Most AI models exist in a vacuum, generating probabilities without regard for the physical machine running them. agnt/smth.sh changes that.
By executing ./smth.sh, you deploy a closed-loop inference engine that treats your local system not just as a host, but as a core component of the model itself.
Your computer "feels" your activity through heat, electrical current, and scheduler latency. agnt/smth.sh captures these physical signals alongside the visual data.
Standard AI models are prone to "hallucination"—generating outputs that look plausible but are logically or physically impossible. agnt/smth.sh solves this through a rigorous validation layer.
Because our training data includes the physics of the machine, the model can verify its own logic. If it attempts to predict a visual state that contradicts the known energy expenditure of your hardware, the system flags the anomaly. The result is a model that is inherently constrained by the physical reality of your machine, leading to drastically improved consistency and reliable, verifiable outcomes.
Well, maybe. That's what we are researching in an applied way.
Automated inference and reaction allows you to cut the cord on big tech
With agnt/smth.sh model your system bit by bit, tick by tock, pixel by pixel.
With hash/browser.sh explore your system models, edit the visual representation, link system to experience.
With wotkees/ios sync models 1:1 or 1:Many with helpful, intuitive user interfaces.
The system is packaged as a high-performance binary. To begin your own local telemetry-grounded inference session, simply run:
$ ./smth.sh --init --visual-ingest
Your machine will start building its own truth.