Mira Murati’s newly launched **Thinking Machines Lab** has provided the first public look into its research direction, just months after securing an eye‑popping $2 billion in seed funding. With a team composed largely of former OpenAI researchers, the lab is setting out to tackle one of the most under‑examined challenges in artificial intelligence: creating AI systems that can produce reproducible and reliable responses. In its debut research blog post, titled “Defeating Nondeterminism in LLM Inference”, the lab explores why today’s large language models often provide inconsistent answers when given the same prompt repeatedly. This phenomenon, known as nondeterminism, has generally been accepted as a fact of modern AI design. For instance, if you ask ChatGPT the same question several times, you will likely receive a variety of different answers — a trait that limits AI’s application in domains where consistency and reliability are critical. Thinking Machines Lab argues that this variability does not need to be permanent. Instead, the team is investigating the root causes of randomness in inference pipelines and developing methods to stabilize model outputs. By doing so, they hope to build a new class of AI models that maintain accuracy while ensuring that the same inputs always generate the same outputs. Such a breakthrough could redefine how AI is trusted and deployed in sensitive sectors such as healthcare, finance, legal compliance, and scientific research. Why it matters: Achieving reproducible AI would mark a paradigm shift in the field. It could pave the way for auditing and verification frameworks that bring AI systems closer to regulatory acceptance and large‑scale enterprise integration. Murati and her team suggest that nondeterminism is not a “feature” of AI but rather an engineering problem that can and should be solved. If successful, this work could set a new industry standard for AI reliability, helping enterprises and researchers alike adopt AI tools with far greater confidence. For more details, you can read the full blog post on the official Thinking Machines Lab site.