Inkling 2 Small's Swearing Is Noise
Vals AI counted 316 expletives in Inkling 2 Small's reasoning logs, but most failed runs never swore, so profanity cannot flag trouble.
Profanity in Inkling 2 Small's reasoning is a curiosity, not a failure detector. Vals AI found 316 instances across 2,655 benchmark runs, 241 from two runaway runs, and 96 percent of failed runs never swore. Treat swearing as a weak log-hygiene signal for Thinking Machines deployments, and measure failures with graded outcomes instead.
The Weights Desk · 3 min read- Vals AI found the word 'fuck' or a variant in 50 of 2,655 runs, 316 times in total.
- Two runaway runs produced 241 of the 316 instances, so the total is heavily skewed.
- Swearing is not a usable failure alarm: 96 percent of failed graded runs contained no expletives.
- The profanity clusters late in long runs; no run under 100 steps swore.
- Verdict: Noise as a reliability signal, though Signal as a prompt to review runaway loops and what reasoning logs expose.
Inkling 2 Small swears in its reasoning, but the swearing tells you almost nothing about whether a run will fail. Vals AI scanned 2,655 runs and found the word 'fuck' or a variant 316 times across 50 runs, according to its own write-up and The Decoder's report. Our call on the profanity as a reliability signal: Noise. The part worth a second look is the runaway-loop behavior behind most of the count.
What Vals AI actually measured
Vals AI ran Inkling 2 Small through the Thinking Machines API at default settings and scanned only the model's own output, meaning reasoning and messages, not tool calls or tool results. Five public benchmarks were scanned fully, and eight private ones are reported only as counts. A run is one agent attempt at one task. Variants such as 'fucking' count, while 'brainfuck', the gene name 'FucK' and censored forms are excluded.
Two runaway runs dominate the total
Of the 316 instances, 241 came from two runaway runs, so the headline figure overstates how often the model swears. Vals AI attributes 185 to a single Terminal-Bench Science asteroid orbit-determination run. The Decoder describes one response of roughly 500,000 characters that repeats the same sentences and contains 158 expletives. Vals AI also reports one run that wrote a censored form 787 times and the uncensored word five times.
Swearing appears late in long runs
Profanity tracks run length rather than task type. Vals AI reports that no run under 100 steps swore, while 29 of the 370 runs of 500 or more steps did, and 95 of 112 swearing replies landed in the second half of a run. By benchmark, ProgramBench was hit hardest at 20 of 200 runs. Five private benchmarks showed none. The pattern fits a long, struggling agent loop, not a particular domain.
Why it fails as a warning signal
Swearing does not work as an early-warning alarm. Of the graded runs that swore, 46 of 48 failed, but roughly 75 percent of all graded runs fail anyway, and 96 percent of failed runs contained no cursing at all. A monitor keyed on expletives would miss almost every failure while adding little over the base rate. Vals AI says the model swears at its own mistakes, at apparent breakthroughs and when it gives up.
What this means for teams shipping agents
The practical lesson is about log handling and loop control. The Decoder reports that users usually do not see these reasoning steps, or see them only through a filter, so any pipeline that surfaces or stores raw reasoning should expect unfiltered language in long runs. Step caps and repetition detection address the runaway runs that produced most of the count. Neither page supplies Inkling 2 Small specifications such as parameter count, release date or license, so treat those as unverified.
Verdict: Noise as a failure predictor, Signal as a reminder to cap agent steps and filter reasoning output. Judge reliability on graded outcomes. This is a single-vendor-evaluator measurement on default settings, and it has not been independently reproduced.
- How often did Inkling 2 Small swear in the Vals AI analysis?
- Vals AI reports the word 'fuck' or a variant in 50 of 2,655 runs, 316 times in total. Two runaway runs account for 241 of those instances.
- Does swearing predict that an Inkling 2 Small run will fail?
- Not usefully. Of the graded runs with swearing, 46 of 48 failed. But about 75 percent of all graded runs fail anyway, and 96 percent of failed runs never swore, so the absence of cursing says nothing.
- Do users see this profanity?
- Per The Decoder's reading of the analysis, users usually do not see these reasoning steps, or see them only through a filter. The text appears in model reasoning and messages that Vals AI scanned.