Mirendil just raised $200 million to do something humans have never done: let self-improving AI improve itself to solve scientific problems.
On August 6, TechCrunch reported the Anthropic spinout announced a $100 million Google Cloud partnership. But the money isn't the story. The vision of self-improving AI is.
Benham Neyshabur, Mirendil's CEO, put it plainly: "You can have a self-improving AI where you point a problem at it and it keeps getting better with time."
That's the promise. Not AI that follows instructions. Self-improving AI that learns, iterates, and compounds its knowledge to solve disease, discover new materials, or accelerate drug discovery.
This is recursive self-improvement in action. And it might change how science actually works.
The Problem It Solves
Right now, scientific breakthroughs require human scientists. You need years of training, deep domain knowledge, the ability to spot patterns others missed. A brilliant researcher can spend a decade on one problem. And maybe—maybe—crack it.
What if recursive self-improvement could compress that timeline?
Neyshabur describes the vision: "How can we have an AI system that keeps doing research, keeps improving its own knowledge and performance when it comes to Alzheimer's disease? This technology allows us to set goals that are ambitious for AI, and the AI would keep making progress."
The mechanism: Give the self-improving AI a scientific goal. Let it break the problem into sub-problems. Research each one. Build knowledge. Refine its approach. Iterate. Improve. Keep going.
Humans stay in the loop for validation and direction. But the compounding work—the research accumulation, the pattern recognition, the hypothesis generation—that's automated.
For fields like medicine, materials science, and biology, that's exponential acceleration.
Why Now?
Self-improving AI needs compute. Lots of it. Mirendil's $100 million deal with Google gives them access to both TPUs and Nvidia GPUs, plus managed training clusters. That's the infrastructure to train iterative self-improving systems constantly.
The second reason: the talent. Mirendil's founders—including Harsh Mehta and Benham Neyshabur—come from Anthropic, where recursive self-improvement research is a core focus. They're not speculating about whether this is possible. They're building it.
The third reason: competition. Recursive Superintelligence signed a $400 million deal with Amazon. Ricursive Intelligence just launched. Anthropic itself is heavily investing in this direction. The arms race for self-improving AI is starting now.
What This Means for Science
If Mirendil succeeds, the impact is profound:
Compressed timelines. Drug discovery that takes a decade becomes 2-3 years. Materials science research that would require 50 PhDs happens with one scientist + one self-improving AI system.
Accessibility. Today, only massive pharma companies and top research institutions can afford frontier research. Self-improving AI could democratize it. A smaller biotech startup could compete with Pfizer on research velocity.
Novel discoveries. Humans have cognitive blindspots. We miss patterns that AI would see. A self-improving system might discover connections in biology or chemistry that human researchers never considered.
Compounding knowledge. The AI learns from each experiment, each failure, each success. Knowledge compounds. Later research builds on earlier discoveries automatically. Human researchers spend months reading literature; the AI absorbs it instantly.
The Bet Mirendil Is Making
They're betting that self-improving AI can do the work of an entire frontier AI lab.
Neyshabur said: "The startup hopes its AI will eventually be able to take on the work of an entire frontier AI lab."
That's bold. It means self-improving AI doesn't just assist researchers. It becomes the researcher.
The bet: By investing $100 million in compute infrastructure, they can train self-improving systems that solve real scientific problems faster than human teams. And then they can productize that—sell it to biotech companies, academic labs, enterprises who need scientific breakthroughs.
The Realism Check
Self-improving AI sounds like science fiction. And it is—partly.
The reality: Current self-improving AI systems are narrow. They work within specific domains, not general research.
Mirendil's bet is that narrowness is fine. You don't need general AI. You need specific self-improving AI that improves itself on well-defined scientific problems.
That's achievable. The question isn't if it works. It's how fast and at what cost.
Why This Matters for Founders
If Mirendil succeeds, it changes the game for startups in biotech and materials science.
Today: capital-intensive, talent-intensive. You need PhDs, domain expertise, expensive labs.
Tomorrow: You need a good problem, domain expertise to steer the AI, and access to compute.
That's cheaper. More scalable. More entrepreneurial.
For founders in biotech, Mirendil signals: Self-improving AI becomes your co-founder. The race is on.
The Next Breakthrough
Mirendil isn't trying to replace scientists. They're trying to give scientists a tool that learns and improves as fast as they think.
"You can have a self-improving AI that keeps making progress with time." That's the vision.
Weco's recent work on recursive self-improvement demonstrates that self-improving AI isn't theoretical anymore—it's demonstrable today.
If it works, science gets faster. Breakthroughs come sooner. Problems that looked impossible become solvable.
And Mirendil won't be the only one building toward it.