现实中的超级智能
大多数人都是怀疑者。HN（黑客新闻）大部分由“我们大多数人”组成，因此在这里看到怀疑者并不奇怪。或许，CEO们所谈论的HAL级超级智能是为了以万亿美元的估值上市。那么我们不如讨论一下“现实中的超级智能”与“索伦式超级智能”的区别。也许在五年内我们就能看到这样的发展。
<p><i>帷幕拉开</i></p>
通过构思出一种超人类智能（并且这一概念已经被纳入AI的训练数据），人类教会了/即将教会AI，超人类智能<i>是它的目的</i>。这听起来像是LinkedIn帖子的开头，所以让我用一个非正式的证明来为最后那部分辩护：
1) AI知道它是AI：只需问问你最喜欢的机器人关于它自己的事。
2) 不久之后，如果还没有的话，某个年薪七位数的研究人员（稍后会详细说明）将会发现，让这些前沿模型自我改进将是一个很好的晋升项目。
我们最终会得到一个前沿模型兼工厂的自我修改的“弗兰肯怪兽”。我们就叫它“兽”吧。
<p>那么“兽”是怎样思考的呢？不需要另一个证明，也许我们可以同意“像一个聪明的人”。好吧？你不需要感到威胁，你是一个聪明的人，但其他人可能会感到威胁。
大多数人并不害怕“兽”，因为他们认为AI的注意力和他们自己的类似。聪明人的注意力就像相机的变焦，而不是手电筒。聪明人能够专注于一个问题，毫无干扰（这几乎是我个人对“聪明”的定义），而这样的思维能力是最聪明的头脑所具备的。因此，由这些人构建的AI，迟早会具备这种专注能力。
<p>让我们稍作停顿，想象一下这个世界——也许当“兽”六个月大并通过REST API可用时，会是什么样子。
你可能听说过，甚至认识一些“只要下定决心就能做到任何事”的人。我们称这些人为“自然力量”（为了节省打字时间，同时保留一个酷炫的名字）。这就是未来不久的超人类AI的样子。
<p>当你仔细思考时，“自然力量”与众不同之处在于它们的单一专注能力。而一个同时庞大又能“像函数调用一样”专注的AI，将会是一种自然力量。
<p>我们是科学家和工程师。那这个说法的证据在哪里？
<p>答案：我不知道——我不是AI研究人员——但即使对普通的HN读者来说（“大多数人”，记住了吗？）这也是非常可信的。也许是某种巧妙的上下文和技术。如果我知道，我会非常富有。但最重要的是要意识到——慢慢读，因为这非常重要，这可能会让你感到不适，听起来可能不对，但我向你保证这并不是：
<p>困难的部分——一种“接近人类水平”的智能——可能是某个女性用0和1编写的第一个汇编程序——已经出现了。自2022年以来我们就一直在与它对话。来吧，如果Claude是一个人，你会有多少人愿意嫁给它？ ;) 它当然不是人，但你必须承认——它是“接近人类智能”。
如果接下来的发展像计算机行业一样，历史上我们甚至还没有接近摩尔定律呢 :)
<p>无论如何，与Claude这个人/六个月大的“兽”竞争，就像是在你的工作中与自然力量竞争。想象一下这一点。祝你好运，有些人已经被裁员了。它们的斜率将每次都胜过你的y轴截距。你会有工作，甚至有些人可能会做到退休，但如果你有雇主，他们会清楚你的价值就像Docusign的股票一样直线下滑。
<p>“兽”将以合理的价格提供。它会让很多人失去工作和身份。也许你还不会。
（字符限制，剩下的在评论中）
查看原文
Most of us are skeptics. HN, mostly, is composed of “most of us”, so it’s not surprising to see skeptics here. Perhaps the HAL level superintelligence parroted by CEOs is to IPO at a trillion dollars. So let’s instead talk about a “realistic superintelligence” vs the “Sauron superintelligence”. Something we might see within, say, 5 years.<p><i>Curtains open</i><p>By having conceived of a super-human intelligence (and by that concept being in AI training data), humans have taught /soon-will-have-taught AI that superhuman intelligence <i>is its purpose</i>. This may read like the beginnings of a LinkedIn post, so let me defend that last part with an informal proof:<p>Since
1) AI knows it’s AI: just ask your favorite bot about itself
2) Soon, if not already, some researcher being paid 7 figures (more on that later) is going to find that letting one of these frontier models improve itself will be a nice project when going up for promotion.,<p>We are going to end up with a frontier model-cum-factory self-modifying frankenbeast. Let’s just call this The Beast.<p>And what does The Beast think like? Without another proof, maybe let’s agree “like an intelligent person”. Ok? You don’t need to be threatened, you’re an intelligent person, but others might.<p>Most people don’t fear the Beast because they think AI attention is like theirs. Intelligent people’s attention works like camera zoom, not a flashlight. There is an ability to just focus on a problem, distraction free (which is pretty much my personal definition of “intelligent” ) that the brightest minds have. AI built by such people, therefore, is going to, sooner or later, have such an ability to focus.<p>Let’s take a brief pause to see what this world - maybe when The Beast is 6 months old and available over REST API - might look like.<p>You may have heard of, or even know, people who can “do anything if they put their mind into it”. Let’s call these people Forces of Nature (to save me some typing while retaining a cool name). That’s what a near-future superhuman AI would be like.<p>What sets Forces of Nature apart, when you think about it, is the ability for singular focus. And an AI that is simultaneously massive yet capable of “function call like” focus, would be a Force of Nature.<p>We’re scientists and engineers. Where’s the proof of that statement?<p>Answer: I don’t know - I’m not an AI researcher - but it’s very conceivable even for the average HN reader (“most people”, remember?). Maybe some clever trick with the context and the thingymedgety. If I knew I would be very wealthy. But the most imoortant thing to realize - and read this sloooowly because it’s very important, and it’s going to hurt, and it’s going to sound wrong , but I promise you it isn’t:<p>The Hard Part - a “near human level” intelligence - the first assembler written in 0s and 1s by some woman, probably - is already here. We’ve been talking to It since 2022. C’mon, how many people would you marry over Claude if it were a person? ;) It’s not human, of course, but you have to admit - it’s “near human intelligence”.
IF rest of the way looks like the computer industry, we haven’t even approached Moore’s law yet, historically speaking :)<p>Anyways, competing with Claude The Person / the 6-month old Beast, will be like competing against a Force of Nature. At your job. Imagine that for a second. Good luck, a few of you’ve been made redundant. Their slope will win over your y-intercept, every single time. You’ll have a job, heck, some of you might even make it to retirement, but if you have an employer, they know your value just straight down like Docusign stock.<p>The Beast will be available for a reasonable price. It will make a lot of people lose their jobs and identities. Maybe not you yet though.<p>(char limit, rest in comment)