News
Short posts prepared from approved sources with clear links back to the original articles.
RSS feedProduction AI agents can loop or skip work while services remain online. StackGen recommends nested traces, hard call limits and append-only audit logs to expose behavior and contain runaway costs.
LinkedIn’s multi-teacher pipeline trains a 600-million-parameter job ranker about eight times faster. Online serving supports experiments, while cached teacher outputs reduce repeated computation at scale.
Nvidia’s beta Personal AI Router distributes separate inference requests among compatible machines on one network. It works with Ollama and LM Studio but does not combine GPUs into shared memory.
AI researcher Timnit Gebru says industry warnings about human extinction divert attention from present harms such as autonomous weapons and concentrated power. Her argument challenges labs’ safety framing.
Anthropic says it blocked accounts that obscured biological research involving avian influenza and other risks. The company cannot establish malicious intent because the same information can support vaccines.
Google’s ToolGrad reverses synthetic agent-data generation by finding a working tool chain first. Small Gemma models trained on 500 examples improved on unseen tools, though results come from benchmarks.
Apple researchers trained a protein-design model directly on sequences and 3D structures in one stage. The simpler approach remained competitive across joint design and unconditional generation tests.
Apple researchers propose evaluating video captions with human-checked questions instead of reference wording. The method separately measures factual accuracy and coverage of important visual details.
Apple researchers built a text-to-ASL-gloss system that tracks people, places and concepts across connected sentences. It improves discourse consistency without sacrificing sentence-level quality.
Microsoft’s LLM Speech 2607 improves mixed-language transcription, names, numbers and punctuation while cutting latency. The server-side update needs no application architecture changes.
Suno’s v6 music model is its first developed with record-industry support and includes licensed partner data. It adds free and experimental variants, multimodal prompts and edits that do not regenerate an entire song.
Massachusetts will require new data centers above 25 megawatts to supply clean electricity or fund protections for ratepayers. The state is also pausing tax-exemption applications while regulators implement the rules.
AWS proposes a turn-level Agent Evaluation Metric that separates root errors from later responses that inherit them. The first implementation scores truthfulness and completeness while leaving room for safety and instruction-retention measures.
Palantir is integrating Nvidia’s open Nemotron models and cuOpt optimization software into Foundry for supply-chain planning. Nvidia’s operation, spanning millions of parts and thousands of suppliers, is the first deployment target.
A small study of AI-assisted coding found that approved specifications did not help reviewers catch more bugs. They did make detected drift easier to trace to a concrete requirement, at additional time and cost.
OpenAI’s new Data agent connects ChatGPT Work to business databases and analytics tools. It can investigate metrics and build dashboards while applying the permissions of each connected account.
d-Matrix plans to integrate its next-generation Raptor processors with Nvidia NVLink, MGX racks and Spectrum-X networking. The move gives the chipmaker a ready-made deployment stack instead of requiring a separate rack design.
MERIT evaluates whether long-term memory improves an agent’s actions, while tracking token and dollar costs. Its early results show strong gains on dependent tasks but unstable retrieval when stored facts are updated.
Researchers propose X-CoSD, a lossless inference method that lets a small on-device model draft text for a server model even when their vocabularies differ. It also cuts the token-distribution data exchanged during verification.
Public agencies are seeing sharp increases in complaints, claims and petitions as AI makes filing easier. Researchers warn that services must distinguish abuse from legitimate demand instead of treating every automated submission as spam.
Universal Music Group and ElevenLabs are developing a service for remixes and new versions built from licensed recordings. Artists can opt in, and the companies say participating creators will be compensated.
Clearview AI prototyped a tool that could expand a face-search lead into a profile of aliases, associates and activity. The company says police never used it and it is not planned for release in its current form.
AWS released sample code for an instruction-driven detector that identifies personal information with different language models. Teams can change entity types without retraining, but accuracy, cost and hallucinated labels still need monitoring.
Workflow1111 recreates most AUTOMATIC1111 features as 11 connected media pipelines on one Gradio canvas. Developers can duplicate the project, swap models and expose each output through an API.
The NSA, CISA and FBI accuse six Chinese companies of using large-scale query campaigns to distill US models. Their proposed defenses include stronger identity checks and quietly routing suspected accounts to less capable systems.
DeepSeek’s V4.1-Flash model reduces the cache used for long contexts and activates less compute while reading input. The open model targets agents with frequent tool calls, though reported strengths and limits still need independent testing.
Meta describes an enterprise-agent design that keeps expert knowledge in version-controlled files and turns corrections into tested updates. The approach aims to preserve specialist judgment without retraining the underlying model.
A University of Pennsylvania lab uses specialized models, ChatGPT and Codex to search genome data for potential antimicrobials. AI can prioritize candidates quickly, but laboratory testing remains essential before any medicine emerges.
Cloudera is integrating Mistral models into its hybrid data platform for deployments in public clouds, private infrastructure and isolated environments. The partnership targets regulated enterprises that need tighter control of data and models.
Skild AI says its S1 robot model can learn multistep physical tasks from a single video without retraining. The approach could reduce the data collection needed when factory work changes.
Cymphony combines identity, data and activity records to show what employees and software agents can reach. Its funding includes a $25 million Series A that values the startup above $100 million.
Instinct can now create accounts, contact businesses and follow up through a dedicated mailbox instead of using the owner’s inbox. The agent asks for input when needed but otherwise handles the exchange itself.
The new Sketch canvas lets users draw a rough composition and describe the desired result. ChatGPT Images 2.5 also supports comments on selected areas and promises faster generation across multiple devices.
A 91-country study found students using AI often scored lower in science, especially when it replaced reading or writing. Moderate use paired with training to critique AI output showed a more positive pattern.
Amazon is combining AI and visual effects to align mouth movements with translated dialogue. The feature starts with the English dub of Maxton Hall and has not yet received a wider title list.
Muse is rolling out through apps, the web and WhatsApp to perform digital tasks for users. Meta isolates sessions in virtual machines and requires direct approval for sensitive actions, but can technically access current data.
Kepler Computing says 3D stacking and a ferroelectric material can raise memory density in existing fabs without EUV tools. The startup has tested 2,000 wafers but has not yet reached mass production.
The new UpdateRecord API changes selected machine-learning features without rewriting an entire record. It reduces extra reads and avoids pipelines silently overwriting one another’s changes.
Pathway’s BDH-CQ model refines answers inside a recurrent latent state instead of writing long reasoning traces. The company reports strong ARC-AGI-1 cost efficiency, but broader validation is still needed.
AWS customers can call OpenAI’s GPT-6 Astra through Bedrock or connect it to ChatGPT Work and Codex. The release supports million-token inputs, prompt caching and browser-based workflows.