Welcome to the Genmorphics Blog

Welcome! This is the new home for ideas, lessons, and research from the team at Genmorphics.

Why a blog?

We spend our days working on one of the least glamorous but most consequential parts of the AI stack: training data. Preference labels, instruction pairs, model evaluations, safety tests — the quality of these inputs quietly determines the quality of the models built on top of them.

Most of what we've learned doing this work at scale isn't written down anywhere public. This blog is where we change that.

What to expect

We'll be writing about:

• RLHF and preference data — what makes a good preference label, and how to catch a bad one
• SFT dataset design — instruction diversity, response quality, and common failure modes
• Model evaluation — building evals that actually predict real-world performance
• AI safety testing — red-teaming, policy adherence, and edge-case discovery
• Working with domain experts — why real expertise beats generic crowdsourcing, and how to operationalize it across 40+ languages

Stay in touch

You can subscribe via RSS, visit our website, or log in to the platform.

Thanks for reading — more soon.