Vals finds profanity in 50 of Inkling 2 Small's 2,655 benchmark runs

Vals AI read the thinking logs from 2,655 runs across 13 benchmarks and found profanity in 50 runs, 316 instances in all, according to The Decoder. Two runs that spiraled produced 241 of them, including one 500,000-character response. Of 48 scored runs with swearing, 46 failed, though most scored runs fail anyway.

Why it matters: It is an outside look at how a model behaves inside its reasoning, not just at its final answers.

The Decoder reports about 75 percent of all scored runs fail, and 96 percent of failed runs had no profanity. Weights are still promised for the week of Nov 9.