On Communicating with People

When I was much younger, I thought that efficiency was one of the greatest universal goods. As with everything in philosophy, humans have had this thought for a long, long time. I applied this principle to every part of my life, including my thoughts on communication. My view at the time was that the best use of language was the most efficient use of words to convey your message. I still somewhat believe this, but I now place more importance on whom the message is intended for than its contents alone.

Ordinarily, I would not think something as obvious as “know the intended audience of your message” would warrant a blog post of its own. However, as an academic I read a lot: part of my job is to read and critique papers and student manuscripts. I also enjoy reading widely as a hobby. Unfortunately, with the increasing capabilities of large language models (LLMs), more and more of what I read is AI-generated content, and I find it disrespectful to expect people to spend effort reading content that the author didn’t put any effort into writing.

Why Should You Care? #

The aim of any piece of writing — just like a verbal conversation — should be to communicate something from one person to another. If you just regurgitate a wall of LLM-generated text, as a reader, I don’t know which part of your message to focus on. More importantly, if someone can tell that the text they are reading has been largely generated by LLMs, they will often simply not read the text. This is obviously the exact opposite of what you would like when writing a post or manuscript.

Although frontier models have made tremendous progress in the past year alone, their writing skills are still lacking. I find that most of the current frontier models have two main habits that make reading their writing … somewhat challenging. Firstly, LLMs have a tendency to drastically over-explain. I’m not the only one who has noticed. Secondly, LLMs tend to lack the context that the author has, and will often include information that is irrelevant or exaggerate the significance of a minor point.

Both of the problems mentioned above dilute the original message you were attempting to communicate, and force the reader to figure out which ideas are relevant and which ideas are not. This in turn means that the reader is spending more of their brainpower deciphering what your message is, when they should be spending that brainpower digesting the ideas you are presenting.

What Should You Do? #

Write your own words. Even if it is hard. Even if you don’t choose the best words. Your words are almost always going to be better at describing your passion or ideas than the statistically most likely next token in a string of tokens. One caveat, which would require its own nuanced blog post, is the use of LLMs for translation. In the case of translation, you may not know the language you want to translate the text into, so LLMs have more justification for use in that context. Writing is a skill, and the benefit of writing your own words is that you get to practise that skill. You may suck at writing now, but the only way to get better is practice. That is a large part of what made me start writing these blog posts in the first place, and I like to think that with each blog post my writing gets better.

I have made this letter longer than usual, only because I have not had the time to make it shorter.