Ryan Lin (@_ryanlin10) profile picture

Building, Physics & Philosophy @ Oxford, IOAA Gold

Oxford, England
Joined November 2022
Photo by Ryan Lin (@_ryanlin10)Photo by Ryan Lin (@_ryanlin10)Photo by Ryan Lin (@_ryanlin10)Photo by Ryan Lin (@_ryanlin10)Photo by Ryan Lin (@_ryanlin10)Photo by Ryan Lin (@_ryanlin10)Photo by Ryan Lin (@_ryanlin10)Photo by Ryan Lin (@_ryanlin10)Photo by Ryan Lin (@_ryanlin10)Photo by Ryan Lin (@_ryanlin10)
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.
67
113
18
2,344
84,892
UX designers are now using Orion to improve conversion and engagement It's great seeing the diverse use of our new foundation model Try it yourself github.com/ryanlin10/orion-g…
Replying to @_ryanlin10
Thank you for the Orion Gaze API, it's been genuinely useful. We ran about 50 screenshots through /v1/predict. Setup took minutes, the docs were clear, predictions came back in seconds, and DELETE cleaned everything up exactly as documented. EU-only processing was a big plus for us. It changed a real design decision. Our Run button was getting almost no predicted attention (about 0.1–0.2× chance). Making it bigger, bolder or adding an accent did nothing. Moving it to the top of the image column took it to roughly 4–10× chance, at a small cost to the image itself. We're adopting that placement, and screening variants with Orion is now a standard step before we show design finalists. Some feedback, in case it's useful: • Per-element scores. Attention share for a region or CSS selector would save us the post-processing we do on the grid. • Phone viewports. 390×844 screenshots are accepted, but the grid stays 64×40 and attention piles into the top-left corner, so we couldn't rank phone layouts. Phone-tuned predictions would be great. • Known priors. We noticed a pull toward the left edge and the top of the page. A short note on the model's known biases would help us read the results fairly. Thanks again, and looking forward to where this goes
1
5
336
Ryan Lin retweeted
Replying to @_ryanlin10
Thank you for the Orion Gaze API, it's been genuinely useful. We ran about 50 screenshots through /v1/predict. Setup took minutes, the docs were clear, predictions came back in seconds, and DELETE cleaned everything up exactly as documented. EU-only processing was a big plus for us. It changed a real design decision. Our Run button was getting almost no predicted attention (about 0.1–0.2× chance). Making it bigger, bolder or adding an accent did nothing. Moving it to the top of the image column took it to roughly 4–10× chance, at a small cost to the image itself. We're adopting that placement, and screening variants with Orion is now a standard step before we show design finalists. Some feedback, in case it's useful: • Per-element scores. Attention share for a region or CSS selector would save us the post-processing we do on the grid. • Phone viewports. 390×844 screenshots are accepted, but the grid stays 64×40 and attention piles into the top-left corner, so we couldn't rank phone layouts. Phone-tuned predictions would be great. • Known priors. We noticed a pull toward the left edge and the top of the page. A short note on the model's known biases would help us read the results fairly. Thanks again, and looking forward to where this goes
1
1
1
368
Luke and Avi are incredible, they held conviction on the bet on duplex voice models long before it was mainstream, check out their work
You’re unable to view this Post because this account owner limits who can view their Posts. Learn more
1
11
669
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.
1
1
3
432
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.
1
34
2,263
Orion prevents computer-use agent prompt injections. We reduce attacks 96% by flagging messages that appear in areas with low predicted gaze density.
3
1
25
2,166
You can plot a straight line on the METR plot within error bars
1
2
772
Be careful vibe coding guys
343
Free doughnut method for GTM has officially been patched
301
We are releasing an open-source pipeline to predict your user sentiment from computer actions Connect to your existing session replay tool Prevent loss of signal with high-fidelity mouse capture Pick refined recipes to train sentiment models on your user data Check it out: paradigmmachines.ai/orion
6
8
642
Our keystroke data collection is privacy-preserving, capturing timing (press and release timing) rather than content
17
Synthetic sentiment labelling with VLMs allows us to push past previous data walls VLMs can label based off semantic (prompt) and visual (UX) context
22
The issue with sentiment classification is there hasn't been enough labelled data to do it at scale Sentiment classification demonstrates log linear scaling
209