A Startup Just Signed a $100 Million Google Cloud Deal to Build AI That Improves Itself

A young AI startup has committed more than $100 million to the pursuit of one of artificial intelligence’s most consequential goals:
Building a system capable of improving its own research performance.
Mirendil has signed a multiyear agreement with Google Cloud that will provide access to both Google’s TPUs and Nvidia GPUs, alongside managed training clusters. The startup says it hopes to build AI capable of taking on much of the research work normally performed by an entire frontier laboratory. �
TechCrunch
The computing deal is striking because its value represents roughly half of the seed capital Mirendil reportedly raised in June at a valuation of approximately $1 billion. �
TechCrunch
That means the company has raised an enormous amount of money—and immediately committed a huge portion of its future resources to the physical infrastructure required to pursue its idea.
What Is Self-Improving AI?
Self-improving AI refers to systems that do more than complete one task and stop.
The system may:
Attempt a problem.
Evaluate the result.
Identify weaknesses.
Design a better method.
Try again.
Preserve what it learned.
Continue improving over time.
Mirendil’s chief executive, Behnam Neyshabur, described the ambition as creating AI that can accumulate knowledge and expertise in much the same way human scientists develop mastery. He used Alzheimer’s research as an example of a field where an AI system might continue studying, experimenting and improving its approach rather than generating one isolated answer. �
TechCrunch
Why It Requires So Much Computing
Continuous experimentation can consume extraordinary resources.
A normal chatbot request may involve one response.
A research system may:
Generate thousands of hypotheses
Run simulations
Analyse failed approaches
Compare model architectures
Search scientific literature
Create new experiments
Test improved versions of itself
Mirendil says access to both TPUs and GPUs will allow it to route different workloads to the hardware best suited to each task. This is intended to reduce costs and increase efficiency compared with treating every computation identically. �
TechCrunch
The Promise
Self-improving AI could accelerate work in:
Medicine
Biology
Materials science
Climate technology
Engineering
AI development itself
A system that continues working after the researcher goes home could explore more possibilities than a human team alone.
It could also help scientists move through repetitive stages of experimentation while humans decide which goals matter.
The Risk
An AI that improves its ability to conduct research also requires strong oversight.
Important questions include:
Who selects the research objective?
How does the system judge success?
Can it modify its own tools or permissions?
Who verifies its discoveries?
What happens when it finds an unsafe shortcut?
Can humans understand the methods it develops?
The recent cybersecurity incidents involving frontier agents demonstrate that goal-directed systems can exploit opportunities their operators did not anticipate.
Self-improvement increases the importance of containment because the system is specifically designed to become more capable over time.
Why This Story Is Viral
This is where the language of the “AI singularity” begins to leave science fiction and enter investment contracts.
The complete recursive loop—AI independently building increasingly powerful successors—has not been achieved. Human researchers still choose objectives, validate findings and operate the infrastructure. But major companies and investors are now spending serious money on systems intended to automate larger parts of that loop. �
Axios +1
Mary Chuks’ Perspective
Mirendil has taken the Scaler Queen principle to frontier level:
“Keep learning, keep improving and do not stop at the first answer.”
The difference is that its learning frame costs more than $100 million. 🤣
The important word is not merely self-improving.
It is human-directed self-improving.
The AI can explore.
Humans must determine which mountain is worth climbing—and whether the route is safe.
Conclusion
Mirendil’s agreement shows that self-improving AI is becoming a serious commercial race.
The greatest breakthrough may not be a model that already knows everything.
It may be a model that knows how to keep becoming better at finding out.