AI doesn't just make mistakes, it makes them confidently.
In this episode, we explore the statistical reasons behind why language models hallucinate, why they sometimes invent facts that sound perfectly true.

It turns out the issue starts in training itself: models are rewarded for guessing rather than showing uncertainty. And when evaluation systems penalize “I don't know,” AI learns to bluff to win.

We break down what’s really happening under the hood, and how researchers are rethinking evaluation to make AI more self-aware, and a little more honest.

AI models are great at sounding smart, until they start repeating themselves.
This episode unpacks a clever method called Verbalized Sampling, a way to make AI rediscover its creative spark. Instead of sticking to safe, predictable answers, this approach asks the model to think in possibilities, generating multiple ideas and weighing their confidence in each.

The result? More diverse, surprising, and realistic outputs, no retraining required.

Tune in to see how one simple prompt can help AI sound a little more… human.

Transformers can write stories, generate code, and chat like humans, but ask them to multiply 477 × 591, and they often stumble. Why does this happen?

In this Hindi episode of Data aur AI Pe Charcha, we explore why even powerful Transformer models struggle with basic math. Panelists explain how AI models process numbers, where they lose track of long-range connections, and what new techniques like Implicit Chain-of-Thought are doing to fix it.

It's a fascinating look at how machines "think," why multiplication breaks their logic, and how scientists are helping them finally get the math right.