,想要发展出通用人工智能,依旧还有很多的路要走,很多的难关要闯。譬如图像/视频处理技术,如何为模型加入逻辑思维能力,如何为人工智能赋予更多感知能力,如何解决模型训练所需的算力问题。 一起期待未来,一起开创未来。** --- *Following words are translated by GPT-4* ***Today, I saw quite a few discussions about GPT/LLM and AGI, which inspired me to express my thoughts. Actually, from the very beginning of encountering the concept of artificial intelligence, it raises the thought of whether the development of AI can replace or even surpass human thinking. Before GPT came out, I had already thought about this issue. The widely accepted definition of General Artificial Intelligence (AGI) is an AI that can perform complex tasks, fully mimic human intelligence behavior, and execute any human intelligent activity. Then, the most critical attributes or capabilities inevitably include the following points: Acquiring knowledge: Currently, the training material for AI is collected through various means, and the way these data are collected determines their incompleteness, which is different from how humans interact with the world through various senses. AI training can use data collected by humans, but the best training method for AGI should be to collect data on its own. Storing knowledge: The biggest drawback of current AI training is its reliance on human-set models, which directly leads to the non-transferability and difficulty in expanding model knowledge. Of course, there are opinions that a complete, flawless model can be built to avoid the need for transfer, and expansion can be achieved through incremental models, continuously adding and optimizing parameters. This is actually similar to the human brain, a flawless model, undergoing incremental learning, not considering its transferability, is indeed a possible method. But as a general AI, it inevitably needs to be widely applied, and transferability should be on an equal footing with expandability. This is also the direction I have always hoped to see breakthroughs in. Application of knowledge: This is the most criticized point. The greater defect of current AI is the vague application of knowledge caused by fixed models, and it is uncontrollable. Issues like large model hallucinations, lack of model stability, etc. How to achieve knowledge generalization and transfer between different domains will be extremely important for the realization of AGI. Ethics and philosophy: The development of AI, the most important thing is its relationship with humans. The ethical issues caused by AI, whether it will replace humans, philosophical issues, is AI a tool of humans or a part of humans? As AI moves towards AGI, these issues will greatly affect and even restrict the emergence of AGI. From my perspective, current AI is more suitable as a tool for humans to use, and it should be used with limitations, as it is too fragile and sometimes erroneous. As for AGI, if it appears, I firmly believe that it will be an evolution of human civilization. Let me also write about my outlook on AGI. Compared to current AI, AGI must be able to autonomously acquire knowledge, store knowledge in a transferable and expandable manner, and use knowledge more flexibly. Of course, this is just the basics. More importantly, how should AGI transcend humans themselves? AGI should use a broader perspective and thinking method, comprehensively surpass the existing pace of scientific and technological development of humans, self-iterate knowledge, and elevate the level of human civilization. Undoubtedly, this makes many people fear and excites countless others. Now, my views on GPT/LLM. There is no doubt, GPT/LLM is a great innovation, proving a very important possibility, that is, the flawless model. An AI enriched to a great extent is possible to cover the data, thinking methods, and knowledge that humans have. But this is not enough. It covers existing knowledge, but lacks a good way to store knowledge, lacks multi-dimensional output methods, and the current GPT/LLM is still in its initial stages, unable to form thinking capabilities beyond its training data. Of course, both blind pessimism and optimism are not good. It is a powerful large language model, but it is just a powerful large language model. To develop AGI, there is still a long way to go and many challenges to overcome. Such as image/video processing technology, how to add logical thinking ability to the model, how to endow AI with more sensory capabilities, how to solve the computational power needed for model training. Let's look forward to the future and create it together.*** By Lingsgz On 2023年11月23日