If annoyance in software marketing had to be summarised, I guess it more or less could be done using a single word - CoPilot. Ever since someone at Microsoft came up with the term, the entire Microsoft marketing has been after people with that stupid logo that does not make sense against its name. They have it in the Start Menu on Windows, on the taskbar, on the notification area, on the sidebar in edge, in Office apps. If that is not enough, you get reminded that it exists, using notifications, by the OS itself.
TL;DR version of this post
After being persuaded by Windows way too many times, I gave CoPilot a try to solve a small problem that I was having with writing shell scripts. It could not solve the problem I had. But when I told it back “thanks, but I solved it”, it asked me to share back how I solved the problem. I did not tell it back. And that gave me an insight (which I have mentioned in the summary section).
State of LLMs
Before I gave up on LLMs for anything productive, I had actually given them a try. They are all 💩! They constantly lie, creatively and confidently. They make up stuff. They write bad code, use non existent functions and libraries, can’t write configuration files and what not. Yet they are all useful. Often, if I am starting to learn a new tool, concept or technology, the bigger problem is not that I don’t know the tool, it’s that I don’t know the terms around it. Like, if I am new to nginx, it is crucial to know that the term “server blocks” is the same as “VirtualHost” from Apache Server world. They give me terms that I can search around with. That’s the real value I have derived so far - learning the terms around a tech.
ZSH is awesome
If you are reading this, you probably wanted to read the longer version. So here is what I was doing. I was building something using Shell scripts. I don’t remember what it exactly was but it must have been either Zshy or one of the extentions I was working on.
Now, with ZSH and LLM, I was already skeptical because most LLMs back in the day told me stuff like “ZSH allows you to run your last command starting with the prefix you have already typed on the shell by pressing the up button”. That is NOT a ZSH feature. That is a OhMyZsh feature. OhMyZsh is an extremely famous ZSH shell plugin manager and framework. But LLMs are trained on the data that is already there on the internet. So they got trained on the shitty blog posts that wrote it that way.
I did not have much hopes anyway that CoPilot would be able to help.
Let’s give CoPilot a try
One of the reasons Google Chrome is so successful is not because it is simply superior. It is also because at the time when it was born, Google’s homepage (google.com) was something almost everyone typed in the address bar before anything else. And Google, being Google, would give them a sweet little button on the top right corner asking if they wanted to use Chrome. I ignored it for hundreds of times before I clicked it once (and was mind blown and I switched to it and stayed on it for about a decade before moving back to Firefox).
Similar thing happened with Copilot.
So I was facing trouble and I felt a bit lazy. After some Googling and failing to locate the right solution, I eventually thought “I am already paying for it, let me try to use it” (I have a Microsoft 365 subscription). And I lousily told it my problem. It gave me a lousy answer. Not much help upfront but with iterations of stating my problem more clearly, trying out suggestions it gave me and making a little bit of progress, I finally found enough hints to rewrite a part of my code that solved my problem.
Great. So what did I do? Well, CoPilot helped, right!? I gave me hints to solve the problem so I went back to it and like a good samaritan, I told it a “thank you for the help, I solved it myself”.
It said something like “That’s great. I am glad I could be of help. Would you like to share how you solved it?” I was in no mood to tell it though and I thought “I should have closed the window instead of thanking it”. But staring at that response for about a minute, a thought came to me which I have summarised below.
Summary
Artificial Intelligence models, without exception, are “trained” on existing data set. LLMs that can write code were trained on code. But the way I see things moving, I see an increasing trend of the behavior that I had shown myself - I went to a LLM (CoPilot, in this case). A human never saw my query. A human never answered it. The LLM asked me for the solution but it did not get any. Most people would just close the conversation and not give it that feedback (and they should not!?).
A piece of knowledge was gained by me but was lost to the world, and for good reason.
The solution to my problem is now known to me. Anytime the problem comes back at me, I will use that knowledge. But no one else will know about it. There is no post on stackoverflow, no email, no discord message anywhere and I did not tell the LLM the solution either. So no AI model will be trained on that knowledge.
In a time when people are just VibeCoding, learning something, discovering something about how to solve a problem is the real deal. I now know something about solving a problem in ZSH that no one else publicly does. Maybe that is where the value is going to shift towards - in knowing and being able to do stuff that LLM’s can’t figure out.