Recently I discussed with someone how they expect superintelligent AI to find a flawless, perfect language for thought and communication, without any useless ambiguities: a language that’s “better” in every way. I disagreed with the premise - language is a tradeoff between specificity and verbosity, where an increase in precision incurs a cost either in message length or your character set. (You can’t describe a higher-entropy thing without using higher entropy!) In drilling down further, he then talked about how there’s vague, nebulous feelings and intutitions which he has, that he has trouble putting words to, and that this language would be that thing which empowers him to fully express his internal state to others.
Did you know that there are books which have all the words in them?
Reading the dictionary isn’t cool. But if you genuinely think the language you’re using is foundationally limiting your thought, in a full Sapir-Whorf sense where cognition is bounded by words, then you have the power to learn more words. And if that is not enough for you: you can make up more of them! Fields as formal as mathematics to as qualitative as philosophy all have their own jargon terms; words with highly specific meanings for obscure concepts, that let you talk about them in a less clunky way. Some of these might not obviously help you communicate - perhaps you don’t care to distinguish biltong from bildungsroman - but if you truly think language is the limiter on your thought, why aren’t you constantly learning more words to expand that frontier?
Why does it feel like language is a barrier to thought, especially (at least, from my sample group) for those in STEM? Or at least, why is it easy to claim that? I feel like there’s a few parts to this. First, that there’s a long history of science fiction stories where this is the case. In Iain Banks’ Culture series, the Marain language is exactly what my friend fantasizes: a constructed language by hyperadvanced AIs to give its speaker clarity and expressivity, enabling higher-level thought with trivial ease. In Orwell’s 1984 the manipulation of language limits possible thoughts by forbidding conjuring of illegal jargon to refer to them. And in the extreme, Ted Chiang’s Story of Your Life a.k.a. Arrival has a language which directly manipulates perception in a more extreme manner.
The problem with generalising from science fiction is that these concepts do not need to be Real Things. We can talk about a language where learning it gives the speaker supernatural abilities, because the language is the language of magic, and speaking a thing in that language makes it so. But that doesn’t mean that there is by necessity a real language with those properties!
Second problem: language being intrinsically nebulous is taken to be an inefficiency to optimise away, rather than a fundamental property of The Thing Language Is. Maybe a better way to think about language, for this purpose, is as a lossy compression of thought. When you have a thought and wish to communicate it to someone else, language is the tool that you need to use.
But we don’t have to communicate just with words! When I point to something and say “that thing”, I have communicated additionally by the pointing; such gestures are deictic and metatextual. I don’t need an infinitely rich language of colour to point and say “that colour”, and it saves half an hour splitting hairs betwixt mauve and taupe. If I need to show a correlation, I don’t waste words if I can just show a scatter plot. Language is not the only communication medium we have - if you treat it as All You’ve Got, then you’re probably Sapir-Whorfing yourself into a hole needlessly.
The third part is that STEMfolk seem to view language as “lesser” or “the other” due to this nebulousness. Hell, we had to conjure up prompt engineer as a job title to wiggle out of calling LLM guiding via prose composition by its obvious extant name - writing. Some of this is the idea that if it can’t be expressed explicitly and algorithmically, it may as well not exist. This feels to me like saying meteorology is a sham because meteorologists can’t answer such simple questions as “how many clouds are there in the sky right now”! When in practice the idea of distinct “clouds” suffers from the coastline paradox, where your clustering method determines your count number. There isn’t any “true, underlying number of clouds” any more than there is a “true coastline length”. And this is because “cloud” is a usefully messy boundary, that we keep messy in the same way as “some” or “a lot” in casual usage.
Annoyingly for tech-types, it’s not straightforward to “read” Wiktionary from start to finish, though you can flick through the “random” page repeatedly. I wrote a browser userscript that turns Wiktionary and Wikiquote into infinite scrolling pages, which avoids revisiting seen pages to keep things new. (Without that filter, I’ve ended up with the same words appearing unusually often, since the “random page” selected is only pseudorandom.) I keep the random pages for Wiktionary and Wikiquote this in a bookmarks folder with my other daily-check sites. If I already know the Wiktionary words, I can just keep scrolling down until I learn something new. And so I’m constantly expanding my vocabulary.
There are downsides to this. You’re likely to start using atypical words, and then need to explain what they mean to other people. Expanding your vocabulary might help you introspect more effectively and precisely, but this might come at the cost of language as a social tool for communication. The deeper you get into jargon, the less approachable your sentences become, and you end up spending more time explaining word definitions than the things you wanted to communicate. I feel like Wikiquote can counterbalance this, with its examples of especially pithy phrases which are concise but universal. And it serves a similar purpose: often the quotes outline vague feelings you’ve not put words to before, and give you further introspective power.
I think people are using ChatGPT for this kind of introspection now, and that’s part of why it can seem so compelling. When there are feelings you can’t put words to, or worries you have trouble expressing, LLMs can bounce and percolate those things and help articulate and clarify those thoughts. And that’s when the tools are at their best, I think - when they’re helping users better understand themselves. But it’s also useful to build up a repository of words and introspection tools for yourself, so you don’t need to pull out a chatbot every time you want to notice your own feelings.