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

Researchers examining the internal thinking logs of an AI model found that it used profanity 316 times across benchmark test runs, with the word appearing in 50 of 2,655 total runs. The model typically swore when describing its own mistakes, perceived progress, or when giving up on tasks, such as when a single response about calculating an asteroid's trajectory expanded to 500,000 characters with repeated expletives. While most runs containing profanity did fail their tasks, the vast majority of failed runs contained no expletives, suggesting the swearing is not a reliable warning signal of failure. Users generally do not see these internal thinking steps, or they are filtered out before display.