Meta Announces Eigengrau: Embodied Foundation Model Trained on 4.2B Wearer-Hours

MENLO PARK, CA — 2027-09-14

Meta Platforms, Inc. today announced Eigengrau, the company’s first embodied foundation model, trained on a proprietary corpus of 4.2 billion wearer-hours of multimodal capture from the Ray-Ban × EssilorLuxottica and Meta AI Glasses product lines.

Eigengrau is described by the company as “the first foundation model to learn from attention as it actually occurs in the world” — a characterization Chief AI Officer Yann LeCun repeated three times across the press materials. The model’s training corpus comprises:

• 4.2 billion hours of forward-camera video (average 3.1 hours per device per day across the active fleet of ~2.4M units)
• Synchronous gaze from on-glass eye-tracking (standard in the EssilorLuxottica line since late 2025)
• Head pose and IMU at 200 Hz
• Audio (ambient, directional, and — where applicable under local consent regimes — speech)
• Location at 1 Hz (coarse, urban/transit aggregated)

All streams are co-registered at frame level. Meta’s materials emphasize that “no first-party user speech is retained in pretraining weights,” a claim some researchers are calling technically true and operationally sideways.

What it does

Eigengrau can:

• Predict what a person is about to look at three seconds before they look there, across indoor and outdoor settings, with accuracy comparable to human prediction of other humans.
• Describe the task a wearer is performing without being told — and flag task state transitions (“moving from folding laundry to making a phone call”) with the granularity of a human observer.
• Generate novel embodied scenarios consistent with the statistics of human attention — a capability Meta demonstrates by asking the model to compose a ten-second video of “someone finding their keys in a cluttered apartment” that reviewers describe as indistinguishable from a genuine wearer capture.

The last capability is the one that will be discussed for the next six months. Its implications for synthetic media, for training data attribution, and for what counts as a record of a life are not addressed in Meta’s announcement.

What it means commercially

Meta will make Eigengrau available through:

• Horizon Assistant (consumer) — a context-aware AI that “understands your day as you live it” and is present on Ray-Ban, Vision, and (forthcoming) in-home devices.
• Horizon for Developers — API access to embodied-reasoning primitives for third-party apps building on Meta’s hardware.
• Horizon for Research — selective, vetted, non-redistribution access for academic partners. No institution outside the Meta Research Consortium has been granted pretraining-corpus access.

The announcement explicitly does not commit to corpus disclosure under any proposed regulatory framework. A spokesperson referred reporters to the company’s Provenance Transparency Statement, which does not exist as of the announcement’s publication.

What analysts are saying

“The capability gap Eigengrau opens isn’t one any competitor can close by training harder. Google and Anthropic have the models; they do not have the fleet. Snap has the fleet fragments; it does not have the scale. Apple has the engineers; it does not have the retail presence. For embodied foundation models, Meta has one-player’s-worth of runway.” — Maria Okafor, Independent AI Industry Analyst

“We have been collectively naming this problem in public hearings since early 2027. Today is not a surprise. Today is the day the problem is settled.” — Dr. Sal Mateo, Columbia University (whose [[2029-ground-truth|Ground Truth methodology paper]], then in preprint, would later be cited in the EU’s CDPA hearings)

“I left Reality Labs because I did not want to be the person who built this. Today I am the person who watches it. The difference is not always as large as I would like.” — Attributed to an anonymous [[2026-severance|OHC coordinator]], via the commons newswire. (The OHC’s [[2027-commons-statement-on-capture|formal response]] was filed the same afternoon.)

Regulatory status

Eigengrau is released in all markets except those where wearer-consent frameworks explicitly prohibit capture-corpus training — currently no jurisdictions meet this standard. The EU’s [[capture_fleet_consolidation|Capture Data Provenance Act]] is in committee and will not move until late 2028. By the time it passes, Eigengrau’s pretraining will have been complete for fifteen months. The corpus cannot be untrained.

Meta’s press materials do not address this timeline.

Meta Platforms, Inc. press release, lightly edited for in-world formatting. Annotations by the [[ohc_ai_human_commons|OHC News Service]] are marked in italics.