Hi All,

So after this discussion, I've been back and forth between work related
projects and fun ones, ironically both relating to using AI-coding agents.

I decided to use ChatGPT and GPT-Codex to make a game from scratch for the
M100. Honestly, it's just a prototype at the moment, because I haven't
figured out how to exactly make it fun yet. I'll explain what I did.

1. I used ChatGPT to brainstorm on winter-themes games that could possibly
run on the M100. We settled on a game where you have to shovel
your driveway. It's essentially a tile-moving game with stacking mechanics
/ templated obstacle overlays.

2. I used GPT-Codex to write the entire thing in BASIC first. I created a
Codex-Skill which it is able to invoke as needed. It can add to its own
context relevant information for whatever task it's working on at the time.
I included various documentation I converted to markdown.

   - 8085 Op-Code Manual from Intel
   - BASIC Keywords for M100 - and their descriptions on how to use them
   - General approaches for integrating BASIC with Assembly
   - Sim8085 (an Intel 8085 CPU simulator written in Python) - Which Codex
   can use to evaluate its ML subroutines, and assemble them.
   - Documentation from the Bitchin100 Wiki (System Memory Map, Variable
   Memory Layout for BASIC, etc)

3. Most of the issues come from the fact that Codex cannot directly run
BASIC. Thus it does things like creating test programs in Basic, print
debug info onto the screen, and ask me to read it back to it when running
in VirtualT. If there were a way to expose virtual-t to Codex directly, and
allow it to "see" the screen prints, and provide keyboard input, it would
be able to iterate on things by itself.

I think many would be surprised at how well it performs given the
limitations. A $20 a month GPT subscription goes a long way with Codex. It
used up around 80% of my weekly token limit making this game.

It was able to rewrite some of the slow-initial generation loops in ML and
integrate them. Likewise I was able to understand the ALTLCD print routine
I used in Text Sweeper, and incorporate it into this game, as well populate
ALTLCD.
When I started to get concerned about the amount of RAM the game uses, GPT
was able to suggest bit-packing together two separate arrays into one, and
then modified the ML subroutines and BASIC to handle the adjusted structure.

Overall the experience was positive. I would say with High-Thinking
enabled, complex problems took a long time to solve (5-10 minutes) but
Codex was able to iterate, and test, and iterate and test, until it was
happy with the outcome. To the point where I went downstairs, did the
dishes, made a cup of coffee, and came back upstairs to review the work.

Feel free to try it out from the release here, already minified and
tokenized:
https://github.com/Grimakis/Snowed-In/releases/tag/v0.1.0

If you want to see the readable source:
https://github.com/Grimakis/Snowed-In/blob/master-protected/SNOW.DO
https://github.com/Grimakis/Snowed-In/tree/master-protected/assembly

-George






On Fri, Dec 12, 2025 at 4:13 AM B 9 <[email protected]> wrote:

> On Sun, Dec 7, 2025 at 11:37 PM Joshua O'Keefe <[email protected]>
> wrote:
>
>> In any case, if you'd like to get in touch I'm happy to help. One of my
>> hobbies a few years ago was helping folks with normal, small-scale home
>> systems be able to do LLM inference with readily available tools. For
>> everyone else: I'll let folks know if anything list-relevant comes out of
>> poking some of my code generation models. I'm... not especially hopeful.
>> The exercise is fun, though.
>
>
> I'm very curious how this exercise will turn out. It is an interesting
> challenge as proper programming back in the day often involved discovering
> clever, problem-specific optimizations to save space and time. Our Model
> T's are compute engines from that more civilized age when people
> handcrafted elegant assembly code that was precise, succinct, often
> beautiful, but almost never reusable. Generative AI works on statistics and
> such idiosyncratic patterns may be difficult to discern. Even worse, the
> code was rarely well commented, and so the mapping from natural language to
> implementation may be especially difficult for an LLM.
>
> While training on BASIC (or C) would be easier, it is the Model T's
> machine language which is the most tedious to code and where an LLM would
> (theoretically) be the most helpful. I agree with John that the lack of a
> large assembly language corpus may be a significant barrier. (There's the ROM
> disassembly with commentary
> <https://github.com/eriktier/RomT102Disassembly/blob/master/t102rom.asm>,
> but what else?)
>
> Please do keep us posted with any developments.
>
> —b9
>

Reply via email to