Microsoft Patente

Regenerable items serve as identifiers or keys or both. A regenerable items tuple is generated by combining context-specific identifiers into a string, hashing the string, and designating individual hash portions as respective regenerable items. Regenerable items maintain artifact tracking continuity after a reinstall, a reboot, or another intervening event without retrieval of a previously generated value from a registry or state file. Some regenerable items conform to a universally unique identifier specification. Some regenerable items are free of reliance on nondeterministic content, and some are free of reliance on time-dependent content.

Methods, apparatuses, and products for dynamic constraint- and resource-based artificial intelligence (AI) deployment provisioning, including: receiving a request to create an AI deployment in a cloud computing environment, wherein the request comprises one or more constraints for a selecting a model to be executed in the AI deployment; determining, in response to the request, a resource availability of the cloud computing environment; generating, based on the resource availability and one or more optimization objectives for the AI deployment, a deployment specification, including identifying a model that satisfies the one or more constraints and is able to be executed in the cloud computing environment based on the resource availability; and generating, in the cloud computing environment, the AI deployment based on the deployment specification.

This document relates to utilizing spare cores to mitigate hardware failures on a computing device. For instance, in some cases, tenant workloads can have sensitive data that executes on a given core of a device. In some cases, a hardware fault in the system may occur that affects the core running the tenant workload, and there may be an action that can be taken in mitigation software or firmware to fix that hardware fault. However, because of the sensitivity of the data on the core experiencing the fault, it is not always possible to run the mitigation algorithms on the core that is affected by the fault. The disclosed implementations can utilize a spare core to execute a hardware mitigation algorithm on another core that can clear the fault, without accessing any sensitive data on the core that is affected by the fault.

Components in a system-on-chip (SOC) may communicate through a mesh interconnect. Such components include processing circuits that may initiate memory transactions and memory circuits that may be access but memory transactions. The SOC also includes access circuits to control memory transactions directed to memory circuits through their component interfaces. The access circuits may receive memory transactions and generate indications of whether a processing circuit that initiated the memory transaction has authority to access a region of memory in the memory circuit according to the memory transaction. In some examples, execution of memory transactions may proceed according to the generated indications. In some examples, a desired access (e.g., reading, writing, executing) to a targeted memory address may be indicated by an instruction in the memory transaction.

A technique processes input information using a machine-trained model in a context-specific manner. The technique includes, for a particular processing block: receiving a set of input tokens; transforming the set of input tokens into a set of output tokens; and generating a control decision based on the output tokens. For a first state, the technique appends at least one of the output tokens to the input tokens to define a new set of input tokens and then repeats the transforming and generating operations. For a second state, the technique passes at least some of the set of output tokens to a next processing block. The token(s) passed to the next processing block convey supplemental information that does not necessarily have a verbal counterpart. The technique has the end effect of invoking transformation operations in a context-based as-needed basis, which, in turn, reduces latency and the consumption of resources.

Disclosed are techniques for determining the probability that a computing resource will be compromised during a cyberattack. In some configurations, a collection of computing resources is modeled with a resource graph. Computing resources are represented by nodes while connections between computing resources are represented by edges. Random walks are performed through the resource graph. Different random walk algorithms may be applied to identify different properties of the resource graph. For example, nodes that are visited frequently by the random walks may correspond to computing resources that are likely to be compromised during a cyberattack. Various security operations may be performed in response to determining that a computing resource is likely to be compromised, including increased security measures, introduction of honeypots or honeytokens, sending or prioritizing security alerts, etc.

Load balancing and distributing issue requests for resolution by assignable resources is described. A multimodal model determines attributes of an input issue request. A lower dimensional representation of the input issue request is generated based on the attributes; and compared to other lower dimensional representations of other issue requests to identify prior similar issue requests. The multi modal model outputs indications of similarity strength for the prior similar issue requests, and the resources assigned to resolve the prior similar issue requests. An expertise level for each assignable resource is determined based on the similarity strength and the previously assigned resources. A routing metric for each assignable resource is determined based on a combination of the expertise level and an availability of each assignable resource; and the input issue request is routed to an assignable resource in the assignable resources based on the routing metric.

Systems and methods for provisioning a sovereign cloud root certificate, retrieve, during a boot process of a sovereign cloud, a root certificate from a first source; measure a cryptographic measurement of the root certificate in a Trusted Platform Module (TPM) of the sovereign cloud, which is stored in a Platform Configuration Register (PCR) of the TPM; generate a Certificate Signing Request (CSR) including the TPM-stored measurement for a leaf certificate based on the PCR; transmit the CSR to a certificate authority for verification; wherein the CSR is forwarded to a Hardware Attestation Service (HAS) for attestation; receiving an attestation result at the certificate authority, the attestation result indicating whether the CSR is valid; and based on the attestation result, receive by the sovereign cloud, the leaf certificate when the attestation result is positive, or a rejection of the CSR when a discrepancy is detected.

A system implements techniques for executing a collaboration session between at least one human participant, at least one robotic device participant, and at least one artificial intelligence agent participant. The collaboration session can be executed in relation to a mission to be completed within a geographical environment. The system described herein generates a simulation environment in the context of the collaboration session. The simulation environment enables the simulated execution of a simulated version of an actual robotic device for validation purposes. For instance, a dangerous task to be executed by the actual robotic device may be simulated before actual execution to ensure the safety of the task is validated. Or, a complex task to be executed by the actual robotic device may be simulated before actual execution to ensure the effectiveness and/or efficiency of the task is validated.

An adaptive tactile feedback touchpad with customizable texture simulation adapts feedback to provide realistic, nuanced sensations associated with different textures by sensing touch input attributes, such as input location and force, and determining additional touch attributes, such as direction and velocity, from a time-series set of touch inputs. Touchpad haptic actuators implemented in one or more axes of a touchpad can be controlled with varying amplitudes (e.g., weights) to provide consistent tactile sensations regardless of the location and direction that a user touches and moves a finger on the touchpad. A haptic waveform can be selected, for example, based on the simulated texture and input velocity while a waveform amplitude can be determined based on input force, direction, and velocity. The haptic actuator(s) can be controlled by the selected waveform and determined amplitude to provide realistic tactile sensations regardless where the user interacts with the touchpad.

Methods, apparatuses, and products for securing a data lake using artifact-level security, including: storing, in a data lake of a data analytics platform, for one or more data artifacts of a plurality of data artifacts stored in the data lake, artifact-level security data defining permissions to access a corresponding data artifact by one or more roles, wherein the data lake is accessible to a plurality of workloads in the data analytics platform; receiving, by the data lake, a request to access a particular data artifact of the plurality of data artifacts; and controlling access to the particular data artifact using the artifact-level security data for the particular data artifact.

Techniques are described herein that are capable of increasing efficiency of a kernel using streaming multiprocessor-level time estimation. Tiling strategies for performing respective implementations of a matrix multiplication operation are defined by taking into consideration dimensions of first and second matrices that the matrix multiplication operation is configured to multiply. Estimated amounts of time or estimated latencies, which are associated with a kernel performing the respective implementations of the matrix multiplication operation using the respective tiling strategies, are calculated. The kernel is configured to implement an identified tiling strategy to perform a subsequent implementation of the matrix multiplication operation as a result of an estimated amount of time associated with the identified tiling strategy being no greater than an estimated amount of time associated with each other tiling strategy.

Methods, apparatuses, and products for securing data lake tables using row-level security, including: accessing row-level security data for a table stored in a data lake of a data analytics platform, wherein the row-level security data comprises, for one or more roles, a corresponding expression applicable to the table; identifying, for each role of the one or more roles, a subset of a plurality of rows of the table accessible to a corresponding role by satisfying the corresponding expression; and storing, in the data lake, filtering data identifying, for each role of the one or more roles, the subset of the plurality of rows of the table accessible to the corresponding role.

A method, computer program product, and computing system for developing a custom operating system (OS) and software development kit (SDK) package. A feature of an application is determined and a subset of operating system (OS) software components of a plurality of OS software components available for the OS is identified, the subset of OS software components being required to implement the feature. An OS is assembled with the subset of OS software components. A subset of software development kit (SDK) software components from a plurality of SDK software components available for a SDK, the subset of SDK software components being associated with the subset of OS software components. The SDK is assembled with the subset of SDK software components and a software package including the OS and the SDK is generated.

One or more follow-up response recommendations relating to a question-answer event processed by an artificial intelligence assistant are generated. The one or more follow-up response recommendations identify questions from a predefined response library. A weighted average of embeddings is generated for each question-answer conversation of historical question-answer events collected via telemetry. A historical question-answer conversation includes a chronologically ordered sequence of question-answer events. A self-supervised weighted embedding dataset is generated and includes the weighted average of embeddings for each historical question-answer conversation and at least one corresponding semantic nearest neighbor question from the predefined response library. Topic weights of a follow-up question recommender are tuned by training the follow-up question recommender using the self-supervised weighted embedding dataset. The follow-up question recommender executes in an inference mode on the question-answer event to generate the one or more follow-up response recommendations of questions from the predefined response library.

The technology described herein enables a remote desktop application to provide visual content to displays associated with multiple client devices, not just the primary client device serving as the active endpoint. Remote desktop software allows a user to access and control a computer or server from a remote location. Conventionally, a remote desktop session is limited to display of content through displays communicatively coupled to the primary client. The technology described herein expands display options to include displays associated with a user's other client devices that have a remote desktop application installed. For example, a user may access a virtual machine on a primary client but use the screen on a secondary client to view additional content provided by the virtual machine.

Dynamically loading endpoint data during System-on-Chip (SoC) validation in processor-based devices is disclosed herein. In one exemplary embodiment, a processor-based device, by executing an SoC validator, obtains a path for an endpoint node of an endpoint tree data structure. The path comprises node identifiers including a root node identifier of a root node, intermediate node identifiers of corresponding intermediate nodes, and an endpoint node identifier of the endpoint node, and corresponds to a hierarchical path from the root node of the endpoint tree data structure to the endpoint node. The SoC validator traverses the endpoint tree data structure from the root node to the endpoint node based on the path, and retrieves value data for the endpoint node based on the traversal. The SoC validator then generates an endpoint object representing the endpoint using the value data, and performs an access operation on an endpoint using the endpoint object.

A qubit island including a plurality of topological superconducting wires, a trivial superconducting wire that couples the topological superconducting wires, and one or more electrostatic gates located above the topological superconducting wires and the trivial superconducting wire in a thickness direction. The plurality of topological superconducting wires each have respective topological wire widths that are greater than a trivial wire width of the trivial superconducting wire. The one or more electrostatic gates are operated at a gate voltage that places the topological superconducting wires into a single sub-band regime and places the trivial superconducting wire into a depleted regime.

Technology is disclosed for adapting a dynamic audio stream to modify sounds of a sound class based on user preferences in systems including gaming systems. A user may provide, via a graphical user interface, at least one sample instance of a sound class and a transformation type. The system may use the instances of the sound class to train a generative artificial intelligence (AI) model to identify instances of the sound class in a dynamic audio stream (e.g., from game instances on a gaming instance) and to transform the instances of the sound class with the transformation type in the dynamic audio stream. The output stream from the generative AI model can be used as the audio output. As the user hears other instances of the sound class in the output audio stream, the user can provide feedback to tune the generative AI model for the user-specific instances.

Solutions disclosed herein provide for near real-time language model interaction. Efficient management of the historical state reduces token counts in prompts, enabling use of (near) real-time invocation of language models. Examples segment a transcript (or other streaming representation) of a conversation into an earlier portion for summarization and a later portion. The summary and later portion are included in a prompt for the language model, along with an instruction to return an update to a real-time contextual analysis product (e.g., an updated agenda or other data product) based on prompt. Using the summary, or a sliding time window, the token count in the prompt is reduced, enabling a more rapid cadence of updates. The update is provided as an output (e.g., an agenda update or an updated summary of the conversation), and updated summary may be efficiently used as the summary in the following prompt (i.e., the next update iteration).

Examples are disclosed relating to a circuit (600) for controlling voltage overshoot in a computing system. In one example, a circuit (600) comprises a network of shunt devices (602) arranged into a plurality of branches (604). Each branch (604) of the plurality of branches (604) includes shunt device(s) (602) connected to an enable pin (606) associated with the branch (604). Each shunt (602) is configured to induce current through a transistor (722) connected between a power node (726) and a ground node (728) when the shunt device (602) is activated. The circuit (600) comprises a controller (608) connected to a plurality of enable pins (606) corresponding to the plurality of branches (604) of the network. The controller (608) is configured to receive a computing processor voltage, generate a difference value indicating a difference between the processor voltage and a reference voltage, and send enable signal(s) to enable pin(s) (606) to activate the shunt devices (602) based at least on the difference value.

Systems, methods, and computer program products are described herein related to on-demand battery life with adaptive performance control, which achieves a target battery life (e.g., based on user request) by dynamically regulating a battery discharge slope. A power mode is implemented with adaptive power level (PL) limits (e.g., average PL limit and maximum PL limit) to manage battery life over time intervals. The discharge rate of the battery is controlled by dynamically adapting PL limits of the power provided to the computing device to align the remaining battery life with a target battery life. PL limits can be adapted based on a hardcoded mapping of a target performance levels or based on a machine learning model. Computing device performance is adaptively enhanced on demand for usage scenarios by temporary adjustment of PL limits when estimated battery life is within a margin of target battery life.

A mobile computing device (10) is configured to charge a component battery (14, 32, 52). The mobile computing device (10) comprises a power supply unit (20), a computing device battery (12), a processor (18, 504), and a memory (15, 508, 512) storing instructions (16) executable by the processor (18, 504) to control a rate of charging the component battery (14, 32, 52) from either the power supply unit (20) or the computing device battery (12) based at least in part on a magnitude of throttled power provided to the processor (18, 504).

A method for processing a glass preform for hollow core fiber comprises providing a length of glass preform for hollow core fiber from which a portion is to be removed in order to terminate the preform forming an end face. The preform comprises a transverse cross-sectional structure comprising a hollow core surrounded by a plurality of capillaries defining a plurality of voids encased by a jacket tube, wherein the hollow core and the plurality of voids extend longitudinally along the length of the preform. The method further comprises positioning the preform using a preform holder attached to a diamond wire saw, wherein the preform holder is configured to clamp the preform equally on both sides of a desired location for cutting. The preform is cut in the desired location using the diamond wire saw.

A medium voltage, modular rack system includes: an IT (Information Technology) rack configured to house a plurality of compute resources; a power panel rack comprising a plurality of power connectors and configured to operatively couple the compute resources to a low voltage DC power source; an energy storage rack comprising one or more energy storage devices, wherein each of the energy storage devices is configured to store energy from and provide energy to the low voltage DC power source; a power conversion rack comprising power conversion devices configured to convert a medium voltage power source to the low voltage DC source; and a medium voltage rack configured to receive a medium voltage AC source and a medium voltage DC source.

Methods, systems, and computer storage media for providing speech-text alignment management using a speech processing engine in an artificial intelligence system are described. The speech processing engine includes a speech-text alignment engine that operates as intermediary between raw spoken language and its corresponding textual representation, ensuring real-time processing and translation tasks. The speech processing engine utilizes a speech encoder to convert spoken language into digital signals that represent its acoustic features. The data then passes through the speech-text alignment engine, which includes a Connectionist Temporal Classification (CTC) layer for temporal and sequence alignment and a modified q-former layer for fine-tuning the speech embeddings to closely match the latent space of text embeddings used by Large Language Models (LLMs). This alignment process enables translating or transcribing spoken content into multiple languages without delay. The speech processing engine also incorporates token swapping to support adapting to various speech patterns and accents effectively.

Systems and methods for automatically reducing regression for a software update applied to a population of nodes in a computing environment. A regression detector performs a health analysis of the software update and detects a software regression attributed to the software update with high confidence by performing a combination of data analyses. In some examples, a time window-based observational study, a control-based observational study, and an anomaly detection analysis are performed for identifying various regression conditions. When the identified regression conditions match a set of high-confidence regression conditions configured for the health analysis, a software regression is detected. In further examples, the regression detector transmits an event based on the detected software regression to prevent the software regression from propagating to additional nodes in the computing environment.

Microscopic light emitting diodes (micro-LEDs) systems having in situ current measurement circuits are described. An example micro-LED system includes a set of micro-LEDs formed in a display substrate and a set of pixel driver circuits formed in a backplane substrate, coupled to the display substrate, where a respective pixel driver circuit is to provide current to a respective micro-LED during a first mode of operation for the set of micro-LEDs. The micro-LED system further includes a current measurement circuit, formed in the backplane substrate, comprising an operational amplifier configured to drive a source-follower transistor. The micro-LED system further includes a set of pass transistors to, on a per pixel driver circuit basis, selectively redirect current from one or more of the set of pixel driver circuits to the current measurement circuit allowing for in situ measurement of the redirected current within the backplane substrate.

A system processes encapsulated packets by computing a hash from the inner packet's IP header values, selecting, based on the hash, an appliance IP address (AIPA) from a group of AIPAs, and replacing the outer packet's destination IP address with the selected AIPA before forwarding the packet to the corresponding network appliance. In another aspect, a method for applying a user-defined policy involves receiving the policy which identifies traffic subject to the policy and a rule to be enforced on the traffic. Based on the policy, a group of AIPAs is determined and appliances associated with the group of AIPAs are configured to enforce the policy. The first hop switch is then configured with the AIPAs, enabling it to select an AIPA address based on a hash and replace the outer destination IP address of the traffic with the selected virtual IP address.

A disclosed method provides for distributing portions of a database query among different processors of a same host device. The method includes determining query fragments collectively executable to implement the database query; identifying one or more subplans for executing the query fragments; accessing capabilities of a hardware coprocessor of the host device; determining, based on the capabilities stored for the hardware coprocessor, whether a select subplan of the one or more subplans is viable for implementation by the hardware coprocessor; and in response to determining that the select subplan is viable for implementation by the hardware coprocessor, selecting and executing a first distributed execution plan that includes the select subplan and that delegates execution of the select subplan to the hardware coprocessor.

The present disclosure generally relates to systems and methods for adaptively retransmitting data packets utilizing a biased moving average retransmission time. Systems and methods described herein avoid the risks of retransmission flooding and congestion collapse common to existing out-of-order packet transmission systems by generating, updating, and utilizing a biased moving average retransmission time that adapts to a specific latency of a data flow. By tailoring fast and regular timeout events to this biased moving average retransmission time, the systems and methods described herein can dynamically adapt to in-the-moment network conditions to ensure that data packets reach their intended endpoints.

A processor system may include multiple circuits that are each powered by their own power rails. Incomplete cycling of system power may fail to discharge memory circuits, leaving that data vulnerable to unauthorized access upon reboot. A processor system includes a boot detection circuit that, upon detecting a threshold voltage on a first power rail in the bootup sequence, captures voltage measurements on each of the remaining power rails in multiple time periods, wherein the captured voltage measurements may be analyzed to determine whether the voltages on the power rails fully cycled to a sufficiently low voltage and remained there for a sufficient time to ensure that memory circuits adequately discharged. The power system may proceed to generate an indication of complete reset or incomplete reset depending on the voltage measurements stored in the memory. In this way, the confidentiality of data previously stored in the memory circuits is protected.

Microscopic light emitting diodes (micro-LEDs) systems with bridge circuits for scaling voltages and currents are described. An example micro-LED system includes a set of micro-LEDs having a respective voltage terminal configured to receive a first positive voltage supply and a common cathode terminal configured to receive a negative voltage supply. The micro-LED system includes sensing circuits to sense voltages and currents received from a selected subset of the set of micro-LEDs, where each of the sensing circuits is configured to receive a second positive voltage supply, different from the first positive voltage supply. The micro-LED system includes bridge circuits to: (1) during sensing of any voltages received from a selected subset of the set of micro-LEDs, scale voltages to scaled voltages, and (2) during sensing of any current received from a selected subset of the set of micro-LEDs, limit current flowing through a respective sensing circuit.

Examples are disclosed that relate to performing reasoning processes using generative language models. One disclosed example provides a method of performing a spatial reasoning task. The method comprises, iteratively, at a reasoner agent, receiving query results from a retriever agent, and based upon the query results, generating a reasoner prompt. The method further comprises inputting the reasoner prompt into a reasoner language model, receiving a reasoner output from the reasoner language model, and sending a query to a retriever agent. The method further comprises, at the retriever agent, receiving the query from the reasoner agent, generating a retriever prompt, and inputting the retriever prompt into a retriever language model. The method further comprises receiving an output from the retriever language model, querying scene data, and receiving one or more results of the query, and sending the one or more results of the query to the reasoner agent.

An isolated platform for responsible Al is described. In various examples, a method is performed by a computing device. Output is generated by executing at least part of an application in an isolated virtual machine with an input-output virtualized accelerator, where the application is an artificial intelligence application. The isolated virtual machine is used to check the output. In response to the check being successful, the output is returned. In response to the check being unsuccessful, an error message is returned.

Methods, systems, and computer program products are provided that implement vector-related requests using query operators. For example, a system includes a database, a parser, a converter, an optimizer, and an execution engine. The database is associated with query operators that are non-vector-specific. The database stores a table with vector embeddings. The parser is configured to parse a request indicating a vector operation associated with the vector embeddings. The converter is configured to convert the vector operation into a logical operator tree comprising a representation of the request as a logical flow of the query operators, enabling the vector operation without vector-specific executable code or operators. The optimizer is configured to convert the logical operator tree into an executable plan. The execution engine is configured to execute the executable plan against the table with vector embeddings.

In various examples, an image search is performed based on a query image where the result includes images with semantic information that matches the query image. For example, image encoders are used to extract information from images that are stored in an index. Continuing this example, when a query is obtained the index is searched to determine potential matches, and a comparison is performed between the query image and the potential matches to determine images with semantic information matching the query image.

Methods, devices, systems, and computer program products are provided for adaptive determination of stylus tilt by selectively using gyroscopic and sensor data for accuracy of tilt determinations as signal conditions change. A device comprises an orientation detector that determines stylus tilt using a first tilt detector based on a first set of conditions detected at a first time. The device detects a second set of conditions at a second time. The device adapts tilt determination from the first tilt detector to a second tilt detector for determination of tilt at the second time. The device determines tilt at the second time using the second tilt detector. Adaptive tilt detection can vary tilt calculations by using different equations, using the same equation with different weights, blending the results of calculations using multiple equations, etc., to maintain the accuracy of tilt determinations as signal conditions change.

Examples solutions for evaluating network connectivity between computing resources within a computing environment include: generating a network packet for a proxy ping at a first host server that provides a first virtual machine (VM), the network packet being originally generated by the first host server and not generated by the first VM, the network packet being created to include a proxy ping flag within a header of the network packet; capturing a sending timestamp; receiving a response network packet generated by a second host server in response to interception of the network packet without the network packet being sent through to the second VM, the network packet having been identified for interception by the second host server based on identifying the proxy ping flag within the header of the network packet; capturing a response timestamp; and computing a latency value based on a difference between the sending and response timestamps.

An anomaly detection model is tuned by training an anomaly detection condition tuning model using features of training input data provided to the anomaly detection model, training anomaly detection conditions used by the anomaly detection model to identify a detected anomaly in the training input data, and training data feedback identifying the detected anomaly as true or false. A first tuned training data condition is extracted for each feature by executing the anomaly detection condition tuning model on the features. A condition regression model is trained using second training data including the features of the training input data and each first tuned training data condition. Second tuned anomaly detection conditions are generated by executing the condition regression model on the new input data. The anomaly detection model is executable on the new input data using the second tuned anomaly detection conditions to identify an anomaly in the new input data.