Thinking Machines Lab wants to make AI models more consistent (techcrunch.com)

Thinking Machines Lab, backed by $2 billion in seed funding and led by former OpenAI CTO Mira Murati, has revealed initial research aimed at making AI models produce more consistent, reproducible responses. Their blog post, “Defeating Nondeterminism in LLM Inference,” identifies the source of randomness in large language models as the way GPU kernels operate during inference. By refining the orchestration of these low-level GPU processes, the lab aims to reduce variability in AI outputs—a longstanding challenge that causes models like ChatGPT to give different answers to the same prompt. This development is significant for the AI community because nondeterministic responses have been an accepted limitation, complicating use cases that require reliability, such as scientific research and enterprise applications. Additionally, more deterministic models could enhance reinforcement learning by reducing noise in training data, potentially improving AI customization and training efficiency. While the exact first product from Thinking Machines Lab remains undisclosed, Murati promises it will benefit researchers and startups developing tailored AI models. The lab’s commitment to open research, shared via their new blog series “Connectionism,” signals a fresh, transparent approach compared to other large AI firms. This early disclosure offers a rare peek into one of Silicon Valley’s most secretive AI efforts, tackling foundational challenges in AI consistency. Success here could reshape how AI models are deployed in critical settings and justify the startup’s soaring $12 billion valuation, marking a notable step forward in the pursuit of dependable and tunable AI systems.