LLMs force you to treat them like a pet

Have you heard the old proverb, "treat your servers like cattle not pets"? If not check out Pets vs Cattle, the history. In short, handling servers, we used to treat them like pets, we'd care for them, nurse then back to health when necessary, remove any bugs they may pickup, that sort of thing. As you can infer, treating servers as cattle is the opposite to this; you kill them when they're not on it, you herd them all as a group, individuals are disposable. These genAI tools we have need babysitting. how can you treat something like cattle if it's not even deterministic. surely, the point/goal is to be able to let agents spin up & do things concurrently, like threads. that cannot happen with something stochastic.

you have to stop them pissing all over what you're building at any moment

you can't just give them rules and expect them to follow. they are nondeterministic beasts, just like cats & dogs. If you aren't making sure, your dog will eat its own shit - which ruins a perfectly pleasurable walk. Same with the LLM code generators; if you press enter a couple times without thinking too hard on what it's doing, it'll literally shit all over the place and you won't recognise what you're left with. it's a complete misnomer to think you can let these things run alone. though, people do claim with enough tests you can, still nondeterministic, so this cannot be true.

we should start documenting our experiences because they're hilarious.

I just had a funny one, I fat fingered into a diff tmux session without realising, thought I'd exited it completely & continued to let the AI do what I had asked of it, without reading what it was saying. It got stuck and couldn't start the next server because the port was already in use. So, started another process for the server on a different port... This is the thing, "they" don't actually know what "they" are doing. "They" are a complex (some may say, convoluted) statistical model, built to generate the most probable next token.¹

Footnotes

1. Bender, E. M., Gebru, T., McMillan-Major, A., & Shmitchell, S. (2021). On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? 🦜 In Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency (FAccT '21), pp. 610–623. https://doi.org/10.1145/3442188.3445922 ↩