Harsh Mehta, a researcher who initiated and led Anthropic's automated pre-training research project, has departed to co-found Mirendil, a new AI lab focused on self-accelerating research, with public endorsement from Martin Casado, a general partner at Andreessen Horowitz (a16z). The announcement, made on June 6, 2026, signals that a16z is backing the venture as part of a broader trend of frontier AI researchers spinning out specialized companies.

What Is Mirendil and Why Does It Matter?

Mirendil represents a new wave of AI startups founded by alumni of major AI organizations like Anthropic, OpenAI, and Google DeepMind. Rather than competing directly with these giants on general-purpose models, Mirendil is focusing on a narrow but high-leverage problem: automating the research process that creates foundation models, the large language models (LLMs) that power AI systems.

The company's mission centers on what researchers call "self-accelerating AI R&D." This refers to using AI agents to conduct, evaluate, and iterate on the research processes that produce foundation models. Instead of human researchers manually designing training runs, selecting data mixtures, and tuning hyperparameters, automated systems perform these tasks with minimal human oversight, effectively using AI to build better AI.

"We left Anthropic to start Mirendil because we believe democratizing self-accelerating AI R&D is a bottleneck for accelerating science: Any lab trying to use AI in drug discovery, chemistry, biology, or robotics should have access to this technology in a safe manner," stated Martin Casado.

Martin Casado, General Partner at Andreessen Horowitz

Why Is a16z's Involvement Significant?

Casado's public endorsement carries substantial weight beyond a typical social media mention. As a general partner at a16z, he has championed investments in foundational infrastructure companies and has been an outspoken advocate for the autonomous-agent economy. His framing of Mehta's early vision positions Mirendil as a company built around a thesis that has been validated by Anthropic's continued investment in the same research direction.

The involvement of a16z matters because the firm has been among the most active investors in the AI agent infrastructure space, having led funding rounds in companies across the autonomous-systems stack. This connection ensures Mirendil will have access to capital and distribution networks that reach across the broader AI ecosystem.

How Does This Fit Into the Broader AI Landscape?

Mirendil's formation reflects a significant shift in how AI research is being organized and commercialized. The emergence of these "neo-labs" has important implications for the agent economy's supply chain. If automated pre-training research and development can be productized or offered as infrastructure, it could lower the barrier for building specialized foundation models, expanding the ecosystem of models available to autonomous agents.

The timing is notable because automated pre-training R&D has become mature enough to be spun out and commercialized independently. Multiple frontier labs have reported that autonomous agents are now responsible for meaningful portions of their internal research pipelines, a development that compresses iteration cycles from weeks to hours.

Key Details About the Leadership Transition

Mehta's departure from Anthropic marks a significant leadership change at the frontier AI safety company. According to Casado's announcement, Andrej Karpathy, previously the head of AI at Tesla and a founding member of OpenAI, now leads the automated pre-training R&D project at Anthropic. Karpathy is one of the most recognized figures in deep learning, underscoring the strategic importance Anthropic places on this research direction.

Mehta initiated and led the automated pre-training project at Anthropic more than a year before launching Mirendil. His departure to start a competing venture, with a16z backing, suggests confidence that the technology can be developed faster and more efficiently in a startup environment.

How to Understand Automated Pre-Training R&D's Impact

• Research Acceleration: Automated pre-training R&D compresses the time it takes to develop and improve foundation models by using AI agents to handle tasks that previously required weeks of human researcher effort, reducing iteration cycles to hours.
• Democratization of AI Development: By productizing this technology, Mirendil aims to make advanced AI research capabilities accessible to labs working in drug discovery, chemistry, biology, and robotics, not just well-funded frontier labs.
• Feedback Loop Enhancement: Automated pre-training R&D creates a self-improving cycle where AI agents improve the very models that power them, sitting at the core of what researchers call the "self-accelerating AI" thesis.

What Remains Unknown?

While Casado's endorsement strongly suggests a16z is backing Mirendil, no funding amount or valuation has been publicly confirmed as of the announcement. The use of "we" in Casado's post indicates a16z involvement, but specific financial terms remain undisclosed. The company's website and formal launch details have not yet been made public, though Casado's post encourages interested researchers and engineers to apply or reach out.

This development underscores how the AI research landscape is evolving. Rather than concentrating all cutting-edge research within a handful of large organizations, venture capital is now backing specialized startups founded by researchers who have proven their expertise at frontier labs. For the agent economy, a dedicated company focused on automating model research could accelerate the pace at which new, more capable agent-optimized models reach the market.