Question
Tell me something interesting about AI and LLM
Answer
Here's something fascinating: **LLMs can "think" in languages they weren't trained on.**
Most people assume an LLM like GPT-4 or Claude processes everything in English (or whatever language you're using). But research shows that these models often perform *better* on complex reasoning tasks—like math, logic, or multi-step planning—when they're prompted in a language they were *less* exposed to during training, such as **Frisian** or **Zulu**.
Why? Because the model's "internal reasoning" isn't tied to any one human language. When you ask a question in English, the model might get distracted by the vast web of English idioms, cultural references, and common phrases it learned. But in a low-resource language, it has fewer noisy associations to sift through. It essentially defaults to a cleaner, more abstract "thought process" that's closer to its core mathematical and logical architecture.
This is a bit like a polyglot who finds it easier to solve a puzzle in a language they don't speak fluently—because they're not distracted by the emotional or cultural baggage of their native tongue.
**Bonus twist:** Some researchers have even found that prompting an LLM in a *made-up* or *nonsense* language (like "klingon-style gibberish") can sometimes yield better results than English, as long as the syntax is consistent. The model doesn't need to *understand* the words—it just needs a clean, structured pattern to route its reasoning through.