Pretraining at scale without a pretraining team (mirendil.com)

Mirendil has announced the completion of its first pretrained model, Mirendil-Base, created by a single researcher in just 3.5 months with less than $10 million in compute. This project challenges the conventional view that building competitive AI models requires large teams of specialists. By leveraging AI agents to execute over 99% of the coding, experimental runs, and diagnosing of failures, the researcher was able to focus on strategic decisions. The model was built entirely from scratch using standard libraries, trained solely on publicly available data, and achieved notable efficiency improvements through its innovative architecture and tuning processes. This development is significant for the AI/ML community, signaling a potential shift in the landscape of model pretraining, which has traditionally been limited to large organizations. As Mirendil-Base demonstrates that capable models can be developed with reduced manpower, it opens up opportunities for domain-specific teams—such as in robotics or healthcare—to create tailored AI solutions. The shift in reliance from skilled personnel to computational resources also suggests a new paradigm where the potential for individual researchers and smaller teams to drive AI innovation is greatly enhanced, ultimately accelerating advancements across various fields in AI.