Yesterday I wrote about the @PrismaXai roadmap.
Today I studied Eval Engine, their data evaluation system
AI models watch videos and give a score from 0 to 1. How valuable the data is for training robots.
Four criteria: - is there movement - how interesting the content is - frame quality - uniqueness
Then I realized an important thing
DINOv2 was trained to match pieces of an image and guess hidden fragments. Simple tasks.
But it learned on its own to find wings in birds and airplanes. To see similarity between an elephant and a statue of an elephant. Without a teacher
In robotics, it’s easy to collect video, but expensive to label.
The solution: make the model compress video.
To compress a video with a banana well, you need to understand how bananas are peeled. To describe the movement precisely.
So a simple task - compression creates an understanding of actions.
Eval Engine creates a closed loop
High-quality data → higher score → higher reward → operators try harder
I like the approach. Not just a platform for work, but a system with internal incentives for quality.
Tokens align incentives. Operators get paid for quality, not for time. Companies get filtered data without intermediaries.
Robotics is bottlenecked by data. PrismaX solves this problem in a decentralized way and with the right incentives.
English
