大型语言模型（LLMs）如ChatGPT，展现出类似外星智能的特质，拥有成熟的语言能力和初级的世界知识。人类与不同智能体如动物、智力障碍者及自闭症患者的互动经验，启示我们如何与AI共处。丹尼尔·丹尼特的“意向立场”理论指出，将机器拟人化有助于我们更轻松地与之交互。此外，我们采用还原论、整体论等不同立场来理解世界，而视觉和语言理解属于“简单”的、前科学的任务，机器学习正是通过收集相关性和基于少量证据得出结论来实现这些任务。

🤖LLMs如ChatGPT具有类似33岁大学毕业生的语言技能，但世界知识却如同3岁婴儿，这种差异源于当前技术和资源的限制，未来将会有所改善。

🧠人类与各种不同智能体（如动物、智力障碍者、自闭症患者等）的互动经验表明，我们能够适应并与不同类型的智能进行交流。

👤丹尼尔·丹尼特的“意向立场”理论提出，如果将机器拟人化能使我们更容易与之交互，那么就应该这样做，例如给汽车起名字或认为Roomba害怕某个房间。

🌐我们采用不同的立场来理解世界，包括“还原论立场”（科学方法）和“整体论立场”（直觉和经验）。视觉和语言理解属于“简单”的、前科学的任务，无法用科学方法完成。

💻机器学习通过收集相关性和基于少量证据得出结论，使机器能够执行这些“简单”任务，类似于婴儿学习语言和视觉的方式。

LLMs such as ChatGPT are alien intelligences. They have the language skills of a 33-year old college graduate and the world knowledge of a 3-year old infant.

In humans, our various basic skills are rarely as separable as they currently are in our machines. TBH, at the current state of the art, a computer learning language is major victory achieved at enormous cost, taxing the limits of our global computing capabilities. We don't have the resources to let our machines learn math and physics. Yet. Future generations of machines will. And better algorithms, like mine, will help.

But we already have lots of experience with alien intelligences.

We seem to be getting along fine with dogs, cats, and parrots. And mentally handicapped people. And infants. And mentally impaired elderly. And significantly less or more educated people than ourselves. And (asking for a friend) many people are on the autism spectrum, which may also require "adjustments to interaction protocols".

So it is not a problem. We adopt to other kinds of intelligences all the time.

Daniel Dennett in "The Intentional Stance" explains that this ALREADY extends to machines. His definition is elegantly backwards from conventional thinking about how to relate to machines:

If it makes YOUR interaction with a device EASIER, then by all means, treat that device as if it has intentions.

"My car is mad at me and won't start. I need to let him cool down a bit." If you have a name for your car, you are already doing this.

"That damn Roomba is scared of my bedroom"

This position is what he calls "Adopting an Intentional Stance".

There are other such stances we can adopt. The two biggest and most famous stances are "The Reductionist Stance" where we do things scientifically, and "The Holistic Stance" where we only need Understanding of the problem to do "simpler" things without science, planning or forethought. We "just do it", to coin a phrase. This is how we do nearly everything. Pre-scientifically.

As it happens, vision and language Understanding are two of those "simpler" things. These can not be done Scientifically. We learn these as infants, before taking any Science classes.

And Machine Learning is simply the recent discovery of how to make machines do such "simple" things pre-scientifically.

By gathering correlations and jumping to conclusions on scant evidence.