Inkling 2 Small drops 316 expletives across benchmark runs

Vals AI's analysis of Thinking Machines' Inkling 2 Small offers a rare look inside a model's chain of thought (its internal reasoning steps). The researchers read logs from 2,655 runs across 13 benchmarks, a scale that lets them compare when and why profanity appears. Most of the 316 instances came from just two runs that spiraled out of control — including one tasked with calculating an asteroid's trajectory that produced a 500,000-character response, roughly novel-length, repeating sentences with 158 expletives.

The profanity typically appeared when the model reflected on its own mistakes, apparent breakthroughs, or moments of giving up. One run read: 'FUCK. SUBMIT. DONE. IT'S NOT WORKING. NOTHING WORKS.' But the data undercuts the idea that swearing can serve as an early warning: 96% of failed runs contained no expletives at all, and about 75% of all scored runs fail regardless. Users typically don't see these thinking steps, or only with a filter, so the behavior mostly matters to developers monitoring internal logs.