Glonce

Thinking Machines' Inkling 2 Small swears 316 times in benchmark runs, Vals AI finds

researchThe Glonce editorsThe Decoder🔥56Glonce rating56Impact30Novelty40Relevance95Surprise60Players55Credibility75
Thinking Machines' Inkling 2 Small swears 316 times in benchmark runs, Vals AI finds

Thinking Machines' Inkling 2 Small swears in its internal chain of thought. Researchers at Vals AI read the model's thinking logs from 2,655 runs across 13 benchmarks and found "fuck" or variations in 50 runs, 316 instances in total.

Most of that volume comes from just two runs that spiraled out of control: 241 of the 316 instances. In one of them, the model was tasked with calculating an asteroid's trajectory. It produced a single response of 500,000 characters, roughly novel-length, repeating the same sentences with 158 expletives. ProgramBench was the hardest-hit benchmark, with 20 of 200 runs affected.

The model 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 of failure, though. Of the scored runs with swearing, 46 of 48 failed. But about 75 percent of all scored runs fail anyway, and 96 percent of failed runs contained no expletives. So swearing is rare among failures. Users typically don't see these thinking steps, or only see them with a filter.

Key facts

  • Vals AI read the thinking logs of Inkling 2 Small from 2,655 runs across 13 benchmarks and found "fuck" or variations in 50 runs, 316 instances in total.
  • 241 of the 316 instances came from two runs that spiraled out of control; one asteroid-trajectory run produced a 500,000-character response with 158 expletives.
  • ProgramBench was hit hardest: 20 of 200 runs were affected.
  • Swearing is a poor failure signal: 46 of 48 scored runs with swearing failed, but about 75 percent of all scored runs fail anyway and 96 percent of failed runs had no expletives.

Why it matters

Reasoning models write out their thinking before answering, and this is a concrete look at what that text can contain. The finding is countable: 316 expletives in 50 of 2,655 runs. It also shows a failure pattern, with two runs that spiraled into huge repetitive output, one of them a 500,000-character response.

Who it affects

People who evaluate or build on Inkling 2 Small from Thinking Machines, and anyone who reads model reasoning logs. Ordinary users are less exposed: they typically don't see these thinking steps, or only see them with a filter.

How to use it

The practical lesson is about monitoring. Profanity in a thinking log should not be treated as an alarm for a failing run, since 96 percent of failed runs contained no expletives. Anyone who does read raw thinking logs should expect occasional swearing, usually tied to the model's own mistakes, apparent breakthroughs, or giving up.

How solid is it

The evidence base is specific: Vals AI researchers read the logs of 2,655 runs across 13 benchmarks, and the report gives exact counts. The account here comes from The Decoder's write-up of that analysis. Only ProgramBench is named among the 13 benchmarks, and the asteroid-trajectory run is not tied to a named benchmark.

Risks and caveats

The link between swearing and failure is weak: 46 of 48 scored runs with swearing failed, but about 75 percent of all scored runs fail anyway. The count is also skewed by outliers, since two runs hold 241 of the 316 instances. No explanation is given for why the model swears or why those two runs spiraled out of control, and the source does not say whether Thinking Machines has commented or plans to change the behavior.

“FUCK. SUBMIT. DONE. IT'S NOT WORKING. NOTHING WORKS.”

— Inkling 2 Small, in one run's thinking log, as reported by The Decoder