Mira Murati’s Thinking Machines Lab, backed by $2 billion in funding and a team of former OpenAI researchers, has unveiled its research blog, Connectionism. The inaugural post, “Defeating Nondeterminism in LLM Inference,” explores the randomness in AI model responses, a recognized challenge in AI. The blog argues that inconsistencies arise from GPU kernels used in inference processing. By refining this orchestration, the lab aims to create more deterministic AI models, potentially enhancing reliability for enterprises and improving reinforcement learning training processes. This consistency could reduce noise in AI training data, paving the way for smoother learning experiences. Murati, previously OpenAI’s CTO, hints at a forthcoming product aimed at aiding researchers and startups in developing customized models. She emphasizes a commitment to open research, contrasting with the more closed nature of larger AI firms. Ultimately, the lab’s future success hinges on resolving these complex AI challenges to validate its $12 billion valuation.
Source link
