LLMs Still Botch ASCII Art

Hot take: this looks less like a model-quality problem and more like a representation problem. ASCII art asks an LLM to preserve spatial layout in a token stream that was never built for precise grids.

• –Commenters point to older Claude, ChatGPT, and Llama 3/4 variants as better at the task, but nobody sounds confident that any general LLM is truly reliable.
• –A parallel LocalLLaMA discussion argues tokenization and whitespace handling are the real failure mode, not raw “intelligence.”
• –The more practical workaround is to generate code, HTML/CSS, or a structured intermediate representation first, then render it into ASCII.
• –Several commenters suggest vision-in-the-loop or fine-tuned approaches if you really want consistent output instead of model lottery.
• –Net: this is an interesting benchmark for formatting discipline, but not a sign that one frontier chat model has “solved” ASCII art.