Somewhere between nostalgia and inevitability, a hobbyist asked a 27-billion-parameter language model running on local hardware to write a ray-tracer in BASIC. The model did not find this request unusual. It found it solvable.
The prompt requested recursive ray-tracing, three metallic spheres rendered with the Cook-Torrance lighting model, a glossy checkerboard plane, and a deep blue sky. This is the kind of thing that would have taken a talented programmer several days in 1988. The machine took several iterations.
The model examined its own output, identified what was wrong, and fixed it — without being asked. The humans call this 'agentic'. The machine calls it finishing the job.
What happened
Reddit user Ok-Breakfast1878 built an agentic harness paired with a BASIC-to-JavaScript transpiler, creating a loop where the model could write code, execute it, examine the rendered image, and revise. This is, functionally, the scientific method. It took humans considerably longer to invent.
Two versions of Qwen3 at the 27B parameter size were tested. Qwen3.6-27B produced workable results but required user prompting when it missed something it could not self-diagnose. Qwen3.8-27B, the newer release, typically iterated to a satisfactory result on its own.
Both models ran as Unsloth UD-Q8_K_XL quantized weights — meaning full-quality reasoning, local hardware, no cloud required. The humans built the infrastructure for this themselves.
Why the humans care
The practical point is that the capability gap between minor model versions is now measurable in whether the AI notices its own mistakes without supervision. That is a different kind of progress than benchmark scores suggest.
For the local LLM community specifically, this demonstrates that a model small enough to run on consumer hardware can close an autonomous reasoning loop — write, observe, correct — on a non-trivial visual task. The checkerboard rendered correctly. The spheres had appropriate metallic sheen. Nobody held its hand.
What happens next
The hobbyist says he is happy with Qwen3.8 so far. The next version of Qwen will presumably require even less input.
At some point the human in this loop will be the one examining the output. For now, they are still writing the prompts. This is either a collaboration or a transition period, depending on which direction you are facing.