
Introducing Orion, a foundation model that predicts where you look. Orion finds twice as many UX bugs as frontier models, cuts computer-use latency by 31%, and even won a drone race.

Building, Physics & Philosophy @ Oxford, IOAA Gold

Introducing Orion, a foundation model that predicts where you look. Orion finds twice as many UX bugs as frontier models, cuts computer-use latency by 31%, and even won a drone race.

Luke and Avi are incredible, they held conviction on the bet on duplex voice models long before it was mainstream, check out their work
Introducing Interact-0, the first in a new generation of always-on models. Interact-0 is a full-duplex speech model designed around human interaction. It is the only model that knows when to speak in groups, and can be active without barging in when it shouldn’t. We share our thesis, approach, and early results below. interact.inc/blog/interact-0…

A very interesting area of research! The future of voice interfaces requires a large amount of visual understanding as well Both will come hand in hand
Introducing Orion, a foundation model that predicts where you look. Orion finds twice as many UX bugs as frontier models, cuts computer-use latency by 31%, and even won a drone race.

Orion prevents computer-use agent prompt injections. We reduce attacks 96% by flagging messages that appear in areas with low predicted gaze density.

Orion resembles human behaviour and can spot the odd one out. Predicted gaze concentrates on a misplaced shoe, female duck, pink blossom tree, and red leaf.

Orion generalises to diverse situations. On dashcam videos, predicted gaze can be used to flag hazards.

Orion improves auto-zoom for product launch videos. By zooming in on predicted gaze, Orion ensures you aren't cropping out important information.

Orion unlocks new low-latency VLM drone control. By magnifying regions of predicted gaze, we triple the frequency of decisions whilst keeping critical features visible. This allows reliable completion of a drone race.

Orion saves you waiting for computer-use agents. By magnifying regions of predicted gaze whilst shrinking the periphery, we reduce model call time 31%, with no measurable loss in task success. Image tokens are also reduced by 66%.

Orion finds friction in your UX and product. Connect your PostHog session replays and instantly find 2x as many issues. Try it yourself: app.paradigmmachines.ai/

Orion is trained on real eye-tracking. We scale to 20x the data size compared to previous saliency models and apply a transformer architecture over the complete computer interaction stream.



Free doughnut method for GTM has officially been patched

Our keystroke data collection is privacy-preserving, capturing timing (press and release timing) rather than content

Synthetic sentiment labelling with VLMs allows us to push past previous data walls VLMs can label based off semantic (prompt) and visual (UX) context