Foundation models made in India?
Is Kalaido.ai legit? Or is this another one fine-tuned using OpenAI models or wrapped around them? The demo looks good tbh
You can’t really just fine tune an OpenAI model. The weights are not open, they did have a fine tuning program open but it was invite only afaik.
Training large foundational models is hard for startup’s, mistral did it but they raised an obscene amount of money.
All this is besides the point - I don’t think that an LLM trained from scratch on Indian languages solves any relevant problems.
I don’t think anyone is doing it from scratch. That would take too much money and resources. It is all about tuning now
Costs are estimated to be anywhere between $10M to $100M plus time factors too.
I was initially of the opinion it cannot be done. But this startup turned out to be an arm of a bigger $100M giant. I think they might have the resources to do it.
“It is trained on a public dataset of 70 Mn images.” They would have cut corners for sure
See more comments