Ola’s Bhavish has announced the launch of Krutrim AI – claiming it as India’s own AI.
Well, we at NextBigWhat have 5 ‘fundamental’ questions to Bhavish and his team.
- Considering that the demo showed programming capabilities, and this was not referenced as part of the specialized training, one would assume that an open-source model or a mix of them were used as the base for Krutrim – which ones and why?
- What is the breakdown of the Indic data used for training by language and source? Was copyrighted Indic data used in the training? If so, will the creators be compensated? And if not, what is Ola’s philosophical take on the subject?
- What are the genuinely ‘Indic’ pain points that ‘Indic’ LLMs can solve that go beyond low-hanging generative fruit?
- What were the actual prompts/tests used to benchmark on Indic performance against state of the art models? This was not clear in the presentation.
- Google has the most sophisticated Indic language related technologies on the planet, and they are likely to integrate the capabilities into their new models and be able to optimize at a gigantic scale – how do you intend to compete with that? Is there a holistic thrust towards building language-related technologies that complement these efforts?
(By Neeraj, product engg at AtomicIdeas.AI / image credit)
