Mirendil signs $100M+ Google Cloud deal for self-improving AI research

In short: Mirendil has signed a multi-year Google Cloud partnership worth more than $100 million to get the computing power it needs for “self-improving” AI research.

What happened

Mirendil, a small AI research lab, has signed a multi-year deal with Google Cloud that is worth upwards of $100 million, according to Mirendil CEO Benham Neyshabur.

The agreement gives Mirendil access to two kinds of specialized computer chips used to train AI, Google’s TPUs and Nvidia’s GPUs. It also includes managed training clusters, which are groups of machines set up to work together (like renting a whole gym of exercise equipment instead of a single treadmill).

Mirendil is working on “self-improving AI,” also called recursive self-improvement. In simple terms, this is AI that tries to get better over time by using its own results to improve its next attempt, similar to how a person learns by doing and then adjusting.

Mirendil says it wants these systems to help automate parts of scientific and AI research, including work in medicine, biology, and materials science. Neyshabur gave Alzheimer’s disease research as one example of a problem where an AI system could keep making progress over time.

Why it matters

Training advanced AI systems can require huge amounts of computing power, which is expensive and hard to secure. Deals like this show how important access to cloud computing has become for AI startups, and how big cloud providers like Google are competing to supply the “rentable horsepower” behind the next wave of AI research.

Source: TechCrunch AI