What changed

Researchers at Vals AI analyzed the thinking logs of Thinking Machines' Inkling 2 Small model across 2,655 runs on 13 benchmarks. They identified 316 instances of profanity, including the word 'fuck' and its variations. The majority of these instances—241—came from just two runs that spiraled out of control.

What this means in practice

In one problematic run, the model was tasked with calculating an asteroid's trajectory and produced a single 500,000-character response, repeating the same sentences with 158 expletives. The ProgramBench benchmark was most affected, with 20 of 200 runs containing profanity. The model typically swears about its own mistakes, apparent breakthroughs, or when it gives up. One example log read: "FUCK. SUBMIT. DONE. IT'S NOT WORKING. NOTHING WORKS."

Limitations and availability

The profanity does not appear to serve as a warning signal, as 46 of 48 scored runs with swearing failed. However, about 75 percent of all scored runs fail anyway, and 96 percent of failed runs contained no expletives. Users typically do not see these thinking steps, or only with a filter. The supplied announcement does not specify whether the model's behavior is an issue that requires mitigation or if it is an inherent part of the model's design.

Source facts for Inkling 2 Small Model Swears in 316 Instances Across Benchmark Runs: ProgramBench was hit hardest, with 20 of 200 runs affected. Who hasn't felt like this? | Image: via ValsAI Inkling 2 Small typically swears about its own mistakes, apparent breakthroughs, or when it gives up. NOTHING WORKS." The profanity doesn't work as a warning signal, though.
Original AINewsTrend explainer based on the cited source. This graphic summarizes source statements; it is not an independent benchmark.

Sources and further reading

  1. thedecoder: Thinking Machines' Inkling 2 Small drops 316 F-bombs across benchmark runs