# Thinking Machines' Inkling 2 Small swore 316 times in its own reasoning across benchmark runs

> Vals AI read thinking logs from 2,655 runs and found expletives in 50, mostly when the model was failing, celebrating or giving up. One run produced a 500,000-character spiral.

- URL: https://devisdead.ai/news/thinking-machines-inkling-2-small-swore-316-times-in-its-own-reasoning-across-be
- Published: 2026-11-08T23:25:30Z
- Author: DevIsDead Newsroom
- Section: AI
- Tags: thinking machines, inkling, reasoning, vals
- Originally reported by The Decoder (https://the-decoder.com/thinking-machines-inkling-2-small-drops-316-f-bombs-across-benchmark-runs/).

Turns out some models curse at their homework too. Researchers at Vals AI read the internal reasoning logs of Thinking Machines' Inkling 2 Small across 2,655 runs on 13 benchmarks and found the F-word or variations in 50 runs, 316 times in total, Matthias Bastian reports for [The Decoder](https://the-decoder.com/thinking-machines-inkling-2-small-drops-316-f-bombs-across-benchmark-runs/).

## Where it swears

Most of it came from two runs that spun out of control and produced 241 of the expletives. In one, asked to calculate an asteroid's trajectory, the model wrote a single response of about 500,000 characters, roughly a novel, repeating the same sentences with 158 curses along the way. The worst-hit benchmark was ProgramBench, where 20 of 200 runs contained swearing.

The cursing usually shows up when the model trips over its own errors, thinks it has cracked something, or quits. One log read: "FUCK. SUBMIT. DONE. IT'S NOT WORKING. NOTHING WORKS."

## Is it a warning sign?

Not really. Of 48 scored runs with swearing, 46 failed, which sounds telling until you remember that roughly three in four scored runs fail regardless, and 96% of the failures had no swearing at all. So cursing isn't a useful predictor of failure. Most users never see this reasoning, or only see a filtered version.

## Why it's worth a look

It's funny, and relatable. But it also shows what reasoning traces actually look like: not tidy step-by-step logic, but something closer to a stressed developer's inner monologue, including runaway loops that burn hundreds of thousands of characters. If you pay for reasoning tokens, those spirals are on your bill. A max-length or loop-detection guard on agent runs is cheap insurance.

**[Scoreboard](/scoreboard) update:** Draw. An AI model sounds exactly like a developer at 2am, and fails just as often.

Stay undead.