According to many true believers, the greatest promise of AI is its potential to accelerate scientific discovery. Thanks to algorithms, once-in-a-generation breakthroughs could one day become commonplace. It is thought that by extracting patterns from mountains of data too large for the human mind to comprehend, AI scientists could ultimately help solve some of humanity’s most dire technological problems: climate change, cancer, and even (according to hardcore transhumanists) death itself.
But science is a crowdsourcing enterprise, relying on a global community of researchers who can freely access and develop each other’s research. In contrast, the AI industry is currently dominated by a small number of research institutions, whose proprietary code is insulated from each other and the wider world.
A fast-growing startup called Mirendil now hopes to bridge the gap between scientific discovery and access to frontier AI.
The company recently raised $200 in seed funding, giving it a total valuation of $1 billion. Funding was provided by VC firms Andreessen Horowitz and Kleiner Perkins, as well as Nvidia. Based in downtown San Francisco, it currently has a technical staff of approximately 20 researchers. The company’s website lists several job openings with starting salaries up to $500,000.
The startup’s name means “Friend of the Precious” in Elvish, and it joins a growing list of inspired technology company names. lord of the rings— set out to build what is a long-sought, and sometimes cause for anxiety, technological goal within Silicon Valley. It’s an AI that can build better versions of itself, a process known in the tech world as recursive self-improvement.
All AI and machine learning algorithms inherently have some degree of self-improvement, as they are trained to learn from their mistakes over time and adjust their output accordingly. But some of the newest and most advanced models are taking the process to a new level by largely replacing human software engineers and revising much of the code that is executed. It points to a possible future where each new version of the model builds its own successor. Depending on who you ask, the feedback loop could lead us to either a post-scarcity utopia or a hellscape ruled by misaligned superintelligent AI overlords.
Even the current frontrunners in the AI race, Anthropic and OpenAI, have publicly called for the creation of a global oversight board to monitor recursively self-improving AIs and, if necessary, force unilateral slowdowns to prevent humans from losing control. Two of Mirendil’s co-founders, Behnam Neyshabur and Harsh Mehta, previously worked at Anthropic. They left the company in January.
Meanwhile, Microsoft is turning its recursive self-improvement tendencies into an enterprise AI sales pitch. in ×post The company’s CEO, Satya Nadella, wrote earlier this month that an “agent system that improves over time” could soon become a key asset for companies. “I think of this as a mountain-climbing machine,” he wrote.
Anthropic’s Fable 5 opened earlier this month but was quickly shut down following orders from the Trump administration, with security guardrails in place that prevent it from responding to questions about potentially dangerous topics like cybersecurity and chemistry. However, its restrictions were so strict that it often refused to get involved in harmless scientific research problems.
Mirendil believes the problem is not recursively self-improving AI per se, but rather the fact that access to such frontier capabilities is currently gated to a small number of well-heeled research institutions, such as Neishaboor and Mehta’s former employers. So the company has set out to build a self-improving AI system specifically for open source developers.
“Today, laboratories seeking to utilize AI in drug discovery, chemistry, biology, and robotics need to become cutting-edge AI laboratories,” Mirendil writes on his website. “Our goal is to democratize cutting-edge AI research and development and make it widely accessible. Our work will accelerate all science and technology efforts that rely on AI. ”
In other words, the idea is to put recursively self-improving AI at the frontier level into the hands of as many independent research institutions as possible, with the ultimate goal of accelerating scientific progress. ““The most direct path to maturity and significant impact for the AI industry is to let engineers and researchers outside the lab do the actual AI work – to explore the frontiers of their fields of expertise,” Andreessen Horowitz said in Mirendil’s investment announcement. “Call it Vibe Research.”
