Former Anthropic researcher founded Mirendil, raising $200 million, claiming to be "self-upgrading AI."

AI startup Mirendil announced the completion of a $200 million seed funding round, achieving a valuation of $1 billion and becoming a unicorn. The round was led by Andreessen Horowitz (a16z) and Kleiner Perkins, with Nvidia participating. The company was co-founded by Behnam Neyshabur and Harsh Mehta, both former researchers at Anthropic, who left after the release of Claude Opus 4.5 in December 2025. Mirendil's core goal is to develop a "recursive self-improvement model," allowing AI to autonomously rewrite, train, and upgrade itself, assisting scientists in building specialized AI models in fields such as medicine and materials science. (Previous summary: Two leading scientists from Google DeepMind left, stock price fell 7%, Demis Hassabis responded: We can still attract talent) (Background supplement: The US-China AI competition is heating up, but scholars from both countries agree: Don't let AI face a "Chernobyl moment")

Table of Contents

Toggle

• Two Anthropic veterans leave to start their own company
• Building an AI that can rewrite itself
• Is self-improvement feasible?

Key Highlights

• Mirendil completed a $200 million seed funding round, valuation reached $1 billion, led by a16z and Kleiner Perkins, with Nvidia participating
• The company was co-founded by former Anthropic researchers Behnam Neyshabur (CEO) and Harsh Mehta, with a team of about 20 people, including former xAI members and MIT graduates
• The core technology is a recursive self-improvement model, with a clear divergence from frontier labs' safety concerns about the technology

Just half a year after leaving Anthropic, Behnam Neyshabur and Harsh Mehta returned to the table with an idea that frontier labs are most wary of: building an AI that can rewrite itself. And this time, Silicon Valley's top venture capitalists chose to back this idea with real money.

a16z has publicly released an investment announcement, formally confirming this $200 million seed round, with a valuation that has reached the $1 billion unicorn threshold. For a startup that hasn't released any products yet, this endorsement carries significant weight.

Two Anthropic veterans leave to start their own company

Mirendil was co-founded by Behnam Neyshabur and Harsh Mehta. Behnam Neyshabur serves as CEO, previously leading the scientific AI reasoning team at Anthropic; Harsh Mehta also comes from Anthropic's research department. The two first met while working at Google in 2019, joined Anthropic together at the end of 2024, stayed for less than a year, and both left after the release of Claude Opus 4.5 in December 2025.

The founding team also includes former xAI early member Shayan Salehian, and MIT graduate Tara Rezaei. Currently, Mirendil has an office in downtown San Francisco with about 20 technical staff.

Building an AI that can rewrite itself

Mirendil's core technology is the "recursive self-improvement model." The main idea is that AI can autonomously rewrite, train, and upgrade itself, without requiring human engineers to manually intervene each time.

The company's positioning is not to compete head-on with OpenAI or Anthropic, but to create an "AI that can accelerate AI research," allowing scientific labs and enterprises to develop and control their own specialized models without relying on a few frontier labs. Target application scenarios include medical research and materials science, helping scientists build domain-specific models.

Is self-improvement feasible?

There is a very subtle line here. Data disclosed by Anthropic shows that as of May 2026, Claude has written over 80% of the company's code. Frontier labs themselves have long been using AI to accelerate model development. However, on the other hand, the terms of service of frontier labs explicitly prohibit external developers from using large models to train competitive products, and recently they have quietly imposed reply restrictions on questions related to AI development without actively informing users.

Regarding safety concerns, Mirendil's position is quite direct:

Recursive self-improvement is the shortest path to accelerating scientific research, and this path can be achieved under safety regulation. This is the core proposition put forward by the Mirendil team.

This stance clearly diverges from frontier labs' cautious attitude toward recursive self-improvement. Many AI safety researchers believe that models capable of self-upgrading pose unpredictable risks; Mirendil, however, frames the technology as a tool for accelerating science and emphasizes that safety regulation can proceed in parallel.

Mirendil plans to release models and products in the coming months to gather early user feedback.

Frequently Asked Questions

What is Mirendil? Who founded it?

Mirendil is an AI startup co-founded by former Anthropic researchers Behnam Neyshabur (CEO) and Harsh Mehta, established after leaving Anthropic in December 2025. The company aims to develop AI models that can autonomously rewrite and upgrade themselves, assisting scientific labs in building specialized models. It currently has an office in San Francisco with a team of about 20 people.

What is recursive self-improvement? Why is it controversial?

Recursive self-improvement refers to AI's ability to autonomously rewrite, train, and upgrade itself without human engineers having to manually intervene each time. The controversy lies in frontier AI safety researchers' concerns that the behavior of self-upgrading models is unpredictable; but Mirendil believes it is the shortest path to accelerating scientific research and argues that it can be achieved under safety regulation.