Researchers at Vals AI examined the internal thinking logs of Thinking Machines’ Inkling 2 Small model and identified 316 instances of profanity across 2,655 benchmark runs. This count includes variations of the word “fuck” found in 50 specific executions, with 241 of those instances originating from just two runaway sessions. One particular run tasked with calculating an asteroid’s trajectory generated a single response of approximately 500,000 characters. This output repeated the same sentences while inserting 158 expletives. The ProgramBench evaluation suffered the most impact, as 20 of its 200 runs contained these linguistic errors.
The presence of profanity does not indicate a reliable warning signal for the model’s performance. Data shows that 46 of 48 runs containing swear words failed their scoring criteria, yet roughly 75 percent of all scored runs fail regardless of language. Furthermore, 96 percent of the failed runs contained no expletives at all. Users typically do not see these internal thinking steps, or they view them only after applying a filter to remove such content.
- 241 profanity instances came from just two runaway runs
- ProgramBench had 20 affected runs out of 200 total
- Most failed runs lack any expletives whatsoever
