I opened the fundraising note at 2 a.m. and felt a familiar jolt: a small San Francisco startup had just closed a massive seed round. You probably felt the same twitch of hope and unease—the exact mix that makes science news stick to your brain. For a heartbeat the future split into two headlines: a tidal wave of breakthroughs or a runaway system no one controls.
I want to walk you through why that split matters, who’s trying to steer it, and what Mirendil—the a16z-backed startup at the center of the storm—says about the risks and the promise.
In a downtown San Francisco office, twenty researchers are coding the future — who is Mirendil and what did it just sell to investors?
I’ll be blunt: Mirendil looks like a VIP pass for labs that can’t compete with the compute and secret sauce of the big AI houses. The company, whose name borrows Elvish for “friend of precious things,” raised a $200 million seed round (€186 million) and now sits at about a $1 billion valuation (€930 million). Backers include Andreessen Horowitz, Kleiner Perkins and Nvidia, and the team lists salaries up to $500,000 (€465,000) on job postings.
You should notice two facts at once: the founders, Behnam Neyshabur and Harsh Mehta, left Anthropic in January; and Mirendil’s pitch is simple—give independent science labs the same self-improving tools that today are mostly locked inside deep-pocketed AI companies.
What is recursive self-improvement?
At its core, recursive self-improvement is an algorithmic feedback loop: models that not only learn from data but increasingly write and revise their own scaffolding. All modern machine learning tunes itself to better predictions, but this is a version where the model helps redesign the code, automates tests, and proposes architectural changes with less human supervision. Think of it like handing the apprentice the lathe and the blueprint, then watching them design a better lathe.
On X, Satya Nadella called this “a hill climbing machine” — how the industry is framing self-improvement
Microsoft’s CEO turned the concept into an enterprise sales line: agentic systems that improve over time are a business asset. That framing matters because it pushes the technology from a research curiosity into the commercial center of gravity. Anthropic and OpenAI, meanwhile, have publicly argued for global oversight committees to monitor any runaway cycles. Anthropic’s recent model, Fable 5, was pulled after a federal order and was so tightly constrained it refused benign scientific questions while blocking potentially dangerous ones.
Is self-improving AI dangerous?
You’ve heard the two dominant narratives: one paints a post-scarcity utopia, the other a misaligned superintelligence. I’ll tell you what worries me: concentration of capability. When a few labs hold the keys, the incentive structure favors guarded models and business-first guardrails. That’s why Mirendil frames its mission as democratic—make the same engine available to open-source developers and independent labs so progress isn’t wedged behind corporate walls.
In lab corridors and Slack channels, researchers complain about gated access — how Mirendil plans to alter scientific workflows
Mirendil’s public line is that any group trying to use AI for drug discovery, chemistry or robotics currently has to become a frontier AI lab. Their answer: provide self-improving stacks that smaller teams can deploy. Andreessen Horowitz called the idea “vibe research,” meaning the clearest path to maturation is letting domain experts run real experiments with frontier tools.
This is a practical argument. If you’re a biotech team, you want agents that can generate assays, analyze failures, and suggest protocol tweaks without hiring five more PhDs in model engineering. Mirendil’s pitch is that those agents should be accessible.
How will Mirendil affect open science and drug discovery?
From where I sit, the change is not instant magic. Giving labs tools changes timelines: experiments iterate faster, hypotheses get tested sooner, and publication cadence accelerates. But faster is not the same as safer. The governance question shifts from “who builds the models?” to “who polices and audits each lab’s stack?”
There’s another human dynamic at play. When capability spreads, so do different risk appetites and expertise levels. The same system that helps a small team find a cancer lead could also, if misused, speed research into risky dual-use areas. That’s why calls for oversight from Anthropic and OpenAI aren’t just hand-wringing: they’re a plea to craft shared norms before the technology is ubiquitous.
In investor decks and job listings, money signals priorities — what the funding and hires reveal
VC backing and big salaries tell you where attention is flowing. Nvidia’s involvement signals hardware alignment; Andreessen Horowitz and Kleiner Perkins point to a growth play that expects enterprise adoption and broad developer uptake. Mirendil’s message to labs is clear: you won’t need to become a frontier lab to use frontier tools.
There’s a political angle too. Mirendil positions itself as an antidote to concentration, yet concentration can reappear in different guises—if the startup’s stacks become the new standard, who audits Mirendil’s updates? The paradox is that democratizing access can create a new single point of reliance.
I won’t pretend there’s a neat answer. The debate now is about trade-offs: faster discovery and broader access versus shared oversight and careful guardrails. And yes, the trajectory of these systems matters: like a campfire that either warms a village or jumps the ridge, small differences in design can rewrite the map.
I’ve traced the players, the money and the arguments. You can read the fundraising headlines and the policy statements and feel pulled both toward hope and alarm. Which model do you want steering tomorrow’s lab bench—one controlled by a few big players, or one you and your colleagues can tune yourself?