Mirendil, an AI research lab, has entered into a long-term partnership with Google Cloud, as first reported by TechCrunch. This collaboration aims to secure extensive computing resources necessary for the lab's innovative self-improving AI research.
This agreement highlights two significant trends dominating the AI landscape: major cloud providers are actively pursuing startup collaborations through substantial infrastructure commitments, and AI enterprises are aggressively securing computing agreements to support their growth.
According to Behnam Neyshabur, co-founder and CEO of Mirendil, the partnership's value exceeds $100 million—approximately half of the funding Mirendil secured during its seed round at a $1 billion valuation in late June.
Through this partnership, Mirendil will have access to Google’s Tensor Processing Units (TPUs) and Nvidia GPUs, along with managed training clusters designed to enhance its self-improving AI projects. The startup aspires for its AI to eventually perform the comprehensive functions typical of a leading AI research laboratory.
Self-improving AI, often referred to as recursive self-improvement, involves AI systems that continuously refine their capabilities. This area of research has been a focus at prominent labs, including Anthropic, which counts several of Mirendil’s founders among its team. Additionally, new startups like Recursive Superintelligence and Ricursive Intelligence are emerging with a similar focus.
Mirendil envisions that this innovative process could automate significant portions of scientific and AI research, driving advancements in various domains such as medicine, biology, and materials science.
Neyshabur believes that AI can emulate the learning processes of human scientists, gradually expanding their knowledge and enhancing their efficiency. He stated, “You can implement self-improving AI where any problem presented to it will see gradual improvement over time.” He exemplified this by questioning how an AI could continuously conduct research and enhance its understanding regarding complex issues like Alzheimer’s disease. “This technology empowers us to pursue ambitious goals for AI, ensuring consistent progress,” he elaborated.
Nonetheless, training self-improving AI necessitates vast computing power. Harsh Mehta, another co-founder, explained that the key to effective training lies in efficiently pairing specific workloads with suitable hardware.
“These models excel in optimizing various workloads across different chips, ensuring each task is executed on the most appropriate platform,” Mehta noted. He emphasized that Google’s variety of hardware offers crucial flexibility, enabling Mirendil to effectively align workloads with the right accelerators, ultimately reducing costs for both the lab and its customers.
This adaptability is central to Google’s approach to AI infrastructure. Amin Vahdat, Senior Vice President and Chief Technologist of AI and Infrastructure at Google, remarked that the progress of AI extends beyond pure chip performance; it also involves orchestrating comprehensive systems of intelligence that transcend conventional scaling limitations.
Neyshabur added that Mirendil's software solutions optimize the performance customers can derive from Google’s hardware, providing the tech giant with a competitive advantage. In exchange, Google benefits from having a strategic partner that is pioneering cutting-edge recursive self-improving AI technology, which could potentially be offered to enterprise clients in the future.
