Mira Murati’s Thinking Machines Study: Your LLM Isn’t Creative, It’s Just Broken

Mira Murati, the ex-OpenAI CTO, and her team just proved AI’s “randomness” is a server-side flaw, changing how we must build every LLM.

We’ve all been there. You ask ChatGPT a question. You get a decent answer. You’re curious, so you ask the exact same question a minute later, and… you get a completely different answer.

Classic AI, right? We’ve been told this is normal. It’s part of the “creativity” of the model, a result of a process called “sampling.”

But what about when you want a straight, consistent answer? The tech-savvy among us know the trick: you set the “temperature” to 0. This is supposed to turn off the creativity, making the AI’s choices totally deterministic. It should always pick the most probable next word, no exceptions.

So, you set temperature to 0. You ask your question. You get an answer. You ask it again.

And you get a different answer.

:(

This isn’t just a small glitch. It’s a serious problem that has confused developers and frustrated researchers for years. It makes scientific experiments impossible to repeat and turns debugging into a complete…