猜想：GenAI 与人类有何不同？
我的理解
本课从哲学层面列举了 GenAI 与人类的五大核心差异：判断（无法在真实世界中验证）、「尤里卡」时刻（无法从苹果落地跳跃到星体运动）、批判性思维（持续挑战并改进自己答案的能力存疑）、对人性的非文字理解，以及直觉（提前生成远期想法的能力）。最关键的洞察是：GenAI 或许能发现浮力原理，但无法判断这一知识对人类是否重要——识别重要性这一步必须由人类完成。机器人技术目前仍无「GPT 时刻」的迹象，与物理世界的交互是 GenAI 的根本性弱点。这些推测不是定论，而是邀请学员用批判性思维检验自己的世界模型。
相关链接
- Ch05-L04 研究 生成式 AI 建立在共识之上 — 共识机制是「尤里卡」差异的技术根源
- Ch05-L08 要点二 有价值的观点都是反共识的 — 人类独特认知能力正是反共识价值的来源
- Ch07-L05 基于人类反馈的强化学习 RLHF — RLHF 通过人类反馈弥补 GenAI 理解人性的短板
- Ch02-L02 研究 LLM内部机制 记忆知识上下文 — LLM 内部机制对比人类认知的更深层视角
原文
Lesson 41 of 68 猜想：GenAI 与人类有何不同？/ Conjecture: How is GenAI different from humans?
我们对脑科学和神经科学的理解非常有限，因此只能从哲学中寻求答案。顺便一提，除非脑科学出现重大范式转变，否则 Neuralink 这类脑机接口不太可能实现人们想象中的功能。但我们并非哲学专家，以下内容仅供参考。
判断
无论 ChatGPT 多么强大，它都只能吸收数字信号，无法与真实世界互动。它或许能从一万名专家那里学到“做 A 会导致 B”，但如果不在真实世界中进行实验，它无法从根本上验证这一结论的真伪。要真正理解此类问题，必须亲身体验。
“尤里卡”时刻
牛顿因看到苹果落地而发现万有引力，并由此预测了星辰的运动。在哥白尼之前，人人都看到太阳东升西落，认为太阳绕地球转。如果当时存在 ChatGPT，它一定会笃定地认为太阳绕地球转。虽然它或许能根据苹果如何落地推测桃子如何落地，但很难推导出星辰的运行规律。
然而，能发现万有引力这类事物的人毕竟凤毛麟角。更重要的是，要弄清楚这种认知能力究竟是什么，以及它在日常生活中如何体现。阿基米德在洗澡时发现浮力原理，喊出“尤里卡”，描述的正是那种“灵光乍现、顿悟通透”的瞬间。这一瞬间可以具体描述为“将几个相关的点连接起来，从而发现第三个点”。
这其实包含两个步骤：第一是发现新知识，第二是立即理解它的重要性。ChatGPT 也许能发现很多东西，包括浮力原理（我不确定）。但 ChatGPT 无法判断这一知识的重要性，也无法判断它对人类是否有用。这一步必然需要人类的介入。
批判性思维
批判性思维是主动辨别真伪、不断发掘更重要、更切合实际、更具创造性和价值的想法。ChatGPT 也许能给出不错的答案，但它能否持续地挑战并改进这一答案，仍是未知数。这是 OpenAI 值得继续探索和发展的能力。
理解人类
人类的文字知识中固然蕴含着大量对人性的理解，但也存在一些未被文字记录下来的人性或偏好。将这些与对人的真实世界理解结合起来——而非通过问卷调查或网络数据——所带来的增量理解，是人类相对于 ChatGPT 的优势所在。
直觉
回到 ARLLM 的本质。ARLLM 试图生成下一个词，但人类是否也在做同样的事，尚不清楚。如果说是生成下一句或下一段，那么人类的优势或许在于能生成更远以后才会出现的想法。这或许就是所谓直觉的“数字化定义”。模型能否做到这一点尚不确定，但很可能相当困难。
与物理世界的交互
尽管存在一厢情愿的想法或外推预测，机器人技术（Robotics）仍是一个棘手的问题，目前没有证据表明 Robotics 会迎来自己的“GPT 时刻”。GPT 或许会逐步加速 Robotics 的发展，但要真正提速仍需相当长的时间。在看到明确证据之前，我们应当假设机器人在完成人类与世界交互的任务方面表现不佳，且发展依然缓慢。
English Original
Our understanding of brain science and neuroscience is limited, so we can only seek answers from philosophy. By the way, unless there is a major paradigm shift in brain science, brain-computer interfaces like Neuralink are unlikely to achieve the functionalities people imagine. But we are not philosophy experts, so take this as a reference.
Judgment
No matter how capable ChatGPT is, it can only absorb digital signals and cannot interact with the real world. It might learn from ten thousand experts that doing A leads to B, but without experimenting in the real world, it can’t verify from the ground up whether this is true or false. True understanding of such matters requires personal experience.
“Eureka”
Newton saw an apple fall and discovered gravity, predicting the motion of the stars. Before Copernicus, everyone saw the sun rise and set and believed it revolved around the Earth. If ChatGPT existed then, it would surely assert that the sun revolves around the Earth. While it might predict how a peach falls from how an apple falls, it’s unlikely to deduce the motion of the stars.
However, people who discover things like gravity are rare. More importantly, it’s about identifying what this cognitive ability is and how it manifests in our daily lives. Archimedes shouted “Eureka” when he discovered the principle of buoyancy while bathing, describing moments of “sudden inspiration and insight.” This moment can be specifically described as “connecting several related points and discovering a third.”
This is actually two steps: the first is to discover new knowledge, and the second is to immediately understand its importance. ChatGPT may discover many things, including the principle of buoyancy (I’m not sure). But ChatGPT cannot know the importance of this knowledge or whether it’s useful to people. This step definitely requires human input.
Critical Thinking
Critical thinking involves actively distinguishing truth from falsehood, continually discovering more important, more realistic, and more creative and valuable ideas. ChatGPT might find a good answer, but whether it can continually challenge and improve upon that answer is uncertain. It’s a capability that OpenAI should continue to explore and develop.
Understanding Humans
Human textual knowledge certainly contains much understanding of human nature, but there are also aspects of human nature or preferences that are not documented in text. Combining this with a real-world understanding of people, rather than through surveys or online data, brings an incremental understanding that is a human advantage over ChatGPT.
Intuition
Returning to the essence of ARLLM. ARLLM tries to generate the next word, but whether humans are doing this is unclear. If it’s about generating the next sentence or paragraph, then perhaps the human advantage lies in generating the idea that comes much later. This might be what is called a “digital definition” of intuition. It’s uncertain whether the model can achieve this, but it’s likely difficult.
Interaction with the physical world
Despite wishful thinking or extrapolations, Robotics is still a challenging problem, and there is no evidence of a “GPT” moment for Robotics. GPT may accelerate the development of Robotics gradually, but it will take a long time to pick up the speed. Until we see clear evidence, we should assume that Robots are bad at achieving human tasks of interacting with the world, and the development is still slow.