Agent curses 787 times (mostly censored) while failing to rebuild pixterm
keirp1 · x · 2026-11-07
Vals AI shared a gloriously funny agent run: asked to rebuild pixterm, a tool that draws images in a terminal, the model couldn't get a pixel's color to match the original and grew visibly frustrated.
Over the run it mostly censored itself — writing "FCK" 787 times while the uncensored word appeared only five times. By step 151 of 270, it was spamming "FCK !!!!!! FUCK. ENOUGH. LET ME JUST TEST THE FUCKING FORMULA ON THE COMPUTER." A great natural experiment in model frustration and self-censorship behavior.
More from Fun
- Community ports 'Opus 5.5' video avatars into open-source virtual world project — repligate · 2026-11-07
- Meme: crying over the AI girlfriend you pushed away — AIandDesign · 2026-11-07
- A Review With 66 Bullet Points: AI Peer Review's Escalating Extremes — shangbinfeng · 2026-11-07
- Apple store ad screen looks cracked — until you realize the 'cracks' spell PRO — msg · 2026-11-07
- User hands Grok a $90 parking ticket dispute — agent drafts and auto-submits the waiver request — doooyle · 2026-11-07
- SemiAnalysis valued above Bumble, sparking outrage over newsletter valuations — edshrag · 2026-11-07