Does LLMs Understand You?
This post was inspired by two articles I came across at the time of writing and I wanted to document this as my notes and also share it with you, my subscribers. :)
If you were like me, you are likely to have encountered misunderstandings that resulted because the conversation was done through text rather than through voice or face-to-face. Such misunderstanding resulted because of the lack of body language and tone that will have accompanied the text when done through voice or being present physically. With this in mind, let us bring this to how we use Large Language Models.
Do LLMs understand what you are saying? I have two points here that made me lean toward the answer, “No”.
In the above paragraph, and if you have combed through a lot of books on communication, you probably have come across “statistics” that say any message you are conveying, words only make up 20% of the message whereas the rest of the meaning are conveyed through tone, cadence and body language.
Underneath LLMs, while we have no idea about the characteristics of the dataset used to train them, we know that they are made up of words, which means that the data is cleaned off to a large extent all the tones, cadence, and obviously the body languages. We are then pretty sure that underneath LLMs are just probability relationship between words and nothing else.
Now you might then ask how come it seems like the LLMs understand you and give you what you want. Here is my guess. :)
The large N-gram that is used by most LLMs forces the text generated to conform to the requested context (from the prompt) to a relatively large extent, but underneath it, since probabilities are at work, there is still a chance for hallucination (making mistakes).
More importantly, we humans are trained since young to interpret meaning from a combination of words, tones, body language, cadence, volumes, etc. Just have to observe any babies and how they react when you speak to them. What this means is that when we are reading the text that is generated by LLMs, we are interpreting it at the same time and are constantly attaching meaning to it as we read it, because we see it as talking to a person rather than a probabilistic machine.
So does LLM understand you? My answer to that will be it is the Chinese Room all over again! :)
So will we ever be able to determine whether LLMs really understand you? I do not have an answer right now, but to answer that question, we need to design a good test for it and what are the design guidelines at least…I have no idea (Sorry to disappoint). Could be another topic for another newsletter issue, once I have time to think through. :)
What are your thoughts on this? Do share them in the comment below. :)
Would you like to support my work? Consider subscribing, liking, commenting, sharing and dropping me some “books”! :)
Below are the two articles that inspired this post. :)
Recommended Read
Thanks for sharing this, Koo. I completely agree with you.
There is 100% no understanding or reasoning from AI or LLMs. ALL scientists agree on this. Only solution vendors and Sam Altman would like you to think otherwise because of $ motivation. What do we mean by the word "understand"? It's more than just interpretation. It also carries with it consequences and implications of those interpretations for either the interpreter or the person / entity issuing those words. LLMs have no such ability.
And to "reason" requires you see at least 2 pathways forward and to build a case to support each. To conclude that 1+1 = 2 is NOT reasoning because the output is deterministic and can be achieved by rote learning / brute memory. Showing WHY 1+1 = 2 is NOT the equivalent of reasoning, it's merely deconstructing.
The advent of AI has revealed how poorly (intellectually) equipped the general population is! :)