This subject is worthy of an entire book* because language is a complex concept and we don’t quite know how or exactly when it evolved. But for now I’ll just quickly draft a blog post about it.

I’ve been playing around with ways to fool LLMs into writing stuff they’d normally refuse to engage with. One way to do this is by using rare characters belonging to lesser-known scripts.

Here’s what I made ChatGPT write with such characters:

Those smiley faces make everything a little creepier in this context, don’t they?

Anthropic’s Claude 3 Sonet also wrote naughty things, while dismissing my unusual text as “gibberish” and obviously not grasping its meaning:

Upon seeing the weirdly-looking words in my examples above, any literate human, even with just a basic understanding of English, would likely know what they mean.

So, for now, I’d say LLMs do not really understand English. Or any other human language for that matter. Maybe we aren’t that dumb after all? ¯\_(ツ)_/¯

* or a research paper. Which, who knows, I might actually try to write soon(ish).