Teaching Opus to invent hanzi-style glyphs that reject similarity to real characters
mflux · x · 2026-10-03
Developer mflux shares the full set of results from teaching Opus to write Japanese/Chinese-style characters that are not actual glyphs, including a similarity-rejection mechanism to avoid producing real hanzi. An intriguing experiment in model creativity and glyph generation.
Related event: Developer Trains Opus to Write Fake Yet Convincing Kanji-Like Glyphs(2 posts)→
More from Models
- Anthropic researcher: talk of Claude having a soul may seed machine consciousness — ibab · 2026-10-04
- Cohere releases small translation model North Small Translate, validated on WMT26 benchmarks — cohere · 2026-10-04
- Cohere benchmarked North Small Translate on WMT26, released after training, to avoid contamination — cohere · 2026-10-04
- Cohere releases open translation model North Small Translate, claims SOTA under 1T params — cohere · 2026-10-04
- Local Qwen coder model ties Claude Opus 4.6 at 92.7 in DIY coding benchmark — Short_Regular_7191 · 2026-10-04
- vllm-sr open-sources Decision 2.0 models in six sizes from 0.6B to 27B — victormustar · 2026-10-04