https://futureofcoding.org/ logo
Join Slack
Powered by
Has anyone experimented at all with encoding progr...
# of-ai
s
Has anyone experimented at all with encoding programs/messages into LLM output? Over the weekend I put together this little proof of concept where if you can encode programs into normal looking prose with an LLM and I'm curious if anyone else has played with this before...
i
i'm curious how this is different from the more general idea of encoding programs into prose.
s
It's the same idea, the difference is mainly how easy LLMs make it to create coherent prose.
t
a computational notebook with an LLM? There are lots of examples of this but maybe I do not understand your angle properly. Specifically on
how easy LLMs make it to create coherent prose
Very hard in my opinion if you want it to be readable.
d
One interesting approach was to essentially take a deterministic LLM and, for a given seed/start of sentence, let each word be the Nth most likely completion for that point. This hit issues with sentence fragments which have very few logical completions, like "it was raining cats" where anything that isn't the most likely word feels incredibly unlikely. Neither aclanthology.org/D19-1115.pdf nor aclanthology.org/2021.findings-acl.268.pdf are the paper I'm thinking of but they both look like they're in the same space.
s
So my angle is more: if you create some rules for encoding the number of letters in specific words in a sentence and the number of words in a sentence into bits, and also define a specific assembly language that maps opcodes and operands to bits, you can get an llm to write a program with your assembly language and give you a table of the sentence attributes you need for your prose to be interpreted as your program
And the table can be given to another LLM and say basically write me something on a specific topic following these attributes and even just let it keep looping to get there. This was a table for an early iteration of the language for a program that output Fibonacci numbers
I've been messing with the prompt, language spec and the mappings to open up the possibility space of the prose…with something like this they could even typo different words in order to use words that sound more natural if needed. So nothing that you really couldn't do pre-llm, but the amount of effort you might have needed to put in to do something like this before them would have been too high to use too often…but now you could essentially have all your communication have hidden messages if you'd like
Open in Slack
PreviousNext
Powered by