Former Anthropic Executives Launch Mirendil with $200M Funding to Pursue Recursive Self‑Improvement AI

Origin

In 2019 Harsh Mehta, then a Google researcher, emailed Behnam Neyshabur, a newly‑joined Google researcher known for work on why AI models work. Their shared interest in using AI to accelerate scientific research led to a seven‑year collaboration. Both later joined Anthropic, and after the release of Claude Opus 4.5 in December 2025—an upgrade that markedly improved AI agents’ ability to handle complex tasks—they left to found Mirandil.

Funding

Mirandil raised a $200 million seed round from Andreessen Horowitz, Kleiner Perkins, and Nvidia, valuing the company at $1 billion.

Mission

The startup’s stated goal is to let AI help scientists build their own AI, rather than merely providing generic models. The envisioned platform would enable research teams in domains such as medicine and materials science to train and iterate bespoke models—for example, a model that predicts Alzheimer’s disease risk.

The technical approach relies on recursive self‑improvement (RSI), where an AI continuously refines its own code and capabilities. Neyshabur describes RSI as the shortest path to “AI‑accelerated science” and believes it can be pursued safely under human supervision.

Competitive Landscape

Major AI firms, including Anthropic, increasingly use their own models to speed internal research. Anthropic disclosed that over 80 % of its codebase was generated by Claude as of May 2024. However, their user agreements prohibit external developers from training competing products on these models.

Matt Bornstein of a16z characterizes this “use‑in‑house, not‑share‑out” stance as a rational economic response, creating a structural market gap that an independent company like Mirandil can fill.

Team

The core team includes former Anthropic staff, early‑stage xAI member Shayan Salehian, and MIT graduate Tara Rezaei, totaling about 20 technical personnel based in downtown San Francisco.

Outlook

Mirandil plans to release its first model and product in the coming months to collect early user feedback. The founders envision thousands of laboratories each developing their own AI to tackle the era’s most important problems, with Mirandil providing the enabling platform.