A look at Indian startups like TuluAI, which are building LLMs for low-resource languages by creating data sets nearly from scratch with community involvement
Indian founders are building digital data sets for low-resource languages by involving the community, and count on local relevance to go against big tech platforms. X: @restofworld , @restofworld , @restofworld , and @restofworld . LinkedIn: Rina Chandran X: @restofworld : “Our language is vulnerable and at risk of disappearing. So I took matters into my own hands.” The mission to build better AI tools for India's less common languages https://restofworld.org/... @restofworld : “Most AI systems are built in the U.S. They don't understand Indian languages or contexts. We need our own models that represent us,” said the founder of TuluAI, one of several AI startups hoping to compete with ChatGPT by focusing on local languages https://restofworld.org/... @restofworld : “We don't compete with GPT on scale. We compete on relevance.” The mission to build better AI tools for India's less common languages https://restofworld.org/... @restofworld : India has more than 1,600 languages and dialects. ChatGPT supports around a dozen. That's where these small startups see an opportunity to compete https://restofworld.org/... LinkedIn: Rina Chandran : ChatGPT is huge in India. Some startups are building AI tools for less common languages, painstakingly gathering data …
Context & Ripple Effects
This work addresses a gap long visible in multilingual AI: chatbots' weaker performance outside English and a country whose language diversity complicates foundational-model development, as covered in India's broader AI push.
It also extends a local data-building thread that included Karya's worker-owned AI training-data model. TuluAI and peers are treating community participation and local context as the basis for competing where general-purpose platforms have limited coverage.
First-order effects
- TuluAI and similar Indian startups must create language data largely from the ground up, making community participation a core input to their low-resource-language models.
- Speakers of underrepresented languages gain projects designed around their language and context rather than relying solely on broadly trained platforms with limited language support.
Second-order effects
- Local-model builders can differentiate on relevance and language coverage rather than frontier scale, raising the value of trusted community data collection and curation.
- Big-tech platforms and larger Indian AI providers face a clearer localization challenge: broad language support alone may not meet demand for models grounded in specific linguistic communities.
Third-order effects
- If these efforts produce durable datasets, language coverage could become a more decentralized layer of AI competition, with community relationships and data stewardship acting as defensible assets.
- The approach also makes the governance of community-contributed language data consequential: preservation goals and commercial model development may increasingly need to be reconciled.
The trend: AI competition is splitting between scale-focused general models and locally grounded systems built around language-specific data, distribution, and trust.