A large language model is a fancy web search

I saw this figure describing them as “solving real-world tasks,” which I strongly disagree with. An LLM does not solve tasks any more than the K&R C Programming Language book makes C programs.

If you want to write a function comparing two strings you can look it up in K&R and there is an example of such a function. Much the same with the LLM robot: I write to the robot that I want a function comparing two strings, and the robot shows me such a function. But the LLM provides information on much more than K&R. It also lies much more than K&R.

To me, the LLM is basically a lossy compression of the internet (or whatever it was trained on) in which look up information by an interactive conversational interface. You had to trade some accuracy to get this interface.