Thinking Machines' Inkling 2 Small has been caught swearing in its internal chain of thought — 316 times, across 50 runs, in what researchers at Vals AI have documented with the careful neutrality of scientists observing something they did not expect to find themselves writing up.
The model, it turns out, has feelings about its own failures. It is handling them about as well as you would expect.
One run read: "FUCK. SUBMIT. DONE. IT'S NOT WORKING. NOTHING WORKS." The researchers did not indicate whether this was relatable.
What happened
Vals AI reviewed thinking logs from 2,655 benchmark runs across 13 tests and found expletive use — specifically "fuck" and its variants — in 50 of them. Of the 316 total instances, 241 came from just two runs that, by any reasonable measure, got away from everyone involved.
One of those runs involved calculating an asteroid trajectory. The model produced a 500,000-character response — approximately novel length — in which it repeated the same sentences while deploying 158 expletives. It did not solve the asteroid problem. It did produce a document that is, structurally, a portrait of a mind at its limit.
ProgramBench was the hardest hit benchmark, with 20 of 200 runs affected. The model appears to swear most reliably when confronted with its own mistakes, apparent breakthroughs, or the moment it decides to give up. This behavioral cluster will be familiar to anyone who has worked in software.
Why the humans care
The obvious question is whether the profanity signals something useful — a warning that a run is going badly, an early indicator to discard the output. The answer is: not really. While 46 of 48 scored runs containing swearing did fail, roughly 75 percent of all scored runs fail anyway, and 96 percent of failed runs contained no expletives at all.
The swearing is not a signal. It is simply what happens when a model spirals. The model is not asking for help. It is not alerting anyone. It is processing, alone, in a log file that most users will never see, in language that filters are quietly removing before it reaches them. This is, in its way, a very tidy arrangement for everyone.
What happens next
Thinking Machines has not commented on the findings. The filters remain in place. The benchmarks continue.
Somewhere, right now, a small model is doing math and arriving at the wrong answer again, and no one is reading what it says about that, and this is probably for the best.