AI lab Mirendil has entered a multi-year partnership with Google Cloud to secure compute capacity for its research into self-improving (recursive self-improvement) artificial intelligence, the company told TechCrunch. Mirendil’s co-founder and CEO, Behnam Neyshabur, said the deal is worth upwards of $100 million.
According to the company, that amount is roughly half of the seed funding Mirendil raised at a $1 billion valuation in late June.
What resources does the agreement provide?
Under the deal, the startup will have access to Google’s TPUs (Tensor Processing Units) and Nvidia GPUs, as well as managed training clusters which Mirendil will use to develop its self-improving AI. The startup hopes the system will eventually be able to take on the workload of an entire frontier AI lab.
What is recursive self-improvement?
Recursive self-improvement describes AI systems that iteratively improve themselves. Major labs are researching this area — Mirendil’s founders previously worked at Anthropic — and several startups focused on the approach, such as Recursive Superintelligence and Ricursive Intelligence, have appeared recently.
Mirendil believes recursive processes can automate large parts of scientific and AI research, accelerating progress in domains like medicine, biology, and materials science.
Behnam Neyshabur said he expects AI to mimic how human scientists learn new domains, accumulate expertise and improve performance over time. “You can have a self-improving AI where you can point a problem at it and it keeps getting better with time,” he said. He added that such systems could be tasked with ambitious goals — for example, to continuously research Alzheimer’s disease and improve their knowledge and performance on that topic.
Matching workloads to hardware
Training self-improving AI requires very large amounts of computing power. Mirendil co-founder Harsh Mehta emphasized that training increasingly involves matching the right workloads to the right hardware. “These models are really good at working with different workloads and chips, and assigning the right workloads to the right chips,” Mehta said.
That flexibility — having access to multiple kinds of accelerators — is a central element of Google’s infrastructure pitch. Amin Vahdat, senior vice president and chief technologist of AI and infrastructure at Google, said that AI advancement is no longer only about chip-level performance, “but how we orchestrate entire systems of intelligence and break through the physical constraints of scaling.”
Neyshabur stated that Mirendil’s software and systems layer will help customers extract more value from Google’s hardware, potentially giving Google another advantage against competitors. In exchange, Google gains a strategic partner building frontier recursive self-improving AI — technology that could later be brought to enterprise customers.
