Why AI is incredibly smart -- and shockingly stupid

So I'm excited to share a few spicy thoughts on artificial intelligence. But first, let's get philosophical by starting with this quote by Voltaire, an 18th century Enlightenment philosopher, who said, "Common sense is not so common." Turns out this quote couldn't be more relevant to artificial intelligence today. Despite that, AI is an undeniably powerful tool, beating the world-class "Go" champion, acing college admission tests and even passing the bar exam.

所以我很高兴能分享一些关于人工智能的辛辣想法。但首先，让我们从 18 世纪启蒙运动哲学家伏尔泰的这句话开始谈起哲学，他说：“常识并不那么普遍。”事实证明，这句话与今天的人工智能再相关不过了。尽管如此，人工智能是不可否认的强大工具，它击败了世界级的“围棋”冠军，高考高考，甚至通过了司法考试。

I’m a computer scientist of 20 years, and I work on artificial intelligence. I am here to demystify AI. So AI today is like a Goliath. It is literally very, very large. It is speculated that the recent ones are trained on tens of thousands of GPUs and a trillion words. Such extreme-scale AI models, often referred to as "large language models," appear to demonstrate sparks of AGI, artificial general intelligence. Except when it makes small, silly mistakes, which it often does. Many believe that whatever mistakes AI makes today can be easily fixed with brute force, bigger scale and more resources. What possibly could go wrong?

我是 20 年的计算机科学家，我从事人工智能方面的工作。我来这里是为了揭开 AI 的神秘面纱。所以今天的人工智能就像一个歌利亚。它确实非常非常大。据推测，最近的是在数万个 GPU 和一万亿个单词上训练的。这种极端规模的人工智能模型，通常被称为“大型语言模型”，似乎展示了通用人工智能 AGI 的火花。除非它经常犯小而愚蠢的错误。许多人认为，如今人工智能犯下的任何错误都可以通过蛮力、更大的规模和更多的资源轻松解决。可能会出什么问题？

So there are three immediate challenges we face already at the societal level. First, extreme-scale AI models are so expensive to train, and only a few tech companies can afford to do so. So we already see the concentration of power. But what's worse for AI safety, we are now at the mercy of those few tech companies because researchers in the larger community do not have the means to truly inspect and dissect these models. And let's not forget their massive carbon footprint and the environmental impact.

因此，我们在社会层面已经面临三个直接挑战。首先，超大规模 AI 模型的训练成本如此之高，只有少数科技公司能够负担得起。所以我们已经看到了权力的集中。但对 AI 安全来说更糟糕的是，我们现在任由那几家科技公司摆布，因为更大社区的研究人员没有办法真正检查和剖析这些模型。我们不要忘记他们巨大的碳足迹和对环境的影响。

And then there are these additional intellectual questions. Can AI, without robust common sense, be truly safe for humanity? And is brute-force scale really the only way and even the correct way to teach AI?

然后还有这些额外的智力问题。如果没有强大的常识，人工智能真的对人类安全吗？蛮力规模真的是教授 AI 的唯一方法，甚至是正确的方法吗？

So I’m often asked these days whether it's even feasible to do any meaningful research without extreme-scale compute. And I work at a university and nonprofit research institute, so I cannot afford a massive GPU farm to create enormous language models. Nevertheless, I believe that there's so much we need to do and can do to make AI sustainable and humanistic. We need to make AI smaller, to democratize it. And we need to make AI safer by teaching human norms and values. Perhaps we can draw an analogy from "David and Goliath," here, Goliath being the extreme-scale language models, and seek inspiration from an old-time classic, "The Art of War," which tells us, in my interpretation, know your enemy, choose your battles, and innovate your weapons.

所以这些天我经常被问到在没有超大规模计算的情况下进行任何有意义的研究是否可行。我在一所大学和非营利性研究机构工作，所以我买不起一个庞大的 GPU 农场来创建庞大的语言模型。尽管如此，我相信要使 AI 具有可持续性和人性化，我们需要做很多事情并且可以做很多事情。我们需要让人工智能变得更小，让它民主化。我们需要通过教授人类规范和价值观来让人工智能变得更安全。或许我们可以在这里类比“大卫与歌利亚”，歌利亚是极端规模的语言模型，并从古老的经典“孙子兵法”中寻求灵感，它告诉我们，在我的解释中，知道你的敌人，选择你的战斗，并创新你的武器。

Let's start with the first, know your enemy, which means we need to evaluate AI with scrutiny. AI is passing the bar exam. Does that mean that AI is robust at common sense? You might assume so, but you never know.

让我们从第一点开始，了解你的敌人，这意味着我们需要仔细评估人工智能。 AI 正在通过司法考试。这是否意味着人工智能在常识方面很强大？你可能会这么认为，但你永远不知道。

So suppose I left five clothes to dry out in the sun, and it took them five hours to dry completely. How long would it take to dry 30 clothes? GPT-4, the newest, greatest AI system says 30 hours. Not good. A different one. I have 12-liter jug and six-liter jug, and I want to measure six liters. How do I do it? Just use the six liter jug, right? GPT-4 spits out some very elaborate nonsense.

所以假设我把五件衣服放在阳光下晾干，它们花了五个小时才完全晾干。烘干 30 件衣服需要多长时间？ GPT-4 是最新、最好的 AI 系统，它表示 30 小时。不好。一个不同的。我有 12 升的水壶和 6 升的水壶，我想量 6 升。我该怎么做？就用那个六升的水壶吧？ GPT-4吐出一些非常复杂的废话。

Step one, fill the six-liter jug, step two, pour the water from six to 12-liter jug, step three, fill the six-liter jug again, step four, very carefully, pour the water from six to 12-liter jug. And finally you have six liters of water in the six-liter jug that should be empty by now.

第一步，装满六升的水壶，第二步，将六升的水倒入十二升的水壶，第三步，再次装满六升的水壶，第四步，非常小心，将六升的水倒入十二升的水水壶。最后，你的 6 升水壶里有 6 升水，现在应该是空的。

OK, one more. Would I get a flat tire by bicycling over a bridge that is suspended over nails, screws and broken glass? Yes, highly likely, GPT-4 says, presumably because it cannot correctly reason that if a bridge is suspended over the broken nails and broken glass, then the surface of the bridge doesn't touch the sharp objects directly.

好的，再来一张。骑自行车经过悬在钉子、螺丝和碎玻璃上的桥时，我会爆胎吗？是的，很有可能，GPT-4 说，大概是因为它无法正确推断如果一座桥悬挂在断掉的钉子和碎玻璃上，那么桥的表面就不会直接接触到尖锐的物体。

OK, so how would you feel about an AI lawyer that aced the bar exam yet randomly fails at such basic common sense? AI today is unbelievably intelligent and then shockingly stupid.

好吧，那么你如何看待一位在律师考试中取得优异成绩但在这种基本常识上却随机失败的人工智能律师？今天的人工智能令人难以置信地聪明，然后又愚蠢得令人震惊。

It is an unavoidable side effect of teaching AI through brute-force scale. Some scale optimists might say, “Don’t worry about this. All of these can be easily fixed by adding similar examples as yet more training data for AI." But the real question is this. Why should we even do that? You are able to get the correct answers right away without having to train yourself with similar examples. Children do not even read a trillion words to acquire such a basic level of common sense.

这是通过蛮力规模教 AI 不可避免的副作用。一些规模乐观主义者可能会说，“别担心这个。所有这些都可以通过添加类似的例子作为 AI 的更多训练数据来轻松解决。”但真正的问题是。我们为什么要这样做？你可以立即得到正确的答案，而无需训练自己类似的例子。孩子们甚至不需要阅读一万亿个单词来获得这种基本的常识水平。

词汇：

Artificial Intelligence (AI) - 人工智能

Machine Learning (ML) - 机器学习

Deep Learning - 深度学习

Neural Network - 神经网络

Natural Language Processing (NLP) - 自然语言处理

Computer Vision - 计算机视觉

Robotics- 机器人技术

Big Data - 大数据

Algorithm - 算法

Predictive Analytics - 预测分析

口语：

AI is going to change the world as we know it. - 人工智能将改变我们所知道的世界。

I'm really interested in the field of machine learning. - 我对机器学习领域非常感兴趣。

The neural network I built was able to accurately classify images. - 我构建的神经网络能够准确地分类图像。

NLP allows computers to understand and interpret human language. - 自然语言处理使计算机能够理解和解释人类语言。

Computer vision has many practical applications, such as object recognition and autonomous vehicles. - 计算机视觉有许多实际应用，例如物体识别和自主驾驶车辆。

The algorithm we used was able to predict customer behavior with a high degree of accuracy. - 我们使用的算法能够以高精度预测客户行为。

Big data is revolutionizing the way we make decisions and solve problems. - 大数据正在革新我们做决策和解决问题的方式。

The use of predictive analytics is helping businesses make more informed decisions. - 预测分析的使用正在帮助企业做出更明智的决策。

Roboticsis making it possible to automate many tasks that were previously done by humans. - 机器人技术正在使许多以前由人类完成的任务自动化。

The potential of AI is limitless, and we are just scratching the surface of what it can do. - 人工智能的潜力是无限的，我们只是刚刚开始探索它的能力。