@mos_8502 When I was setting up to see what the 6502 compiler output looked like, I created a simple factorial function, and set up compiler toolchains for llvm-mos and calypsi. After comparing the generated assembly for both, I wanted something larger to work with, so I used claude with this prompt:

create a c program for the commodore 64 that bounces a sprite around the screen.

It initially generated code that used cc65 and included the c64.h file, but when it updated the Makefile, it "realized" I was using llvm-mos, so it removed the header and defined macros for all of the registers it accessed.

The code was decent, it was a single main function, kind of long, but not horrible. The sprite data it generated was a little more diamond-ish than I wanted, but when I compiled it and ran it in vice it worked.

One thing I did change in the code was that it had repeated the code to cycle the color of the sprite when it bounced, I moved that to a new function instead.

@mos_8502 Also, when openclaw was first released, I set it up on a "spare" mini-pc that I eventually plan to use for a CNC mill, but haven't created the brackets to mount it yet so it's been sitting idle on my desk.

I set it up with no access to anything, and let it use an LLM running on ollama on my laptop. I can only run smaller models there, so the "assistant" was kind of dumb and got things wrong more often than one of the big models might.

It was an interesting experiment, in that I learned a lot more about running local models, what the different quantization choices do, and setting up a telegram bot so that I could communicate to it from my phone.

I tried to get it to set reminders for me, so it could ping me on telegram when something needed to me done, but the weak LLM it was using kept messing that up. Again, I suspect it could be a lot more useful with a stronger LLM, but I wasn't willing to put money into running it, or willing to trust it with credentials for anything real.