Exclusive: Mirendil inks $100M+ Google Cloud deal to scale self-improving AI

Mirendil, an AI lab, has entered a multiyear partnership with Google Cloud to access compute capacity for its self-improving AI research, according to exclusive information from TechCrunch. The deal is worth over $100 million, which is half of Mirendil's seed funding that was raised in June at a $1 billion valuation. The startup will have access to Google's Tensor Processing Units (TPUs) and Nvidia GPUs, along with managed training clusters.

This will enable Mirendil to work on its self-improving AI, which aims to eventually match the work of a full AI research lab. Self-improving AI, also known as recursive self-improvement, involves AI systems iteratively enhancing themselves. Major labs like Anthropic, where Mirendil's co-founders previously worked, are exploring this concept.

The startup believes that self-improving AI can automate various scientific and AI research, aiding progress in fields like medicine, biology, and materials science. Mirendil's co-founder and CEO, Behnam Neyshabur, envisions an AI system that continuously improves its own knowledge and performance, tackling complex problems like Alzheimer's disease.

Training self-improving AI demands significant computational resources, and the lab is focused on matching workloads to the appropriate hardware for optimal results. Google's flexibility in providing various chips allows Mirendil to optimize its workloads, reducing costs for both parties. Mirendil's software and systems layer will help customers maximize the benefits of Google's hardware, giving the cloud giant a strategic advantage in the competitive AI landscape.

In exchange, Google gains access to a cutting-edge AI team that can potentially offer its technology to enterprise clients.