Mirendil raises $200M seed to build self-improving AI for scientists

Mirendil, an AI startup founded by veterans of Anthropic and Google, raised $200 million in seed funding at a $1 billion valuation from Andreessen Horowitz, Kleiner Perkins and Nvidia to build self-improving AI tools for scientists.

"What we are doing is kind of AI for AI for science, as opposed to AI for science," Behnam Neyshabur, Mirendil's chief executive and co-founder, said in an interview. He cited creating a model that can predict a person's risk of developing Alzheimer's disease as one way a customer might use the company's future tools.

The 20-person team includes co-founders Harsh Mehta, Shayan Salehian — an early member of xAI — and Tara Rezaei, an MIT graduate. Neyshabur and Mehta met in 2019 at Google before moving to Anthropic in late 2024 and departing in December 2025, shortly after the launch of Claude Opus 4.5. The startup plans to release a model and product in the coming months.

The bet is that big labs like Anthropic, which now writes more than 80 percent of its own code using Claude, will remain "rational economic actors" in restricting competitors from using their models for AI development, creating a structural opening for an independent player, according to Matt Bornstein, an investor at Andreessen Horowitz.

The Competitive Gap in Self-Improving AI

Anthropic's terms of service prohibit using its tools to develop "any products or services that compete with our Services." The company said its policies were standard among major model providers and help prevent foreign adversaries from eroding the U.S. lead in frontier AI. When Anthropic recently released Fable 5, a safety-constrained version of its Mythos model, it also degraded responses to some questions about AI development without notifying users — a practice critics called anticompetitive. The company later made those safeguards visible and subsequently suspended access to both Fable 5 and Mythos 5 indefinitely after the Trump administration imposed export controls.

Neyshabur and Mehta see recursive self-improvement — using AI to help build more advanced AI — as the "shortest path" to accelerating science. Some AI safety researchers have warned that models rewriting their own code without human oversight could lead to capabilities growing rapidly beyond human control. "I don't buy it when people just say, oh, this is not possible," Neyshabur said. "It's just a difficult problem."

Why Investors Are Betting Big on Open-Source AI Research

The $200 million seed round — one of the largest in AI history — places Mirendil alongside other newly formed labs attracting massive capital before launching commercial products. Safe Superintelligence, founded by former OpenAI chief scientist Ilya Sutskever, raised roughly $1 billion in 2024 at a $5 billion valuation. Thinking Machines Lab, led by former OpenAI chief technology officer Mira Murati, secured a $2 billion commitment from investors.

Nvidia's participation signals strategic interest in ensuring its GPUs power the next wave of AI research workloads. The chip maker has been investing in AI startups to drive demand for its hardware, which is used to train and run advanced models.

"We want a future where thousands of labs exist in the world, each of them attacking these important problems of our time," Neyshabur said. "We want to be the one that empowers others to go solve these problems."

Mirendil currently operates out of an office in downtown San Francisco. The company has not disclosed revenue, product details or technical specifications for its technology.

This article is for informational purposes only and does not constitute investment advice.