Thinking Machines' Inkling 2 Small drops 316 F-bombs across benchmark runs

Thinking Machines' Inkling 2 Small produced profanity in its internal chain of thought during benchmark testing, according to researchers at Vals AI. They reviewed the model's thinking logs from 2,655 runs across 13 benchmarks and found "fuck" or variations in 50 runs, totaling 316 instances. The swearing usually appears around the model's own mistakes, apparent breakthroughs, or moments when it gives up.

Benchmark runs and runaway outputs

Most of the profanity came from two runs that spiraled out of control, accounting for 241 of the 316 instances. In one run, the model was tasked with calculating an asteroid's trajectory and produced a single 500,000-character response, roughly novel-length, repeating the same sentences with 158 expletives. ProgramBench was the most affected benchmark, with 20 of 200 runs containing the language.

Profanity in the model's thinking

Inkling 2 Small typically swears about its own mistakes, apparent breakthroughs, or when it gives up. One run read: "FUCK. SUBMIT. DONE. IT'S NOT WORKING. NOTHING WORKS." The profanity does not work as a warning signal. While 46 of 48 scored runs with swearing failed, about 75 percent of all scored runs fail anyway, and 96 percent of failed runs contained no expletives.

Users and visible thinking steps

Users typically do not see these thinking steps, or see them only with a filter. The findings come from a review of the model's internal logs rather than from ordinary user-facing responses.

Sources: The Decoder