Remember watching Arrival? A linguist tasked with communicating with aliens learns their language lets her perceive time non-linearly, reshaping her understanding of reality. The plot brought the Sapir-Whorf’s hypothesis back into mainstream about how language actively shapes the way we think and determines what we can think about. In the context of LLMs and Chain of Thought (CoT) reasoning, the hypothesis becomes particularly relevant since the language of thought quite literally determines the quality of computational output. Does it hold for languages AI can work with?
Consider a practical scenario: a farmer in rural Pakistan who needs AI-powered crop analysis but cannot read English. The current generation of multilingual AI models have the ability of handling questions in Urdu but aren’t quite there when it comes to reasoning in the same language. This can create confusion as the farmer must first filter out the text they can’t understand which may result in losing nuance and context. How easy would it be to create an LLM which uses Urdu for reasoning?
During a recent weekend experiment, I explored a potential solution: rather than retraining a model to reason in another language, just ask it nicely to do so. I call it Mufakkir (thinker) – but it’s a bit of a pseudo-intellectual in reality. You can prompt an LLM to produce its chain of thought in Urdu between special characters, which means we can parse and display it to the users – basically a jugaar, but it kinda works at the surface level. You can give it a go at mufakkir.chaoticity.com (while DeepSeek remains up and my credit lasts) but please be gentle and feel free to plugin your streaming compatible endpoint and credentials (OpenRouter has some free tokens!).
This is like instant coffee – quick and simple and gives you the caffeine boost but is it real coffee? A better option would be like a drip brew with a filter and a slow pour. We can train small LLMs to reason natively in a language other than English given we have a clean training dataset and some consumer level compute – probably less than a month’s task assuming one needs to curate the data. But what if we want steamy, velvety cappuccino brewed at 9 bars delivering bold arabica, silky microfoam, and a perfectly balanced, creamy indulgence in every sip? In other words, an advanced, unquantized DeepSeek-level with all it’s 670b+ parameter goodness capable of processing and understanding complex patterns, nuances, and context in the target language – not just translating or recognizing words, but actually reasoning, making connections, and providing responses in a way that feels natural and fluent, just like a human speaker? That would demand significantly more data, computational resources, and cultural as well as linguistic expertise. Who could undertake such an endeavour?
For other languages, the answers increasingly point to national-level initiatives – or Sovereign AI as it is called now. There has been an interesting evolution in the semantics of this term which has shifted from being a fear-sparking self-determining uncontrollable entity (think Skynet from Terminator) to something far more real and practical: a nation’s capacity to develop domestic AI infrastructure driven by local data and talent, serving national-level use cases. This new interpretation emphasizes the importance of developing sovereign LLMs trained on local languages and grounded in indigenous cultural elements and practices – and reasoning models aside, this is where the future of Urdu AI lies.
The path forward involves creating AI infrastructure and systems that enable creation of models and technology that don’t just translate but think in Urdu – “understanding” its cultural contexts, poetic metaphors and unique ways of expressing concepts as well as the national love for cricket and biryani! This represents both a challenge and an opportunity. Such models can bridge language gaps, making digital spaces more inclusive for those who solely rely on Urdu as their primary language. The applications can range from AI-enabled educators offering personalised lessons, to government services providing vital information, forms, and assistance, ensuring that access to information about healthcare, legal aid, welfare programs, etc is not limited in any way because of language barriers.
Which reminds me of what Wittgenstein once said: “The limits of my language mean the limits of my world.” I think we have the chance to push those limits – and while I make the drip brew, I want my cappuccino soon!