I had a couple of reasons for wanting to build this. First of all I wanted to try to find some ways to actually enjoy working with AI again as I’ve been getting quite tired of all the doom and gloom that constantly circulates. I wanted to go back to that time where you would just think of those crazy “what if” ideas but then try to use an LLM as a collaborator and stop getting all hung up on the end of the world stuff and just have fun.
So my experiment that I wanted to run was seeing how complex of a system these more modern language models are capable of building when you try to strip things down to very small primitives (in this case only a small 32 instruction set with limited resources as well). You can run experiments to limit the resources (# of turns, RAM, Storage, simulated CPU frequency) and see how far the agent can get. There are a few modes as well such as turn based where you just give the agent, say, 20 turns to see how far it can get, or the more interesting mode (in my opinion) which is to let it run until it hits resource limits so that it is forced to try to make the most of the resources it has.
Some of the fun things I’ve prompted the agent to try and do: - Build a programming language and then use it to build a simple terminal
- Make a basic engine for rendering graphics
I found it really interesting to watch the agent slowly build out primitives to render text into the display and then build its own language and watch it try to test it to make sure that all the arithmetic was solid. I’m sure there are lots of other cool things to try out as well!
Hello HN!
I had a couple of reasons for wanting to build this. First of all I wanted to try to find some ways to actually enjoy working with AI again as I’ve been getting quite tired of all the doom and gloom that constantly circulates. I wanted to go back to that time where you would just think of those crazy “what if” ideas but then try to use an LLM as a collaborator and stop getting all hung up on the end of the world stuff and just have fun.
So my experiment that I wanted to run was seeing how complex of a system these more modern language models are capable of building when you try to strip things down to very small primitives (in this case only a small 32 instruction set with limited resources as well). You can run experiments to limit the resources (# of turns, RAM, Storage, simulated CPU frequency) and see how far the agent can get. There are a few modes as well such as turn based where you just give the agent, say, 20 turns to see how far it can get, or the more interesting mode (in my opinion) which is to let it run until it hits resource limits so that it is forced to try to make the most of the resources it has.
Some of the fun things I’ve prompted the agent to try and do: - Build a programming language and then use it to build a simple terminal - Make a basic engine for rendering graphics
I found it really interesting to watch the agent slowly build out primitives to render text into the display and then build its own language and watch it try to test it to make sure that all the arithmetic was solid. I’m sure there are lots of other cool things to try out as well!