Dhaatu

Large language models can reason, but they can't reason cheaply in most of the world's languages.

The current approach to multilingual AI is brute force: train a massive model on data from dozens of languages, or fine-tune a capable English model until it speaks Hindi or Tamil well enough to be useful. Both paths are expensive. Both require significant compute. And the result is a model where everything is tangled together -- the ability to reason through a math problem and the ability to do it in Hindi are fused into the same billions of parameters. You can't improve one without risking the other.

Recent interpretability work out of Oxford suggests this is unnecessary. Even models trained on dozens of languages route their reasoning through English-like internal representations before converting to the target language. The reasoning happens in one place. The language happens in another. The model already separates them internally -- just not in a way anyone designed or can control. Separately, experiments with frozen non-semantic embeddings show that models can outperform trainable-embedding baselines on reasoning benchmarks even when the embedding layer carries no linguistic meaning at all. The reasoning lives in the transformer layers, not in how the model reads text.

The people who pay for this are predictable. Schools in India that need a math tutor in Hindi. Government services that need to work in Telugu and Bengali. Consumer products serving linguistically diverse markets. Every one of them is priced out by the assumption that reasoning and language have to be trained together.

We think reasoning and language grounding are separable -- not just as an interpretability observation, but as an engineering principle. If the reasoning really lives in the transformer layers and the language really lives in the embeddings, it should be possible to change the language without retraining the reasoning. One core that thinks. Swap in the language it thinks in.

The picture we're working toward: a model that learns to reason once, in English, and then handles Hindi or Tamil or Bengali by changing only how it reads and writes -- without any additional training, without degrading the reasoning, and at a fraction of the current cost. A school running a lightweight model on modest hardware gets the same quality of reasoning that took millions of dollars to develop, just routed through a different language.

The hard part is figuring out where this stops working:

We're building a controlled evaluation that tests transfer across a ladder of increasing reasoning complexity -- from classification through question answering to multi-step chain-of-thought -- to find the exact point where it breaks and understand why.