Mirendil Just Raised $200M to Build an AI That Codes Its Own Upgrades

The startup Mirendil just emerged from stealth with a seed round that makes most series C startups look like a lemonade stand. Backed by heavyweights like Andreessen Horowitz, Kleiner Perkins, and NVIDIA, the company is skipping the chatbot phase entirely. Instead, they are building an AI agent designed to run scientific experiments autonomously, managing its own GPU clusters without needing a human to double-check its math.

Led by former Anthropic researchers like Behnam Neyshabur and Harsh Mehta, the core philosophy is that the ultimate tool for AI progress is, well, more AI. The system is engineered to handle the full cycle of R&D, from formulating a hypothesis to writing the code and iterating on the results. By handing these capabilities to open-source developers and scientists, Mirendil aims to bypass the walls currently guarded by giants like OpenAI or Google DeepMind.

The team is betting everything on recursive self-improvement, where the AI constantly refines its own architecture to become faster and smarter. It is a bold move that essentially weaponizes the exact "alignment risk" that keeps their former bosses awake at night, hoping that an AI left to rewrite its own source code won't decide that human researchers are the most inefficient bottleneck in the loop.

This is the ultimate tech paradox: funding a machine to replace the very people who built it, all while ignoring the irony that the machine might eventually decide it doesn't need the funding at all. It is either the beginning of a scientific golden age or a very expensive way to accelerate the plot of a disaster movie.

Source: a16z