One of OpenAI's first two patents claims inserting text between a prefix and a suffix and retraining on labeled output

2024-01-30Filings

United States Patent No. 11,886,826 B1, OpenAI OpCo, LLC, "Systems and Methods for Language Model-Based Text Insertion," Application No. 18/183,898, filed March 14, 2023, issued January 30, 2024 (inventors Mohammad Bavarian and Heewoo Jun)

United States Patent No. 11,886,826 issued to OpenAI OpCo, LLC on January 30, 2024, ten and a half months after filing, with no pre-grant publication, naming Mohammad Bavarian and Heewoo Jun as inventors. It issued the same day as United States Patent No. 11,887,367, on training a model to perform interface actions from video, and the two are the earliest patents recorded to the company. Claim 1 recites a system with memory and a processor that receives "an input text prompt comprising a prefix portion and a suffix portion," determines a set of model parameters from the prompt, accesses a language model, determines "a set of context parameters" that comprise "at least one of location, person, time, or event," generates output text from those context parameters and the model, inserts the output "into the input text prompt," and then optimizes the language model "by aligning the language model based on the language model output using machine learning," through "one or more iterative cycles of training" on outcome metrics and datasets that "comprise at least one of user-instruction data or user-labeled data based on language model output text." Claim 13 is the parallel method claim, and the patent has twenty claims. The specification describes the prefix and suffix as the text before and after an insertion point, with code examples in which the model writes the missing function between them.

What it changesThe first thing this frontier laboratory took to grant is not a model architecture but a serving mechanism with a training loop attached: the prefix-and-suffix prompt structure, the context parameters drawn from it and the retraining of the model on user-labeled output. That is the layer where the engineering is specific enough to claim and where a competitor's product can be inspected for infringement, and it is where most companies' filings should sit. Two drafting lessons are visible in claim 1. The retraining limitation, with its named data types, is what gives the claim a mechanism beyond prompting a model and inserting the answer; without it the claim would recite a result. And the context-parameter element, limited to location, person, time or event, is the narrowing a litigator will test first, so a product that conditions on other context may fall outside it. A corresponding international application was filed in 2023, so teams building code completion or document infill features should compare those features against this claim now.

More on What the AI companies are filing

← 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.