Mirendil: Building the system that builds systems

AI has promised, for years now, to solve our hardest problems. We think the most direct path to that promise runs through AI itself — specifically, AI that can accelerate its own research. And for the first time, there are real signs of models doing meaningful AI research. Not assisting. Doing it.

The lab that automates that loop will do far more than ship better models. It will build something capable of continually improving itself toward any goal you point it at. Mirendil’s founders are building exactly that, and they’ve been preparing for it their whole careers.

This isn’t their first time at the frontier. Behnam was the co-inventer of the SAM optimizer and did foundational work on why deep learning generalizes. He was fascinated from the start by how models learned. Behnam and Harsh met for the first time in Google through a cold email 7 years ago and the pair have been building together ever since. Behnam was a co-lead of Blueshift where he co-led the effort to build Mirerva, one of the first models that could reason its way through hard math, and then together, they went on to help drive pretraining and reasoning on Math and Code in Gemini. The pair then went on to join Anthropic where Behnam co-led the Discovery team with the goal of building an AI Scientist and the computer-use work behind Claude, while Harsh initiated and led the effort for automated AI R&D itself.

They’ve brought along two builders who match them. Shayan Salehian was an early engineer  at xAI, working across post-training, reasoning, agent infrastructure, and the Grok models. Whatever was most critical, he owned. Prior to that he owned large workstreams at X all the way to the original Twitter. Tara Rezaei is a 23-year old MIT graduate, an early student researcher at OpenAI, an Olympiad medalist, and one of the most ambitious young people in Silicon Valley.

Here’s what they’ve decided to build. Most AI labs are research organizations that happen to use AI. Mirendil is the inverse — a frontier lab rebuilt from scratch around AI. They train frontier models that are exceptional at AI R&D, then redesign the entire loop around them to make it faster, more capable, and more autonomous. Better models do better research. Better research produces better models. The loop is the product.

When this works, it works for everyone. Today, any lab using AI for drug discovery, chemistry, biology, or robotics has to become a frontier AI lab first — expensive, slow, and out of reach for all but a handful of teams. Mirendil’s goal is to put frontier AI R&D in everyone’s hands, so the people closest to a problem can stay on the problem, and science can move at the rate it deserves.

The name means friend of precious things, the hidden and the undiscovered. They call Mirendil the first lab from the future. After the time we’ve spent with this team, I don’t think that’s a stretch.

We’re proud to be backing Behnam, Harsh, Shayan, Tara, and the entire Mirendil team from day one. They’ve already built much of what the field is chasing. Now they’re building the system that builds systems.