The genAI conundrum

Up until this point (including this article), I believe I have never used any genAI tool to write on this blog. Not that I don’t use ChatGPT to help me brainstorm or anything, but I have just never felt the need to do so on the blog. The only tools I use are spell and syntax checkers (I like Grammarly (not a sponsored ad (if you want to sponsor me, you’re welcome (this is turning into LISP)))) to help me fix the inevitable mistakes that my flows of consciousness usually contain.

I usually enjoy writing on this blog, as it reminds me of my internship at the local newspaper back in high school, but lately it feels more like work than anything. To me, blogging has always been a way to organize my notes and help me keep track of what I do. Lately, however, the few notes I take are increasingly dense and unreadable, completely lacking in structure behind what is strictly necessary for understanding the content. Just like with crappy, uncommented code, context usually lives in my mind for about an hour (or until I get preempted and start working on a different task), often making the notes themselves useless at best and unreliable at worst.

As such, I’ve decided to start using AI to help me organize my notes. Since this does not sit well with some people, any article that contains AI-generated text will bear the following header:

This article was written with the support of generative AI tooling. The author assures the reader that he is aware of the article’s contents in its entirety, having reviewed it and almost certainly extensively modified it until he deemed it suitable for publishing.

I believe there’s nothing wrong with using genAI to help you do menial tasks, like rephrasing a paragraph or reformatting a table. AI tools, for one, have saved me countless hours of sifting through log files, generating tables, generating BibTeX entries given unstructured data, etc., albeit with one big caveat: you must already know what you’re doing.

If you feel like this Golden Retriever, then genAI is the wrong tool for you. It may help you understand a topic better, for example, but only if you can detect when it’s clearly making bullshit up because it sounds good. Please try to remember that there’s nothing inherently intelligent about genAI (thankfully), and that it’s just a very good inference engine (i.e., it’s good at estimating a distribution) trained on an insane amount of data. Since it cannot really “understand”¹ the relationship between data, there will be instances when the estimation will be completely and utterly deranged.

Of course, people make stuff up all the time as well, which makes it even more likely for genAI tools to give inexact explanations. This doesn’t necessarily mean that the model is bad - it’s likely down to the fact that the internet is full of people who enjoy explaining things that they don’t understand², and Redditors who’re desperate to write an essay about why you’re wrong without actually knowing anything about the topic at hand (they’ve just read up about it on Wikipedia). Heck, probably they’re not even reading it up anymore - having ChatGPT write the snarkiest comment in human history is way more efficient.

To sum up, I believe AI is like the Force (or those jokes that begin with In Soviet Russia…) - if you don’t learn how to use it, it will end up using you. As master Yoda would put it:

…or something along those lines. Yadda, yadda, yadda. I still prefer the dark side.

1. As far as I know, the model determines the relations between the embeddings by giving them closer coordinates in a virtual 3D space (the latent space) and literally clusters them together. People who understand the mathematics behind AI will probably cringe at my descriptions. I am sorry. ↩

2. Just like the author of this article (i.e., me) is doing right here, right now! You’re welcome :) ↩