Mirendil Inc., a startup developing artificial intelligence models for scientists, has raised $200 million in funding at a $1 billion valuation.

The seed round was led by Andreessen Horowitz. Mirendil stated in its Wednesday funding announcement that Kleiner Perkins, Nvidia Corp. and several other investors contributed as well.

Mirendil is led by Chief Executive Officer Behnam Neyshabur and Chief Technology Officer Harsh Mehta. Neyshabur is a co-inventor of SAM, a popular algorithm for improving the output quality of AI models. Mehta helped launch Anthropic PBC’s efforts to automate parts of its internal research program using custom AI tools.

Building a frontier model involves a significant amount of manual work. Mirendil plans to develop neural networks that can automate much of that work. According to the company, the goal is to create an AI system that can autonomously upgrade itself to increase output quality. The hope is that such a self-improving AI will be capable advancing machine learning faster than manual research initiatives.

Mirendil plans to make its self-improving AI available to scientists in fields such as chemistry, medicine and robotics. The company sees customers using the software to build frontier models optimized for specific research tasks.

Mirendil’s website doesn’t contain any information about its technology. However, one of the company’s job postings states that it plans to develop new variants of existing neural network architectures. The opening indicates that the effort will prioritize the transformer architecture, which is used to build large language models.

One of Mirendil’s priorities will be to develop novel attention mechanisms. An attention mechanism is a module that LLMs use to analyze users’ prompt and identify the most important data points.

Researchers have developed multiple versions of the technology over the past decade. Some popular variants, such as grouped multi-query attention, are designed to reduce LLMs’ considerable RAM usage. Others improve models’ ability to process prompts that contain a large amount of data.

A second job posting indicates that Mirendil will develop its self-improving AI with the help of reinforcement learning sandboxes. Those are simulations in which neural networks hone their skills by interacting with one another. Google LLC has used a similar approach to train its AlphaGo Zero system, one of the past decade’s highest-profile machine learning breakthroughs.

Mirendil will build custom AI tooling to speed up its model development efforts. The company hopes to automate more than a half-dozen tasks including data preparation and debugging. 

“Call it vibe research,” Andreessen Horowitz investors Matt Bornstein and Malika Aubakirova wrote in a blog post. “If it works, it could change the way the AI ecosystem is structured, and support experts across many fields.”

Image: Unsplash