Adept AI

Contribution to AI

Two years before computer use became a shipping feature at the largest laboratories, ACT-1 was shown in September 2022 doing the unglamorous version of it: a Chrome extension that watched a rendered page, treated the interface elements on it as the action space, and clicked, typed and scrolled through ordinary web apps from a single typed instruction. The framing outlasted the demo. It argued that the useful unit of work was an action taken against software people already used, which meant no API integrations, no rewriting of enterprise tooling, and a model scored on whether the task finished rather than on how the prose read. The open releases carried the same bias. Persimmon-8B arrived in September 2023 under a permissive licence with a 16k context, and Fuyu-8B a month later discarded the image encoder that every comparable multimodal system depended on, feeding image patches directly into a decoder-only transformer so that screenshots at arbitrary resolution could be read and objects on screen returned as bounding boxes — vision built for interfaces rather than photographs. The ending became its own precedent. Amazon licensed the models and hired David Luan with much of the team in June 2024 while the company continued to exist, one of the earliest instances of the license-and-hire arrangement that then absorbed several AI startups without triggering merger review. Luan's Amazon lab shipped Nova Act, a browser agent, in March 2025.

Drafted with AI and edited by hand (claude-opus-5, reviewed 2026-08).