Back home

Orion

x+web 0.93
Visit source

Intro

Orion is the gaze-prediction foundation model, and the companion Orion Hosted site-analytics product, that Paradigm Machines introduced on 16 November 2026. Founder Ryan Lin (@_ryanlin10) posted at 2026-11-16T17:00:01Z (01:00:01 on 17 November 2026, Asia/Shanghai): “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.” When this scan re-checked the original post it had about 1,518 likes, 76 reposts, about 49,000 impressions and 1,372 bookmarks. That introducing post sat inside the 04:32 Asia/Shanghai window the same day, but was not stored then; this slot backfills it after checking the post and the official product pages.

The hosted product is at app.paradigmmachines.ai, under the line “See your site through your visitors' eyes.” The page says cursor movement, scrolling and pauses give away where visitors look, and that Orion’s model — trained on real eye tracking — predicts that attention from your session replays. The Orion Hosted terms (version date 2026-11-16) name the provider as Paradigm Machines, Inc., a Delaware corporation run from London, and describe two intake paths: a page script (the Tag), or imported recordings from tools such as PostHog, Mixpanel, Datadog or Sentry. The dashboard shows visits, session replays, predicted gaze and a friction score for each visit. The privacy policy is effective the same day; the contact address is [email protected].

In follow-up posts, Ryan Lin said the gaze model is trained uniquely for human interaction, unlike LLMs; that with more data and compute they now have an emergent “visual world model” that understands design taste; that a technical blog and an API are in the works; and that they have also trained a full-duplex realtime checkpoint. A demo is at paradigmmachines.ai/demo. Do not confuse this Orion with Apple’s or any other product that shares the name.

Advertisement

Explain

Orion is not a generic click heatmap. The official product page is explicit: generic heatmap tools often map cursors and clicks, and the hand trails the eyes, rests during reading, and never reaches most of what is seen. From each replay, Orion predicts where the visitor looked, moment by moment. Scrolls and pauses reveal what is being read; a page’s layout draws the eye to headlines, pictures and buttons.

The model is trained on collected eye-tracking data: where people really looked as they browsed, searched, read and typed. The page gives an example of the three seconds before a click, when the eyes reach the link before the pointer does. Orion then checks the predicted gaze for four patterns that eye-tracking research links to difficulty, rolls them into a friction score per visit, and shows the elements behind the score. Scores add up across visits so versions can be compared in an A/B test. The four patterns are: looking away just before acting, going back to something already used, eyes and pointer far apart, and long looks at buttons or fields.

The three claims in the introducing post — finding twice as many UX bugs as frontier models, cutting computer-use latency by 31%, and winning a drone race — are the author’s statements. The public materials do not fully specify the evaluation size or the comparison setup, so they should be read as official claims, not independently reproduced results. Follow-up commentary also attributes further author claims to the launch thread, including shrinking image tokens sent to a vision model by enlarging predicted hotspots (one recap said about 66%) and flagging prompt-injection messages in areas of low predicted gaze density (the author said a 96% reduction). Ryan Lin also distinguished “Attention insight” from Orion: the former generates static attention maps and does not use click, scroll or mouse data across time.

For coding agents, the product page highlights three outputs: gaze time on every element (for one visit or across every visit to a page), the screen at any moment with gaze drawn on it, and a pointer to the exact banner, button or field so a coding agent knows which code to change. Data can also be downloaded: the replay (rrweb JSON), scores (JSON and Parquet), gaze (NumPy), element boxes, and a WebM gaze video.

Usage

The official product offers two ways in. One is to add a one-line Tag on every page, which records replays, clicks, scrolls and page changes, including on single-page apps, and masks form fields. The other is to connect recordings you already have: the product page says you can sign in to PostHog and let Orion read and score those replays with nothing added to the site, or drop in an export of up to 89 days. The terms also mention imported recordings from Mixpanel, Datadog and Sentry.

There are two ways out. One is to attach Orion to an AI agent as a remote MCP server at https://app.paradigmmachines.ai/mcp, then sign in with Google; there is no key to copy. Examples on the product page include:

  • Claude Code: claude mcp add --transport http orion https://app.paradigmmachines.ai/mcp, then type /mcp in Claude Code and sign in.
  • Codex: codex mcp add orion --url https://app.paradigmmachines.ai/mcp.
  • ChatGPT: in plugins, add a custom MCP server named Orion at the same address. The page says this needs ChatGPT Business, Enterprise or Edu, or Pro for tools that only read; on Business, an admin adds it.
  • Cursor and VS Code: use the corresponding links on the product page, then sign in.

In replies on 17 November 2026, Ryan Lin also said you can call add_site over MCP, that a dashboard control for adding another site had just been added, and that an API is still being built. The other path is to download each visit’s files and run your own analysis. The official stance on trials is that if you are not happy, they will fix issues within 24 hours; you can also email [email protected]. The demo and the product are at paradigmmachines.ai/demo and app.paradigmmachines.ai.