文章探讨大模型是否具备类生命特征

If a spaceship were to land in your garden right now and a stereotypical alien came out of it and started expressing themselves in natural language with capabilities similar to those of frontier LLMs, you would probably have no problem in declaring that a form of alien life just landed on Earth.

Now suppose you study the alien in great detail. You figure out that they are made not of flesh and bones, but of transistors. Their energy doesn’t come from thousands of small chemical reactions inside their body, but perhaps from some form of futuristic nuclear reaction. At this point you would realize that this is not carbon-based life. Their body is quite different from ours, but that does not make them any less alive than us.

Then you go on and study how their society and education work. You start to realize that they get their knowledge through something very similar to a mix of reinforcement learning and self-supervised learning, performed on a huge corpus of information that they initially inherited from another species, and which they later started expanding themselves.

At this point, would you shrug everything off and call this being not an alien form of life but just a stochastic parrot?

Just because in principle we understand how LLMs work and because we can create them from scratch, does it really mean that they are not alive?

Are you out of your mind? They are obviously not alive! They are just meat batteries!

As homework: try to imagine the opposite case. An alien form of life that starts mass-producing humans for some utilitarian endeavour. Would that make humans any less alive?