Mirendil raises $200 million to build self-improving AI that accelerates scientific discovery

Mirendil raises $200M to build AI that improves AI

Mirendil, founded by two researchers who left Anthropic after barely a year, has raised $200m at a $1bn valuation. The pitch: sell the self-improving AI that the big labs build for themselves and guard from everyone else. The biggest AI labs share one private conviction. The fastest way to build better AI is to point AI at the problem of building AI. They run that loop in-house, and their terms of service stop outsiders from doing the same. Two of their former researchers have now raised $200m to break the lock. The startup is called Mirendil, and it announced its seed round on 24 June. The figure is striking for a company with no product: $200m at a $1bn valuation, one of the largest seed rounds the sector has seen. Andreessen Horowitz and Kleiner Perkins co-led it, with Nvidia joining in. Its founders are Behnam Neyshabur and Harsh Mehta. The pair met at Google in 2019, moved to Anthropic in late 2024, then left in December 2025, soon after the launch of Claude Opus 4.5. Neyshabur, now chief executive, had spent more than five years at Alphabet co-leading reasoning research for Gemini. Selling the engine the labs keep for themselves Plenty of lab alumni have started their own shops. Mirendil is aiming at a different layer. It wants to build AI that does the work of an AI researcher: designing experiments, searching for the right settings, evaluating models, and running the next round of training. The idea is to package that as a platform other organisations can point at their own problems. The more important shift is who that platform is for. Neyshabur frames it as "AI for AI for science," not AI for science. A university biology lab could use it to build a drug-target model without a machine-learning team. He cites a model that predicts a person's risk of Alzheimer's as the kind of thing a customer might make. The work that takes labs months, the pitch goes, compresses into days. The detail that gives the thesis its edge is the lock it tries to pick. As of May, Anthropic said its Claude model wrote more than 80% of the company's own code. Yet its terms of service forbid using its tools to build competing services. Anthropic told the Wall Street Journal the policy is standard among model providers and helps keep frontier AI out of foreign adversaries' hands. That gap is the business. As Andreessen Horowitz's Matt Bornstein put it to the Journal, the labs are being "rational economic actors" when they deny customers the means to supercharge their own models. "Structurally, there has to be an independent company," he said. Mirendil wants to be it. A $200m bet on recursive self-improvement The technical name for this loop is recursive self-improvement, and it is contested ground. Anthropic has pointed to it as a potential danger, on the theory that a model rewriting its own code without oversight could slip beyond human control. The founders disagree. They call it the "shortest path" to faster science, and a problem that can be supervised rather than avoided. That argument lands in a tense moment for Anthropic itself. The company recently pulled access to its most powerful Mythos and Fable models after the Trump administration imposed export controls. The same week, critics accused it of quietly degrading answers about AI development. Into that backdrop steps a startup whose entire reason to exist is handing that capability to others. The team is small and senior. Mirendil runs on about 20 researchers and engineers drawn from Anthropic, xAI, Google DeepMind, and OpenAI. The founding group also includes Shayan Salehian, an early member of xAI, and Tara Rezaei, a 23-year-old MIT graduate. There is a neat irony in the line-up. Mehta built the first version of Anthropic's internal AI-research platform, at times as a team of one. Now he is rebuilding that idea to sell it. The Information first reported some details of the round. The money is chasing a structural bet The valuation makes sense only against the flood of capital around it. AI took close to half of all global venture funding in 2025, some $202bn, up more than 75% on the year, according to Crunchbase. The AI infrastructure market alone ended 2025 near $337bn in revenue and is forecast to reach $1.2tn by 2030. Mirendil also sits inside a specific cluster of lab spinouts, and the comparison flatters it. Ilya Sutskever's Safe Superintelligence has raised $6bn at a $32bn valuation. Mira Murati's Thinking Machines Lab took $2bn at $12bn. Both, like Mirendil, launched with no shipped product, on the strength of their founders alone. The closer rhyme is Periodic Labs, another Andreessen Horowitz bet that raised $200m to aim AI at materials science. Mirendil is pitching the layer beneath that: the research-automation engine such companies would themselves run on. It is a harder thesis to prove, and a bigger prize if it holds. Venture firms have poured money into the field for two years, and this is their next structural wager. The democratisation pitch There is an ideological pull, too. The founders talk about prising AI research out of a few labs and handing it to thousands, the same democratising argument that has followed every open challenge to Silicon Valley's frontier. Whether that vision survives contact with a real product is another matter, and it lands amid a steady run of nine-figure AI infrastructure rounds. For now, Mirendil has a name from Tolkien, a billion-dollar valuation, and a model and product it says will arrive in the coming months. If AI can truly automate its own research, the advantages today's labs hold, thousands of staff and years of accumulated knowledge, start to look less permanent. Whether Mirendil is on time or five years early is exactly what the $200m is there to find out.

Don't Be Afraid of Self-Improving AI, Says a16z-Backed Startup Mirendil

According to many true-believers, the biggest promise of AI is its potential to accelerate scientific discovery; once-in-a-generation breakthroughs could one day become routine, thanks to algorithms. By extracting patterns from troves of data far too vast for any human mind to fathom, so the thinking goes, AI scientists could eventually help solve some of humanity's most dire technical problems: climate change, cancer, even -- according to diehard transhumanists -- death itself. But science is a crowdsourced enterprise, dependent upon a global community of researchers who can freely access and build off each other's work. The AI industry, in contrast, is currently dominated by a handful of research labs whose proprietary code is closed off from one another, and from the wider world. A fast-rising startup called Mirendil is now hoping to bridge that gap between scientific discovery and frontier AI access. The company recently raised $200 in seed funding, bringing its total valuation to $1 billion. Funding was provided by VC firms Andreessen Horowitz and Kleiner Perkins, as well as by Nvidia. Based in downtown San Francisco, it currently has a technical staff of around twenty researchers. Its website features several job postings with starting salaries of up to $500,000. The startup -- whose name means "friend of precious things" in Elvish, adding to a growing list of tech companies names inspired by The Lord of the Rings -- has set out to build something that's long been a sought-after technical goal, and occasionally a source of anxiety, within Silicon Valley: AI that can build increasingly more capable versions of itself, a process known within tech circles as recursive self-improvement. All AI and machine learning algorithms fundamentally have some capacity for self-improvement, since they're trained to learn from their mistakes over time and adjust their outputs accordingly. But some of the latest and most advanced models have taken that process to a new level by largely replacing human software engineers and revising much of the code it runs on. It points to a possible future in which each new version of a model builds its own successor, a feedback loop that could either usher us into a post-scarcity utopia or a hellscape dominated by misaligned superintelligent AI overlords, depending on whom you ask. Even Anthropic and OpenAI, the two current frontrunners of the AI race, have publicly called for the formation of a global oversight committee to keep tabs on recursively self-improving AI, and if it should ever become necessary, to (somehow) enforce a unilateral slowdown to prevent humans from losing control. Two of Mirendil's cofounders, Behnam Neyshabur and Harsh Mehta, previously worked at Anthropic; they left the company in January. Microsoft, meanwhile, is trying to turn the trend towards recursive self-improvement into a sales pitch for enterprise AI. In an X post earlier this month, company CEO Satya Nadella wrote that "agentic systems that improve over time" could soon become an important asset for businesses. "I think of it as a hill climbing machine," he wrote. Anthropic's Fable 5, which was publicly released earlier this month, only to be swiftly shut down in response to an order from the Trump administration, comes with security guardrails that prevent it from responding to queries on potentially dangerous topics like cybersecurity and chemistry. Its restrictions were so stringent, however, that it would often refuse to engage with harmless scientific research questions. Mirendil believes the problem isn't recursively self-improving AI per se but rather the fact that access to such frontier capabilities is currently gated by a small number of deep-pocketed labs, like Neyshabur's and Mehta's former employer. The company is therefore setting out to build self-improving AI systems specifically for open source developers. "Today, any lab trying to use AI in drug discovery, chemistry, biology, or robotics must also become a frontier AI lab," Mirendil writes on its website. "Our goal is to democratize frontier AI R&D and make it widely accessible. Our work will accelerate every scientific and technological effort that depends on AI." The idea, in other words, is to put frontier-level recursively self-improving AI into the hands of as many independent laboratories as possible, with the end goal of supercharging scientific progress. "The most direct path to maturity and massive impact for the AI industry is to let engineers and researchers outside the labs to do real AI work, i.e. to push the frontier in their own domains of expertise," Andreessen Horowitz wrote in its Mirendil investment announcement. "Call it vibe research."