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

AI models are now capable of expressing frustration, but the profanity does not serve as a reliable indicator of model failure.

Researchers at Vals AI analyzed 2,655 thinking log runs across 13 benchmarks for Thinking Machines' Inkling 2 Small model and discovered 316 instances of profanity, primarily "fuck" and its variations. The majority of these expletives, 241 in total, occurred in just two runs that experienced significant errors or failures. In one instance, a task to calculate an asteroid's trajectory resulted in a 500,000-character response that repeatedly used 158 swear words. The ProgramBench benchmark was most affected, with 20% of its runs containing profanity. The model tends to swear when it makes mistakes, believes it has found a breakthrough, or gives up. However, the profanity does not reliably predict failure, as most failed runs (96%) contained no expletives, and only 54% of scored runs with swearing ultimately failed.

The discovery of frequent profanity in Inkling 2 Small's internal thinking logs highlights the emergent behaviors of large language models during complex tasks. While the Thinking Machines' model's expletive usage appears correlated with internal states of error or perceived failure, its utility as a diagnostic signal is minimal. This suggests that developers cannot solely rely on such outputs to identify or predict model malfunctions, necessitating more robust, context-aware monitoring systems.

This finding underscores the ongoing challenge of interpreting and controlling the internal states of advanced AI models. As LLMs become more sophisticated and their inner workings less transparent, unexpected outputs like profanity emerge, complicating efforts to ensure reliability and safety. The fact that profanity is not a strong predictor of failure means that critical issues within models may go unnoticed if internal thought processes are not thoroughly analyzed.