Google claims the procedure for sizing a model's depth and width to a compute budget, filed five days before its authors posted the paper

2026-10-15Filings

Listen · Thinking Makes It So, 2026-10-19 · 7 min

All episodes · Subscribe: Apple Podcasts · Spotify

United States Patent Application Publication No. 2026/0320392 A1, Google LLC, "Scaling Neural Network Architectures to Computational Resource Constraints," Application No. 19/485,257, the national stage, entered November 17, 2025, of PCT/US2023/029718 filed August 8, 2023, claiming priority to Provisional Application No. 63/502,790 filed May 17, 2023, published October 15, 2026

Claim 1 is a method for choosing a neural network architecture for a given compute constraint in three steps. For each architecture dimension to be adjusted, it fits a power law relating a performance metric to that dimension and to compute, from evaluations of networks identical except in that dimension, and takes a scaling exponent from the fit. It then picks a base architecture from candidate networks of different shapes, chosen by grid search in claim 13 and trained under a smaller compute budget, taking the best of them on performance against compute in claim 5. Finally it modifies the base architecture's dimensions using the exponents, the base architecture's compute use and the target constraint, in claim 6 by raising the ratio of target to base compute to a power set by the exponent. The dependent claims name depth and width (claim 17), the width of the multi-layer perceptron in a Transformer block (claim 18), image processing as the task (claims 21 and 23), and training a network of the output shape (claim 22); claims 25 and 26 repeat the method as a system and a storage medium. The four inventors posted "Getting ViT in Shape: Scaling Laws for Compute-Optimal Model Design" to arXiv on May 22, 2023, five days after the provisional, and reported the SoViT-400m/14 model that the method produced. The file history shows a request for immediate examination under Section 371(f) and docketing to an examiner on July 16, 2026, with no office action yet.

What it changesFive days is the whole margin: the United States grace period would have forgiven the paper, but the European Patent Office and China's office have no grace period for an inventor's own paper, so without the May 17 provisional the paper would have been prior art against the counterparts in the two largest markets outside the United States, and a research group that files the provisional on the way to the arXiv submission button, as this one did, keeps those rights for the cost of a form. On the United States claim as published I would expect a subject matter eligibility rejection: claim 1 fits curves, takes exponents and ends in an architecture, which is a set of numbers describing a model no one has yet trained, and the eligibility decisions that have favored machine learning claims, Ex parte Desjardins among them, involved claims that train or run a model, not a calculation whose output is a design. The repair is in the application already, which is to move claim 22's training step, or claim 23's processing of image data with the resulting network, into claim 1. Detection is the other problem with a design-time method: it is practiced inside a lab, and the only evidence that a competitor used it is the competitor's own paper, so a claim like this earns its budget as the thing that stops a published recipe from being free for everyone, and the family should add a claim to the model shape that the recipe produced.

More on What the AI companies are filing · Eligibility of software and AI claims · Prior art, public disclosure and obviousness

← All recent developments · Search all developments

Burbage Law is the practice of Jon Burbage, a registered patent attorney. 342 United States patents have issued from applications he wrote; the full list is on the Patents page. More about Jon Burbage.

This is general information about a published decision, agency guidance or piece of legislation. It is not legal advice and does not create an attorney-client relationship. Items concerning offices outside the United States are reported for information only and are not counsel on the law of those countries.