Kathmandu, March 26 -- Generative AI models such as ChatGPT and Sonnet are trained on massive amounts of data that have over time enabled them to answer general questions. But they do not perform as well on narrower domains, even hallucinating or making things up.

Now here comes Anthromind, a startup developed by a 28-year-old Nepali engineer Pratik Karki in the tech-womb at Silicon Valley, which develops datasets to improve the performance of AI models in specific areas.

Large language models are 'distilled' by running them on smaller, higher-quality data. AI is first used to assist with labelling or generating data. After that, human experts look at the responses and explain how they are wrong, helping tune the model just right.

Kark...