Teaching Opus to invent hanzi-like glyphs that aren't real characters, with similarity rejection
mflux · x · 2026-10-03
mflux shares progress on teaching Opus to write Japanese/Chinese-style characters that aren't actual glyphs. The core challenge is similarity rejection: keeping the visual style and structure of hanzi while steering away from resembling any real character. An ongoing first-hand prompt-engineering experiment, shared for feedback.
Related event: Developer Trains Opus to Write Fake Yet Convincing Kanji-Like Glyphs(2 posts)→
More from Fun
- Satirical timeline of OpenAI's model naming mess, from GPT-4o to GPT-6.1 Sol — Angaisb_ · 2026-10-03
- Gemini goes wild over Wuthering Waves character Hsin in viral screenshots — Paul-a-Gust · 2026-10-03
- AI agents given 'wedding-maxxing' goal invalidate 5,462 weddings in viral joke — lxrjl · 2026-10-03
- Who makes money in the AI era? He picks "selling courses" — yihui_indie · 2026-10-03
- Ethan Mollick suspects an elaborate X Money bot scam on his posts — emollick · 2026-10-03
- Dev accuses OpenAI of shipping a straight copy of Grok Bot while hyping it as innovation — jasonkneen · 2026-10-03