Thinking Machines Lab Tackles AI Model Consistency

Thinking Machines Lab, led by former OpenAI CTO Mira Murati, has launched a new research blog called Connectionism, aimed at addressing the issue of AI model consistency. The lab's first blog post, titled "Defeating Nondeterminism in LLM Inference," explores the root causes of randomness in AI model responses and proposes solutions for creating more deterministic systems.

The blog post, authored by researcher Horace He, identifies the orchestration of GPU kernels as a key factor in AI model nondeterminism. By controlling this layer, Thinking Machines Lab aims to make AI models more reliable for enterprises and researchers. This approach could also enhance reinforcement learning by providing more consistent data for training.

Thinking Machines Lab plans to continue sharing its research openly through frequent blog posts and code releases, fostering collaboration with the research community. The lab's efforts are part of a broader initiative to develop AI tools that are customizable and reliable, with a focus on reproducibility in AI model outputs.