Get the newsletter
Google · AI / ML · Filed Feb 12, 2026 · USPTO published Jun 18, 2026 · Jun 18 release

Google's New Patent Lets Every Stage of Its AI See the Full Picture at Once

Most AI models read information in one direction — layer by layer, like a factory assembly line. Google's new patent describes a system where every stage of that process can see what every other stage is doing, all at once.

By · Updated · 3 min read

Google's OMNINET Patent: AI That Reads Its Own Thinking, figure from US 2026/0170303 A1
Figure from the official USPTO publication.
Publication number
US 2026/0170303 A1(US20260170303A1)
Applicant
Google LLC
Filed · Published
Feb 12, 2026 · Jun 18, 2026
Inventors
Yi Tay, Da-Cheng Juan, Dara Bahri, Donald Arthur Metzler JR., Jai Prakash Gupta, Mostafa Dehghani, Philip Pham, Vamsi Krishna Aribandi, Zhen Qin
Status when we published
Waiting for an examiner
Document
20 claims
US classification
706/15
Parent application
Continuation of 17/592,796 (filed Feb 4, 2022)(17592796)
AI training & infrastructure

What Google's OMNINET attention model actually does

Imagine you're reading a novel, but you can only see one chapter at a time and you have to forget each chapter before moving on. That's roughly how today's AI language models process information — they pass data through one layer, then the next, without earlier layers talking to later ones.

Google's patent describes a different approach called OMNINET. Instead of that one-way assembly line, a second AI model sits alongside the main one and watches the outputs of every layer at once — early, middle, and late. It's like having a supervisor who can read all the chapters of that novel simultaneously and pull out insights no single chapter would reveal on its own.

The practical goal is a more complete understanding of language, images, or whatever data the AI is processing. By letting information flow across the entire network — not just forward — the system can catch relationships and patterns that a strictly linear model might miss.

How OMNINET taps every layer at once

The patent describes a two-model setup. The first is a standard multi-layer attention model — the kind that powers large language models like Google's own Gemini family. It processes an input (say, a sentence or a document) by passing it through a stack of self-attention layers (layers that weigh which words or tokens matter most relative to each other).

The second model is the omnidirectional model — the novel part. While the main model runs its normal forward pass, this second model collects the intermediate representations (the internal state of the data at each layer, essentially a snapshot of what the network 'thinks' at that point) from two or more of those layers simultaneously.

It then processes all of those snapshots together to produce its own output. The key insight is that later layers in a neural network typically have high-level, abstract understanding, while earlier layers hold low-level, fine-grained detail. Normally those perspectives never meet directly.

  • Layer 1 output: fine-grained token-level details
  • Middle layer outputs: partial, contextual representations
  • Final layer output: high-level semantic understanding

OMNINET feeds all of these into the omnidirectional model at once, so the final answer can draw on every level of abstraction simultaneously.

What this means for the next wave of Google AI

For Google, this matters most in tasks where context is everything — long documents, complex questions, multi-step reasoning. A model that can cross-reference its own early and late processing stages has a structural advantage over one that can only use its final layer's output. That's relevant for search, for Gemini, and for any AI product where getting nuance right matters more than raw speed.

For you as a user, the change would be invisible but consequential: answers that feel less like the AI picked the most statistically likely response and more like it actually understood the full picture. Whether Google ships this architecture into a consumer product or keeps it as research infrastructure is an open question, but the filing signals real investment in rethinking the transformer's fundamental design.

Source: US 2026/0170303 A1 at the USPTO (PDF). Full patent text and figures are in the official publication.

Get our take in your Top Stories

Liked this breakdown? Add Patentlyze as a preferred source on Google, and our plain-English take shows up more often in your Top Stories the next time Google patent news breaks.

Add Patentlyze on Google One click.
There are more where this came from

We read every patent application Big Tech publishes and send you the ones worth knowing. Plain English, free, every week.

More from Google

All 1,079 Google patents →

More in AI training & infrastructure

All 316 AI training & infrastructure patents →