Thoughts on technology, research, and whatever else is on my mind.
March 29, 2026
A deep dive into benchmarking 67 language models across 23 providers on Hindi hate speech detection, covering accuracy ceilings, false positives, script bias, and the real cost of production moderation.