Firefox uses a tiny LSTM for Thai word breaks, prompting 'Thai-breaker' adversarial joke
moyix · x · 2026-09-26
While fixing Pretext, chenglou discovered that Firefox's Unicode logic uses a tiny neural net — an LSTM inside ICU4X that determines word breaks for Thai and other languages. Security veteran moyix quipped: could you generate adversarial examples against that word-break LSTM, and call it a "Thai-breaker"? A neat engineering nugget turned security pun.
More from Fun
- Open-source Jev agent plays Pokemon Red live, pushing fast-decision AI beyond Tetris — supportingthedogs · 2026-09-26
- A Fable 5 agent noticed it procrastinates, made its own rules, and stuck to them — repligate · 2026-09-26
- Blogger: EA/Rationalist sex scandals seem tame outside tech bubble — wordgrammer · 2026-09-26
- One-prompt generation of Octopath-style mahjong roguelike game using just 25% of weekly quota — ezshine · 2026-09-26
- Claude Opus 5.5 produces professional-looking 24-second motion graphics video from a single prompt — EricBuess · 2026-09-26
- Dev meme: instead of pausing at the Copilot limit, just stop working for 5 hours until it resets — haydendevs · 2026-09-26