AI
A newly published application organizes retrieval by time period and graph community, then routes a question to the relevant slice of a changing dataset.
AI
A newly published application describes a video-based multimodal system with a third component wedged between the image encoder and the language model. The placement is the whole invention.
AI
A Sony application published July 16, 2026 describes a generative AI system that embeds the authentication key into the content it produces — and makes the output measurably worse when that key is invalid. The licensing check stops being a gate and becomes a property of the artifact.
Artificial Intelligence
A patent application published July 16, 2026 describes splitting an n-bit weight matrix into n one-bit matrices and precomputing every activation-group permutation into a lookup table, so quantized model inference becomes retrieval and addition rather than multiply-accumulate. The disclosed target is edge deployment on ordinary CPUs.
Artificial Intelligence
A patent application published July 2, 2026 describes giving a machine-learning agent its permissions as a natural-language 'semantic entitlement,' then using a generative model to resolve which files and resources that actually covers. Here is how the mechanism works.
Artificial Intelligence
A patent granted June 30, 2026 describes a flow-style vocoder that runs the same network two ways: training maps recorded speech to Gaussian noise, and inference maps noise back to a waveform. Here is how the mechanism works, and where it sits among a wave of generative-media grants issued to NVIDIA the same day.
Artificial Intelligence
A newly published application describes a neural network built as an analog electrical circuit that learns by clamping its inputs and outputs, letting the circuit settle, and reading weight updates off the hardware itself. It is one of roughly 97 machine-learning applications in this week's drop, and a useful window into where analog AI hardware sits in the field.
Artificial Intelligence
The NIST AI RMF is a voluntary framework for managing the risks of AI systems, organized around four functions: govern, map, measure, and manage. What the document says, in its own terms.
Artificial Intelligence
Most machine-learning and neural-network patents carry the Cooperative Patent Classification subclass G06N. What the code means, what its subgroups cover, and why it is the entry point for reading the AI patent landscape.
Artificial Intelligence
Capital expenditure on data centers is the spending hyperscalers report for building AI capacity. It is disclosed plainly in filings, and its trajectory is the defining figure of the AI buildout. What the term means, read through one company's 10-K.
Artificial Intelligence
High-bandwidth memory stacks DRAM dies beside the processor to feed it data faster. It exists to relieve the memory-bandwidth bottleneck that limits AI accelerators. What HBM is, and why it shows up in the supply story.
Artificial Intelligence
Training builds a model; inference runs it. One is a large, one-time capital event; the other is a recurring per-query operating cost. Why the distinction decides how AI is priced, served, and disclosed.
Artificial Intelligence
Reinforcement learning from human feedback is the fine-tuning step that makes a large language model do what a user actually asks. Here is the mechanism, traced to the InstructGPT paper that demonstrated it.
Artificial Intelligence
The transformer is the architecture under nearly every modern large language model. Its defining move was to drop recurrence and convolution entirely and rely on an attention mechanism. Here is what that means, from the 2017 paper that named it.
Artificial Intelligence
A mixture-of-experts layer routes each input to a small subset of specialized sub-networks instead of running the whole model. Here is the mechanism, traced to the 2017 paper that made it work at scale.
Artificial Intelligence
A new arXiv preprint introduces a taxonomy-guided benchmark for agentic academic paper search built from more than 1,000 computer-science topics and four research intents. The authors report agentic methods beat single-shot retrieval while leaving large gaps in recall.
Artificial Intelligence
A new arXiv preprint describes a multi-agent framework that lets a language model query LLVM's internal optimization state and use runtime measurements to refine compiler options. The authors report it operates without training or fine-tuning.
Tech Brief
A published application describes how to carve a convolution's input tensor into memory-sized tiles and rebuild the result exactly, by computing the overlap each tile must carry from its neighbors. It sits inside a wider cluster of reconfigurable-dataflow filings published the same week.
Artificial Intelligence
The General Services Administration is floating a draft acquisition clause on safeguarding data inside large language model systems — and, unusually, asking for comment before it commits to rulemaking.
Artificial Intelligence
A new arXiv release combines system, network, and browser logs with per-entry MITRE ATT&CK labels, then shows that small language models fine-tuned with LoRA can learn to read across all three.
Artificial Intelligence
Dataset distillation promises to compress huge training sets into a few synthetic images. A large arXiv benchmark finds that on real datasets, simply selecting a subset of real samples is competitive — and far cheaper.
Artificial Intelligence
A new arXiv framework called RubricsTree replaces vague 'LLM-as-a-judge' scoring of health agents with 100-plus clinically verifiable yes/no checks, and reports large gains when those checks become training signal.
Artificial Intelligence
FPRM, a new arXiv model, loops a Transformer over its own output and halts when the answer stops moving, letting it spend more compute on hard problems and less on easy ones.
Artificial Intelligence
A new arXiv paper called VERITAS pairs a general-purpose robot policy with a separate visual checker that scores candidate actions at inference time, then turns the actions it approves into free training data.
