Agent uses an AI-built typography UI to perfectly solve LLM anagram failures
Future_Candidate2732 · reddit · 2026-09-21
LLMs cheat at character-level tasks like perfect anagrams due to tokenization. The author's recursive tool-use loop points the model at an open-source web typography suite that was itself 100% AI-generated: a raw-text control fails, a fresh chat instance of the same model maps the custom DOM inputs, mathematically isolates remaining characters and solves the anagram perfectly, then autonomously generates a contextual background image, computes X/Y overlay coordinates and exports an MP4. The author sees this as a glimpse of agentic workflows: custom interfaces built by AI to bypass the model's own blind spots. Fully open source.
More from coding & agent
- Dev: coding agents got useful and now I'm working 5x harder — rakyll · 2026-09-21
- Mollick: the most annoying part of long agentic tasks is language drift, not hallucinations — emollick · 2026-09-21
- MiniMax H3 round-up: depth camera control, horror LoRAs, seam fix and Qwen Image 2.1 — optimisticalish · 2026-09-21
- DavidKPiano: the trend of putting agent control flow in Markdown will look silly soon — DavidKPiano · 2026-09-21
- Kev open-source decision models scale to 0.6B/4B/8B, trainable in 40 min on one H100 — TheMoonMidas · 2026-09-21
- Dev argues GPT is too slow for high-APM games; jev-style fast controllers fill the gap — flowersslop · 2026-09-21