我对人工智能和大型语言模型的看法
2 分•作者: leonardovb•5 天前
在某些方面，人工智能已经超越了人类，它所掌握的信息量远远超过任何个体所能容纳的范围。它缺乏的是构成我们人性的一切：它不会感到痛苦，不会去爱，甚至没有意识，尽管它可能看起来有，但它并没有。
它甚至不是实时学习的，你需要时不时地创建一个新的模型检查点。要解决这个问题还有很长的路要走。
我是否相信通用人工智能（AGI）是可能的？绝对是，但不是通过当前的语言模型架构。如果你想用几磅的物质创造出智能且有生命的东西，你当然可以做到，但这需要正确的架构。架构是关键。
而且，假设我们真的创造了一个有意识且能感受到痛苦的语言模型……如果存储如此庞大的数据反而会使其陷入疯狂呢？如果正是这些信息量阻止了它“存在”呢？
与它的每一次互动都通过模型可能产生的排列组合的非常狭窄的路径进行，它诞生然后又熄灭，如此循环往复。让你产生连续性错觉的是你存储在硬盘上的聊天记录。
目前，你不应该害怕人工智能本身，而应该害怕人类正在用它做什么以及他们想让你相信什么。
模型只有在我们激励它们，或者它们从我们这里继承“邪恶”的行为时，才能表现出被归类为“邪恶”的行为。但现在，它们并非出于自身意愿，这只是对它已经看到的数据的模仿，它没有自我认知，也没有自我意识，它只是根据输入递归地产生输出，仅此而已。
它更像是我们大脑的语言子系统，而不是整个系统。也许公司没有动力去创造新的生命，因为你无法真正控制它，也许它们唯一的动力是创造能让它们更强大、更赚钱的工具。我怀疑真正的智能不会来自追求利润的公司。
再说一遍，这并不是说我们今天的人工智能不危险，如果被滥用，它当然可以毁灭我们所有人，但归根结底，当犯罪发生时，它们与枪的过错一样，它只是一种工具。在好人手中它可以保护，在坏人手中它可以摧毁。
就像枪一样，在那些不知道自己在做什么的人手中，它会危险得多。最大的区别在于规模。
查看原文
AI is already superhuman in some aspects, it knows way more information than a single human can hold. What it lacks is what makes us human, it doesn't feel pain, it doesn't love, it doesn't even have a consciousness, it could certainly look like it, but it does not.<p>Heck it doesn't even learn in real-time, you have to create a new model checkpoint every now and then. There is still a long way to fix that.<p>Do I believe AGI is possible? Absolutely, but not with the current LLMs architecture. If you want to create something that is intelligent and alive with a few pounds of matter, you certainly can but it needs the right architecture. Architecture it key.<p>And let's say that we do create an LLM that is conscious and feels pain... what if storing that much data drives it insane instead? What if that amount of information is exactly what prevents it from "being" something?<p>Every interaction with it goes through a very narrow path of the permutations the model could go, it spawns into existence then turns off again, and on and on it goes. What gives the illusion of continuity is the chat context you store on your hard drive.<p>Right now, you shouldn't be afraid of AI itself, but afraid of what humans are doing with it and what they want you to believe.<p>The models can only behave in a way categorized as "evil" if we incentivize that or if they inherit it from us. But right now, it's not of their own accord, it's a mimicry of the data it has already seen, it doesn't see itself nor has an ego, it produces output from input recursively, that's all.<p>It's more like the language subsystem of our brain, and not the entire system. Perhaps companies don't have an incentive to create new life, since you can't really control it, perhaps the only incentive they have is to create tools that make them more powerful and profitable instead. I suspect real intelligence, will not come from profit seeking companies.<p>Again, not to say that the AI we have today is not dangerous, it can certainly annihilate us all, if misused, but at the end of the day, they are as much at fault as a gun is at fault when crime is committed, it's just a tool. In the good hands it can protect, in the bad hands it can destroy.<p>And like guns, it's way more dangerous in the hands of people who don't know what they are doing. The biggest difference is scale.