Generating and Editing Text with Language Models

This European patent introduces a clever way for cameras and tracking systems to better handle moments when an object temporarily disappears from view—like when a person walks behind a tree or a car passes behind a building. Instead of losing track completely, the system learns from past tracking failures. It records where objects typically vanish and where they tend to reappear, building a kind of “map of blind spots.” Over time, this map helps the system make smarter guesses about where a hidden object is likely to go next. It can even adapt how long it waits before giving up on a lost object or decide where to focus its efforts to re-identify something. The result: smoother, more reliable object tracking, even when visibility isn’t perfect.

This European patent describes a method for training a machine learning model, particularly in the context of communication networks like 5G. The primary focus is to improve the accuracy of machine learning models trained using federated learning (FL) by addressing the issue of data heterogeneity among the FL clients. The proposed method involves determining the similarity between the federated learning server's test data and the clients' training data and using this determination to decide whether or not to incorporate a client's model update into the global model.

This patent describes a method for enabling more natural conversations with automated assistants. The core problem addressed is the unnatural, turn-based nature of current automated assistant interactions, which doesn’t reflect how humans converse. The patented invention introduces a “soft endpointing” mechanism that allows the assistant to better understand when a user has paused versus completed their utterance, enabling more fluid and natural conversations. The system can provide natural conversational outputs like “Mmhmm” or proactively ask for clarification, rather than immediately fulfilling potentially incomplete requests, thus improving the user experience.

This patent describes a novel system and method for image inpainting, where missing or undesirable regions of an input image are filled in using guidance from a separate "guide image." The core innovation lies in the use of machine learning models, particularly a Style Generative Adversarial Network (StyleGAN), to combine visual features from both the input image and the guide image in a deep latent space. This approach aims to generate inpainted content that is not only consistent with the remaining parts of the input image but also incorporates desirable visual characteristics from the guide image, offering greater control and improved quality, especially for large or complex missing regions.

This patent discloses systems, methods and devices for prioritizing calendar events with artificial intelligence to resolve scheduling conflicts. When a request to schedule a new event clashes with an existing one, the system compares their "event priority scores," generated by a statistical machine learning model considering various factors. If the new event's score is higher, a selectable option to replace the conflicting calendar event with the new calendar event may be presented.

This patent discloses systems and methods developed by OpenAI for automatically generating and editing text using language models (LMs). The core innovation lies in a flexible approach that takes an input text prompt and user instructions to access a language model, generate output text, and then edit the original prompt by replacing portions with the LM's output. The patent also covers systems for automatically generating and inserting text based on prefix and suffix prompts. A key emphasis is placed on iterative refinement of the LM through training and optimization based on user interactions and labeled data.

This patent discloses techniques for adapting previously-annotated training examples into updated training examples for training machine learning models. The core idea involves identifying a specific part (the "find expression") within a targeted subset of training examples (defined by a "filtering constraint") and replacing it with a new part (the "replacement expression"). This process allows for efficient modification of existing training data to reflect changes in system capabilities, user expectations, or to correct inaccuracies.

This patent describes an innovative system and method for automatically configuring electronic devices using artificial intelligence (AI). The core idea involves leveraging device usage data (telemetry data) as input for machine learning models to predict future events and proactively configure the device's operating system, applications, and hardware accordingly. This approach aims to personalize the user experience, optimize device performance (speed, memory, battery), improve resource allocation, and even enable preemptive actions and automatic remediation. A key aspect is the use of cloud-based machine learning models tailored to device metadata, with the potential for local augmentation using device-specific, non-shared information.

This patent describes a system and method for generating cyber-threat scores by leveraging machine learning (ML) and explicitly considering the quality of the sources providing the threat intelligence. The core innovation lies in training an ML model not only on the presence or absence of threat indicators and their classifications but also on "quality metrics" associated with the sources. During inference, the system identifies votes from various sources on a new indicator, assesses the quality of those sources, and generates a threat score based on the trained ML model, which has learned to weight source reliability. This approach aims to improve the accuracy and reliability of threat intelligence by mitigating the impact of low-quality or unreliable sources.