No LLMs
~ 3 minutes
Published
I don’t use LLMs, not for programming, not for writing, not for anything else. The code i write, the programs i develop, the articles i craft, all of them are written by hand, and no outside pressure will change that.
Why avoid LLMs
Even though some LLMs produce impressive result in some aspects, their existence hurt every part of the internet, software development and society i care about.
Articles and write-ups online tend to be generated more and more, and their content and factuality suffers dramatically. Search engines are less useful than ever, because the results they show are just “good enough” computer guesses of the topics they advertise to inform you on.
Software which is created using LLMs tends to be buggy, incorrect or inefficient - just like the code they’re trained on. Simplicity, correctness and performance are difficult topics, which even experienced and talented engineers struggle with. A “token prediction machine” does not have the capacity to reason about any of these in a meaningful way.
The computing power necessairy for these statistics programs is unreasonable. We thought crypto currencies were bad, but LLMs are even worse. GPU got expensive really quickly, now even memory is unaffordable, so hold on to whatever you got.
The energy consumption of LLMs is unsurprisingly very high, which in and of itself is not that bad. But how the energy is produced, and how the resulting heat is dealt with, is worrysome. In Switzerland, the heat produced by datacenters is used for district heating. Water consumption is negligible, because in closed loops water does not evaporate. Some datacenters even do without air conditioning. But most of the datacenters in the US use evaporative cooling, consuming drinkable fresh water in the process.
“More productive”
Productivity was never an issue, quality however was - and is. If you think the speed at which you vomit out new code is the limiting factor for your or your companies success, then you probably never had a good manager or talented and experienced coworkers. Think about the success of SQLite, small, fast, reliable, and the most deployed piece of software in history.
“New skills for everybody”
No, not true. If the computer does something for you, you don’t have that skill. I can’t draw just because i have a printer. I just can make shitty copies of drawings with it.
“It’s just a tool”
Using LLMs makes you worse at whatever you use them for. To be reasonably good at something requires practise. If you learn to play the piano, you dont let it play the easy songs and only play the hard ones yourself. Tools help you do something, they don’t do it for you.
Yes, it’s debatable if LLMs are “just a tool” or not, but using them can reduce your skills, so be careful.
Personal dislike
The biggest reason why i don’t use LLMs is very simple: i like simplicity, efficiency and correctness - none of which describe LLMs in any way. Also, i do programming in my spare time because i like doing it, why would i want the computer to do it for me? (Same with writing.)
Feedback
If you think i’m missing something, or you have a good reason for me to reconsider, let me know.