Build AI's early egocentric release — 400k action labels, 2.5k clips, 2× open-source dataset size (Eddy Xu tweet, Oct 22 2025)

TL;DR

Eddy Xu (Build AI) tweeted on Oct 22, 2025 announcing the release of “the largest egocentric dataset of physical jobs” — 400k action labels across 2.5k clips, claimed to be 2× the size of any prior open-source egocentric dataset, with a download link in the thread. This sits at the earliest documented point in Build AI’s egocentric-dataset scaling cadence: it precedes the Egocentric-10K release (Nov 10, 2025; 10K hours, 2,153 workers, 1.08B frames) by roughly three weeks, and was eventually superseded by Egocentric-100K (Dec 2025) and Egocentric-1M (Apr 2026). The tweet itself contains no architectural or methodological detail; its value is as the origin marker of the cadence.

Key claims

The release is described as the largest egocentric dataset of physical jobs at the time of posting [tweet body].
The scale is stated as 400k action labels and 2.5k clips [tweet body].
The release is positioned as 2× the size of the prior open-source egocentric dataset [tweet body].
A download link accompanied the announcement (linked from the tweet, not transcribed here) [tweet body].

Method

There is no method to describe — this is a dataset-release tweet with no accompanying paper, model card, or technical report at the linked URL. The mechanics inferred from later Build AI releases (Egocentric-10K → 100K → 1M) and consistent with the framing here: head-mounted glasses worn by factory workers in Southeast Asian manufacturing sites, with action labels presumably derived from worker-task taxonomies. None of this is stated in the Oct 22 tweet itself.

Results

No model results. The artifact is the dataset. Two-point context for the scale claim: by November 2025 Build AI had grown the corpus to 10,000 hours (Egocentric-10K), and by April 2026 to ~1M hours (Egocentric-1M). The Oct 22, 2025 announcement is therefore the smallest of the four documented releases in the lineage and predates the explicit hour-based framing that later releases adopted.

Why it’s interesting

This tweet is interesting mainly as a temporal anchor for the Build AI scaling cadence rather than as a standalone artifact. The already-filed Egocentric-1M — largest egocentric video dataset (Build AI / Eddy Xu announcement) covers Egocentric-1M (the Apr 2026 culmination) and walks through the 10K → 100K → 1M progression; this Oct 22, 2025 tweet documents the step before that progression begins, when the corpus was still measured in clips/labels rather than hours. The growth implied — from 2.5k clips (Oct 2025) to ~2M clips at 180s mean (Dec 2025) — is the kind of cadence that’s notable if it holds. For Luma’s embodied-AI / world-model interests, it’s complementary to Action100M: A Large-scale Video Action Dataset (which scales VLM-labelled internet-video action data to 147M segments): Action100M leans on dense automatic labelling over scraped third-person video, whereas the Build AI lineage relies on first-person capture from workers with custom hardware. Different bet on the data-quality / scale frontier.