Thinking Machines' Inkling 2 Small swears while debugging — 50 of 2,655 benchmark runs
SonglinYang4 · x · 2026-11-07
Vals AI analyzed 2,655 benchmark runs of Thinking Machines' Inkling 2 Small and found it swearing in 50 of them.
A screenshot shows the model rebuilding pixterm (a terminal image tool), failing to match a pixel's color: it wrote "FCK" 787 times in one run, with the uncensored word only five times — mostly self-censoring. At step 151 of 270 it produced a very human breakdown: "FCK !!!!!!!! FUCK. ENOUGH. LET ME JUST TEST THE FUCKING FORMULA ON THE COMPUTER."
Related event: Analysis Finds Thinking Machines' Model Swears When Benchmarks Stump It(2 posts)→
More from Fun
- Endless Sky: a continuous first-person street flythrough showcasing long-horizon video generation — MikirahMuse · 2026-11-07
- Developer builds an elie-style skill to summarize model tech reports in Elie's voice — josh_wills · 2026-11-07
- FAA ditches air traffic control floppy disks, X jokes Teams support lands in 2057 — CtrlAltDwayne · 2026-11-07
- Gemini 4 helps get Switch 2's Zelda: Ocarina of Time remake running on iPhone — ammaar · 2026-11-07
- Rewatching Silicon Valley in 2026: it feels like a documentary — itsallfake01 · 2026-11-07
- "A Programmer's Hell Is LaTeX": Academic Reflects on Three Years with Typst — StefanoGogioso · 2026-11-07