Our Path

The Truth of Slop

Post-trained models show 10x linguistic diversity loss from base models. We think this is a fixable problem.

1. Define Slop: what is AI slop? What is its opposite?

   Catalog Slop

   See our catalog of slop

2. Measure Slop
   1. Slop Detector: Like Pangram, but for slop. Our reading of Sturgeon’s Law says: humans can make slop, too.
   2. SlopBench: There is no slop benchmark. Create a benchmark comprehensively measuring slop output in frontier models.

The Origin of Slop

Identify root cause of slop from first principles: at which point in the AI training process does slop get introduced?

1. Make Slop: having found the root cause of slop, create an exaggerated model with amplified slop tendencies.

The Cessation of Slop

Propose an alternative post-training architecture demonstrating preservation of linguistic diversity.

The Path to Defeating Slop

1. Build a better AI text humanizer skill from first principles: If slop is food pigs eat, we think humanizers are like lipstick on a pig: solving the symptom and not the root cause. That said, we think a superior humanizer provides a serious performance improvement for content and advertising (and a serious experience improvement for users on the receiving end of said content).
2. Eradicate Slop: Implement an alternative post-training architecture, usable by AI researchers, AI companion products, creative writers, and others.
3. Serve RL environments superior to current ad-hoc data labeling flows, for big labs too dug-in on RLHF.