Why Low-Resource Languages Need Dedicated AI

There is a comforting assumption in the AI industry: as frontier models grow, every language rises with the tide. Give it a year, the thinking goes, and GPT-class models will speak fluent Uzbek, Yoruba, and Khmer for free. The data says otherwise — and the reason is structural, not temporary.

The long tail is long

Frontier models are trained on web-scale data, and the web is brutally skewed. English dominates; a handful of European and East Asian languages follow; then the curve collapses. Uzbek — spoken by 36 million people — is a rounding error in a frontier training mix. Models learn what they see, and they barely see Uzbek.

The failures this produces are not cosmetic. Uzbek is agglutinative: meaning is assembled by stacking suffixes, and one wrong suffix flips a sentence from 'he did it' to 'he apparently did not do it.' A model that has seen too little Uzbek doesn't make small mistakes — it makes confident, fluent-sounding errors that only a native speaker catches. For casual chat that's annoying. For a bank, a court, or a hospital, it's disqualifying.

Why dedicated beats general

• Data density: a model fine-tuned on 75,000 lines of native-written Uzbek instruction data sees more high-quality Uzbek alignment signal than a frontier model sees in its entire training run.
• Cultural grounding: dedicated data teaches Navro'z, mahalla, Uzbek legal and administrative conventions — things web-crawled English simply doesn't contain.
• Evaluation: we measure Uzbek specifically, with a 300-prompt native benchmark. Frontier labs don't — Uzbek regressions ship silently.
• Economics: frontier labs allocate effort by market size. A dedicated team allocates all of it to the language that matters.

None of this means dedicated models must beat frontier models at everything — they don't have to. They have to be excellent at the language, honest about the culture, and deployable inside the country. That bar is both achievable and, for most institutional use cases, the only bar that matters.

The question isn't whether the giants will eventually speak Uzbek. It's who owns the data that teaches them — and what gets built locally in the meantime.

We believe every major language community will eventually run this playbook: curate the corpus, write the alignment data, build the eval, serve the model. We're running it first for Uzbek — and documenting the path for everyone behind us.