humanity.ai-A1 Teaches Itself in New Gesture Demo

Gesture recognition is only the beginning. We're actively seeking AI researchers and engineers to join our team. Contact us at: hello@humanity.ai

###

Watch humanity.ai-A1 learn gesture recognition from scratch. In one live session it spawns a new expert, grabs a Kaggle dataset, trains on a 24 GB Mac Mini, and deploys the skill, showing the power of self-learning modular AI.

From “I don’t know” to “I’ve got it” in one uninterrupted session.

What if your AI could recognize when it has a missing skill, go find the data it needs, train itself, and plug the new capability straight into production?

That’s exactly what happens in our latest video, where Evgeniy Usatov (Chariot Technologies Lab’s Head of Patent Strategy) asks humanity.ai-A1—nicknamed “Adam” in this demo—to recognize hand gestures.

The Live Sequence
1) Challenge issued: Evgeniy asks Adam to identify a series of hand gestures. Adam can't do it.

2) Expert creation: Evgeniy asks Adam to go learn hand gestures. Adam spins up a blank “Gesture-Expert” container inside the iCon architecture.

3) Autonomous data hunt: Adam finds and plugs in a gesture-recognition set from Kaggle.

4) On-device training: Using the Mac Mini’s 24 GB of RAM, the new expert fine-tunes a lightweight vision model, logs metrics, and saves checkpoints.

5) Hot-swap deployment: The trained weights are wrapped in the container, registered with the Conductor, and verified by the Verification Module.

6) Skill acquired: Evgeniy flashes the "peace" sign. Adam nails it, then proceeds to recognize gestures for "call" and "rock"—all in the same session.

Why It Matters
This closed-loop learning cycle is the first step toward an AII that can grow like a human apprentice: identify a knowledge gap, study, practice, and deliver—all without disrupting the rest of the system.

Under the Hood
iCon containers: isolate library versions so nothing conflicts.
DisNet orchestration: routes requests to the new expert the moment it registers.
FIONa arithmetic engine: keeps RAM usage low (even while training) by converting control logic into pure math operations.
Audit trail: every dataset, hyper-parameter, and verification check is logged for reproducibility.

What’s Next?
Sign up for our email newsletter at humanity.ai to get the latest updates delivered to your inbox.

Видео humanity.ai-A1 Teaches Itself in New Gesture Demo канала humanitydotai