More than Word Guessers
Part 2, the excitement continues
I wrote about us being next-word-guessers recently.
We are also word predictors
If I say “Why did the chicken _____?” Or “See Spot _____” Or “It was a dark and stormy _____.” Or “Tall, dark, and _____” I bet you’ll know what the next words would be. That could mean that we use clichés too much. That’s one possible reason. But I also think it hints at something about our brains too.
I had some more in-depth understanding of this that I need to share.
When we speak there is more to what we’re saying than just guessing the next “correct” word.
What we say has different layers which I think is very fascinating and very important. I think an LLM may be able to get some of the layers, but it can’t neither understand nor reproduce all the layers.
First of all, we think in ideas, not words.
Try learning a new language and for a while you’re forced to think in words. You’ll learn very quickly that your brain doesn’t normally think that way. It will very odd and clunky at first.
Once you’re fluent—and probably at some point after that even—you’ll be able to start thinking in ideas and that language will come out without having to think about which words mean which idea.
It’s a cool experience!
Next, we speak because we want to. Nobody can make us create words. I mean, we can be forced to speak, but not made to speak. Don’t go down that rabbit hole … yes there are exceptions, but no they are not pretty.
That shows our will.
We decide when we speak, what we speak, and how.
An LLM can’t. It’s made to “speak.”
Next, there is a reason or a purpose for everything we say. Even if we don’t know it sometimes and even if we accidentally blurt out something that we wanted to keep inside—and that speaks volumes in and of itself right there.
And often there are more than one reason for saying something. Our words have can multiple meanings.
Next, we speak with emotion.
Everything we say has some emotion behind it. Even sharing a spreadsheet of data—very robotic—has emotions behind it. Why is this person sharing the data?
And yes there are exceptions to that as well, but that also says a lot about what’s going on.
And just like with our reason for saying something, we can have multiple emotions about something we’re saying as well.
Lastly, (or the last layer that I can think of. There are probably more) is an understanding of the audience.
We will say the same thing differently if we are speaking to squirrels or to ourselves or to our best friend or to someone we fear.
Audience matters.
I think an LLM can understand and replicate this one fairly easily.
What other layers do we speak with?
How well can LLMs replicate each of these layers?
Your observations about human communication are spot-on. Our current capabilities in replicating speech focus primarily on mimicking what is said, rather than delving into the nuances of how and why we express ourselves. The essence of communication is far more complex, rooted in our emotions, thoughts, and the ideas we wish to convey. Words serve merely as the vessels for these ideas. At present,LLMs fall short in capturing this depth, struggling to replicate the human experience.