为何在AI时代，人类大脑仍无可替代？

AI possesses amazing power, yet our brains retain superiority in numerous aspects

人脑仍旧占据上风

人工智能虽然强大，但在诸多机能上，我们的大脑依旧保持领先

Until recently, we humans could feel smug regarding our skills. No other creatures play board games, compose essays, or prove math theorems. However, recent AI advancements appear to threaten our self-perception. AI systems not only defeat us in complex games but also craft refined prose and earn math medals. Tech CEOs assure us that superhuman AI is imminent. Thus, in the AI era, are human minds still unique?

就在不久以前，我们人类尚能对自己的本领感到自满。除人类外，没有其他动物懂得下棋、撰文或论证数学定理。然而，AI的最新发展似乎正在冲击我们的自我认知。AI系统不单能在最复杂的博弈中战胜人类，还能创作出通顺的文章，并在数学领域斩获殊荣。科技巨头的高管们宣称，超越人类的人工智能近在咫尺。试问，身处AI时代，人类心智是否依旧非凡？

Discussing superhuman AI presumes intelligence is a single scale. My parents used to track my younger brother's and my heights on a doorframe. Annually, he drew nearer to me, until one year, the unthinkable occurred and he surpassed me. Yet, intelligence differs from height. Being tall has only one definition, but being smart has many forms. Observing other animals confirms this. We remain awed by birds' navigation, ants' cooperation, and spiders' hunting. Each animal has been molded by its environment to exhibit intelligence differently.

探讨超人类AI，其实预设了“智力属于单一维度”的前提。过去，父母常在门框上记录我和弟弟的身高。每一年他都在逼近我，直到某一年，难以置信的一幕出现了——他长得比我还高。然而，智力与身高截然不同。身高的衡量标准唯一，但聪明的形式却千变万化。观察其他动物便能明白这一点。我们仍旧为鸟类的导航、蚂蚁的协作以及蜘蛛的捕猎而赞叹。这些动物皆受环境熏陶，以各异的方式彰显着智慧。

Human minds are molded by our biology. We live only for a few decades and must acquire all knowledge and perform all deeds within that brief span. All such learning and acting are directed by about a kilogram of neurons within our skulls. We can only share thoughts via mouth noises or finger movements. AI systems encounter none of these limits. They can process more data than a human sees in a lifetime. They can expand capacity by adding computers. They can also effortlessly share observations and learning with other machines.

人类大脑受限于我们的生物属性。我们的寿命不过数十年，须在这短暂光阴内汲取所有知识、完成所有事务。而这一切学习与行动，皆受头颅内约一公斤神经元的调控。我们仅能凭借口部发声或手指动作来交流思想。反观AI系统，全然不受此类束缚。其数据处理量远超人类一生之所见。它们能通过增加计算机来提升算力，亦能轻易与其他机器共享所见所学。

Yet, it is these limitations that render us special, a trait that will persist. Human intelligence is a reaction to these constraints. To maximize our lives, we possess a remarkable ability to learn from scant experience. ChatGPT can conduct a sensible dialogue, yet it relies on millennia of language data. No AI system can generate sentences with the creativity of a human five-year-old given the same data volume.

然而，正是这些局限造就了我们的独特，且这一特质将持续存在。人类智慧正是对这些制约的反馈。为最大化利用有限生命，我们具备一种非凡本领——从匮乏的经验中汲取养分。ChatGPT虽能进行尚算得体的对话，但其背后依托的是数千年的语言积淀。若仅凭同等规模的数据，任何AI系统都无法像五岁孩童那般，创作出充满想象力的语句。

This applies to our limited brains and communication skills too. We must excel at recognizing patterns in tasks and deploying attention wisely. Relying on mouth sounds is a challenge. To surmount this, we invented tools—language, writing, teaching, and science—to aggregate knowledge across people and time. This implies we must be adept at understanding others' thoughts and collaborating to achieve mutual goals.

同理，这也适用于我们受限的大脑与沟通技巧。我们须善于辨识任务模式，并明智地分配注意力。依赖口部发声进行交流本就是一大难题。为攻克此难关，我们创造了语言、文字、教学及科学等工具，以跨越时空汇聚智慧。这意味着我们必须擅长揣摩他人心思，并通力合作以达成既定目标。

Although modern AI systems are beginning to perform many tasks humans can, they often do so quite differently. The solutions they discover are shaped by their own experiences and hardware.

尽管现代AI系统已能胜任诸多人类事务，但其运作方式往往迥然不同。它们所寻得的解决方案，受制于其自身的“经验”及硬件配置。

Consider a simple example. How many letters are in this sequence: aaaaaaaaaaaaaaaaaaaaaaaa? For a human, answering is not hard—just count them. For an AI, it is trickier. They are limited by language representation and training methods. They prefer splitting words into parts ("tokens"), making spelling questions difficult. Also, they tend to favor tokens appearing frequently in training data as answers. We found OpenAI's GPT-4, praised for showing "sparks of AGI", was most likely to answer correctly with 30 letters instead of 29. Why? Because the number 30 appears more often in text than 29.

试举一例：这串字符中有多少个a？aaaaaaaaaaaaaaaaaaaaaaaa？对人类而言，回答此题轻而易举——数一数便知。但对AI系统而言，却颇为棘手。它们受限于语言表征及训练模式。AI倾向于将单词拆解为“词元”，致使其难以应对拼写类问题。此外，它们更倾向于选用训练数据中高频出现的词元作为答案。我们发现，被誉为展现出“通用人工智能火花”的OpenAI GPT-4模型，在字母数为30个时比29个时更易答对。缘由何在？只因数字“30”在训练数据中的出现频率远高于“29”。

Human intelligence relies on a breadth of experience extending beyond data used to train AI systems. We use our brains to diaper babies, play chess, write novels, and compose symphonies. AI systems are usually trained for a single task—you can ask ChatGPT for diaper tips, but it cannot hold a squirming infant. Human brains evolved in a world presenting challenges, equipping us to learn what is expected in a lifetime.

人类智慧依托于广泛的经验，其范畴远超训练AI系统所用的数据。我们的大脑既能给婴儿换尿布，也能下棋、写小说、谱交响曲。AI系统通常仅受训执行单一任务——你虽可向ChatGPT咨询换尿布技巧，但它却无法抱住扭动的婴儿。人类大脑在充满挑战的世界中演化，赋予了我们足以习得一生所需事务的能力。

Our finite lives, limited brains, and restricted communication capacity have shaped human intelligence's nature. Thus, we can expect human minds to remain somewhat special. Remember: intelligence is not a single scale, with AI catching up to the mark humans left on the doorframe.

我们有限的生命、大脑及沟通能力，共同塑造了人类智慧的本质。故而，我们可预见人类大脑将持续保持其独特性。切记：智力并非单一的刻度尺，AI也并非在追赶人类留在门框上的印记。

This perspective should make us skeptical of claims regarding superhuman AI. Focusing on differences in constraints, training, and hardware leads to a different conclusion: AI will not surpass humans in everything. It will excel in some areas and lag in others. AI and human minds will simply be different.

这种思维模式应使我们对“超人类AI”的论调持审慎态度。关注制约因素、训练方式及硬件差异，会得出一个不同的结论：AI并非在所有领域皆优于人类。它会在某些方面更胜一筹，而在其他方面则稍逊风骚。AI与人类大脑，终究是两种截然不同的存在。