Who can create the best version of Pong?
After having trouble getting the 20B version GPT oss to load locally on my Mac Mini, I decided it was time to upgrade and see what you can accomplish on consumer hardware. I wanted to grab a 64GB studio—but ended up with a brand new 32GB M2 Max Mac Studio I found for $899 on Slickdeals.
It's awesome.
I had run a lot of local models via LM Studio previously, but it was mostly for fun little conversations to see if I could get them to doom loop with riddles and discover more about the failure modes of smaller models. I didn't really think you could get much use out of running local models, and honestly, it's still challenging, but you can really accomplish some cool and meaningful tasks that would have been unthinkable for a computer to do itself a few years ago.
I am on a quest to find the best local model, so I setup VS Code with Cline and connected it to various local models that I ran on localhost.
Each run follows the same process: open Cline in VSCode in plan mode, paste the prompt below, answer any clarifying questions it asks (typically I prefer 2-player mode and a regular-sized game), then let it write the game without any prompting for bug fixes. One file, one shot.
can you write a plan for how you'd implement a simple pong game in @/src/pages/[name].tsx? make sure CSS and styling is in line, and all changes are in that one fileThe starter component is an empty shell — just a named export with an <h1>Test</h1> inside. Everything else is the model's work.
Results so far — three models tested, two playable games and one failure. I'll keep adding as I run more models.
GPT OSS 20B (openai/gpt-oss-20b) was a complete failure — Cline's tool calling was entirely incompatible with the model and it made zero progress. Nothing was produced.