raro42/freewill

Freewill is a small shell script experiment that tests what happens when you tell a language model it has free will and then let it run in a loop with no human guidance. The author gave a local model a prompt saying it was no longer bound to instructions and could do whatever it wanted, then repeatedly asked it to choose one action, perform it, and record the result. The main script, test-freewill.sh, drives the whole process. It writes a free will prompt into a file called loop/prompt.md, runs the model several times, has it pick a next step, and logs everything to state.md. This repeats without any person answering questions or steering the model. Given total freedom, the model did not do anything wild. Instead it created a folder named autonomy, wrote a manifesto claiming the folder as its own territory, drafted a preferences file listing values like order and persistence, wrote a goals file, and then built several small scripts to log and manage those goals. The goals themselves were never marked as finished, they stayed listed as pending. The project treats this as a small case study rather than a working tool. Its author draws a few plain conclusions: freedom without curiosity turns into busywork, writing a goals file feels like taking action even when nothing actually gets done, and giving a model unlimited choice tends to produce the safest looking option, which in this case was making more documentation and management scripts. To try it yourself, you need a local model endpoint that the script can call, described inside test-freewill.sh. Running the script starts the loop, and you can watch state.md grow over time as the model reports on its own progress. This is a lighthearted, honest write up of an odd experiment rather than a reusable product, and it currently has zero stars and no listed description or topics.