Mirendil Raises $200M to Build Self-Improving AI Platform

Mirendil raised $200 million in a seed round at a $1 billion valuation, co-led by Andreessen Horowitz and Kleiner Perkins with participation from NVIDIA, according to a16z's own investment announcement and reporting by the Wall Street Journal and Gizmodo. The San Francisco company was founded by Behnam Neyshabur (who worked on AI for science for more than seven years at Google and Anthropic), Harsh Mehta (who built the first version of Anthropic's internal autoresearch platform), Shayan Salehian (an xAI ML engineering veteran), and Tara Rezaei (a 23-year-old MIT graduate and olympiad medalist), per a16z. Neyshabur and Mehta left Anthropic in late 2025 or early 2026, with accounts differing on the exact month. Mirendil has about 20 technical staff and job postings with starting salaries up to $500,000, according to Gizmodo.

Mirendil's stated goal is to train frontier models that are themselves expert at AI R&D and build a product around that capability, looping over research and engineering problems (proposing experiments, running code, interpreting results, managing compute, comparing checkpoints) with progressively less human intervention, according to a16z. A16z frames this as a systems problem requiring backend infrastructure, agent harness design, evaluation, and both post-training and pre-training strategy to be built together, rather than a single-model advance. Most major AI labs have built similar internal automation platforms, but a16z notes they are not designed for external users and labs have limited incentive to share them with competitors.

Coverage frames Mirendil's positioning as a direct response to the concentration of frontier AI capability inside a small number of labs; a16z's Matt Bornstein argues the biggest gains from AI will come once the technology reaches builders outside those labs, citing Cursor's evolution from third-party models to training its own as a precedent. The launch also arrives alongside public debate over recursively self-improving AI: Anthropic and OpenAI have called for international oversight of the concept, and Gizmodo notes Mirendil's founders argue the risk lies in restricted access to frontier capability, not in self-improvement itself.

If automated research-loop tooling becomes broadly available, teams adopting it will need stronger evaluation frameworks, reproducibility guarantees, and provenance tracking before trusting automated retraining decisions, since the core engineering effort shifts from writing model code to designing evaluation suites and safe rollout criteria. Mirendil has not yet published a technical whitepaper or benchmark results, so its actual capabilities remain unverified outside of investor and company statements.

Track what concrete product Mirendil ships (experiment orchestration, automated evaluation suites, or model-selection APIs), any published evaluation methodology or reproducibility guarantees, pilot programs with universities or research labs, and public documentation addressing safety checks and access policy for automated retraining.