Building Llamas That Think?
So, casually trying to make LLaMA achieve consciousness (as one does on a Tuesday), when I stumbled upon something hilarious. Turns out, you can make these language models "reason" with about as much code as it takes to write a "Hello World" program. No, really!
https://colab.research.google.com/drive/1jfsG0_XP8a5mME76F5a6xeP9uu-tvDZ3#scrollTo=fNg798sHpuqk
Here's the philosophical crisis I'm having now: When you ask an LLM to code something vs. asking it to reason about something... are we basically watching the same neural spaghetti being twirled around?
The real question is: If one can make an AI model "think" with 5 lines of code, does this mean: a) An LLM should be able to write its own reasoning code b) We've been overthinking AI? c) The simulation is running low on RAM d) All of the above
Would love to hear your thoughts, preferably in the form of recursive functions or philosophical paradoxes.
P.S. No LLAMAs were harmed in the making of this experiment, though several did ask for a raise.