A project which puts LLM-powered gremlins right into your runtime! hmm

As previously mentioned, I enjoy trying out new tech like claude-code, making esoteric and ill-advised libraries that are maybe-jokes, pushing on meta-models and run-time eval, and like many others I don't mind 5 minutes of fame. So by that measure, TOTAL SUCCESS!

Let's start with what this new Gremllm library that I made can do:

from gremllm import Gremllm

counter = Gremllm('counter')
counter.value = 0
counter.increment()
counter.increment(5)
counter.add_seventeen()
print(counter.current_value)
print(counter.value_in_hex)
counter.reset()

To be clear -- each of those methods is completely made up. At runtime when you invoke a method it is at that moment sent over to the LLM (ChatGPT in this case) to generate the python code, which is then executed. It's like claude-code, but at runtime!

Speaking of which, I built Gremllm using Claude Code, which I've been using a lot lately. I'm not always convinced that it is faster, and I'm not always convinced that it is slower, but it is fun enough that I am making stuff that I might not make otherwise at the very least.

Using Claude Code did have some trickery. The biggest thing is that it kept wanting to run my eval (or in Python it is actually exec) in some sort of sandbox, and I had to be like "no... please do this dangerous thing. Really!". Even now it is doing some things that are more safe than I would have done if it was completely written by hand.