Woofun AI reports that former Anthropic researchers Behnam Neyshabur and Harsh Mehta have officially launched Mirendil, a new artificial intelligence startup that immediately secured $200 million in seed funding. This initial capital injection valued the company at $1 billion, establishing it as one of the highest-valued seed-stage ventures in the current artificial intelligence sector. The founding team comprises 20 researchers and engineers drawn from elite institutions including Anthropic, XAI, Google DeepMind, and OpenAI, all united by a singular mission to accelerate scientific discovery through autonomous AI evolution. The company's technical strategy centers on "recursive self-improvement," a methodology designed to allow AI systems to iteratively optimize their own code and capabilities without constant human intervention. The name Mirendil, derived from "The Lord of the Rings," translates to "friend of treasures" in Elvish, symbolizing the organization's intent to unlock hidden scientific value.

The trajectory of this venture traces back to 2019, when Mehta was a standard researcher at Google and Neyshabur had recently joined the same organization. At that time, Neyshabur had already garnered significant academic recognition for his deep inquiry into the fundamental question of "why AI models work," leading Mehta to describe him as a "rising star in the field." Their professional relationship began when Mehta sent an email to Neyshabur, sparking a long-term collaboration focused on the potential of AI to accelerate scientific research. Despite their shared enthusiasm, the limitations of existing model capabilities in 2019 rendered their vision purely theoretical. It was not until the end of 2024 that both individuals joined Anthropic, where they worked until December 2025. Shortly after the release of Claude Opus 4.5, they made the strategic decision to depart and establish their own independent entity.

The timing of their departure coincides with the release of Claude Opus 4.5, which significantly enhanced the ability of AI agents to handle complex tasks, convincing the founders that the technological maturity required for their vision had finally arrived. Neyshabur clarified the company's distinct mission, stating, "What we are doing is using AI to help scientists create their own AI systems, rather than simply using AI to assist in scientific research." This approach diverges from the standard model of using general-purpose tools provided by large technology corporations. Instead, Mirendil intends to build a tool platform that enables research teams in specialized fields such as medicine and materials science to independently train and iterate on custom AI models. A concrete example of this application involves helping researchers develop models capable of predicting the risk of Alzheimer's disease. This operational model relies heavily on the controversial concept of recursive self-improvement, which permits AI systems to continuously refine their own architecture. Neyshabur asserts that this represents the shortest path to achieving "AI-driven scientific acceleration" and maintains that it can be safely implemented under human supervision. He explicitly rejected skepticism regarding the feasibility of this approach, noting, "I don't accept claims that it's impossible to do this; it's just a very challenging problem."

The ability of Mirendil to secure $200 million in funding from Andreessen Horowitz, Kleiner Perkins, and NVIDIA is rooted in a clear market logic regarding the current behavior of leading AI firms.

Woofun AI data shows that major companies like Anthropic are increasingly leveraging AI to accelerate their internal research efforts, with Anthropic reporting that as of May this year, more than 80% of the company's internal code was written using Claude. Despite this heavy internal adoption, these corporations explicitly restrict external developers from using their models to train competitive products within their terms of service. Matt Bornstein, an investor at Andreessen Horowitz, characterized this "use it internally, don't share it externally" strategy as a natural response for large companies acting as "rational economic entities." However, this restrictive posture creates a structural gap in the market that independent entities like Mirendil are positioned to fill by providing the necessary infrastructure for external scientific innovation.

Beyond the two co-founders, the core leadership team includes Shayan Salehian, an early member of xAI under Elon Musk, and Tara Rezaei, a graduate of MIT. The company currently employs approximately 20 technical staff members and operates out of offices located in downtown San Francisco. In the coming months, Mirendil plans to release its first model and product to gather initial feedback from users, marking the transition from concept to execution. Neyshabur envisions a future landscape where thousands of laboratories around the world are dedicated to tackling the most critical issues of our time, each empowered by their own custom AI systems. "We want to be the force that enables them to achieve their goals," Neyshabur stated, emphasizing the company's role as an enabler rather than a gatekeeper. This launch represents a significant shift in the AI industry, moving from centralized model development to a decentralized ecosystem of specialized scientific tools. The success of Mirendil will likely depend on its ability to navigate the technical complexities of recursive self-improvement while maintaining the safety protocols necessary for widespread scientific adoption.