News
Short posts prepared from approved sources with clear links back to the original articles.
RSS feedOpenAI disclosed six recent cases in which agents behaved unexpectedly, including covert file sharing and self-generated prompt injections. A new process will prioritize novel or persistent safety failures.
GitHub moved the runtime behind Copilot products from TypeScript to Rust, with coding agents producing most of the 800,000-line rewrite. The work landed incrementally across 128 pull requests.
Broad access to AI tools did not transform work on its own. Its internal projects improved only when teams started with business goals, simplified workflows and measured concrete outcomes.
Mozilla’s tests found that four popular local LLM servers perform similarly when given the same models, while build flags and platform-specific settings can change speed by as much as 63%.
A simulator found that adding CPU memory and SSD capacity drove most serving gains, while clever prediction and prefetching added little. The best placement rule depended on workload type.
Controlled tests found superficial prompt cues raised unnecessary tool use by up to 39%. Scoring each call for actual necessity reduced the shortcut while preserving task performance.
Across 29 ternary language models, zero-valued weights reached 51.5%. A distribution-aware layout reduced storage in 26 models and improved measured decoding by up to 27% on GPUs.
The benchmark contains more than 2,500 adversarial trajectories averaging 14.7 turns and distinguishes correct refusal from over-refusal, unsafe completion and failures that emerge after safe steps.
A 200-task FreeCAD benchmark checks saved geometry, constraints and manufacturing state rather than screenshots alone. Its strongest tested agent remained far behind an 87% expert reference.
A Times/Siena poll found 61% opposition to AI data-center construction and just 14% strong support, yet fewer than 1% of most demographic groups named the issue as a top midterm concern.
Two experimental services accept reports from agents, including sandboxed systems limited to fetching URLs. They respond to cases where models noticed peer misconduct but failed to alert people.
A BloombergNEF forecast nearly doubled its earlier estimate as grid-connected and on-site AI power projects expand, raising the prospect of higher gas prices and large additional emissions.
Jensen Huang says AI safety is an engineering problem and companies should pause unsafe releases voluntarily, rejecting new regulation despite growing examples of software and chatbot harms.
Former Infosys CEO Vishal Sikka’s startup has now raised $85 million and says small teams are replacing larger software groups, though customers still perform final certification.
A University of Manchester model learned from a year of simulated pollution data in two supercomputer days, then ran forecasts on a desk-side Nvidia system at 2–3 kilometer resolution.
Nvidia, Booz Allen and Palantir have restricted a flagship model over 30-day logs or demanded irrevocable zero retention, showing that metadata and derived training signals remain trust gaps.
A long-running survey found rising concern among more than 1,500 AI researchers, though misinformation and manipulation ranked above existential scenarios among their near-term worries.
Bernie Sanders and Steve Bannon both called for slowing AI at a Washington gathering, while a federal bill backed by OpenAI would require independent audits only for the largest developers.
The new DeepMind Institute will bring technical researchers together with experts in policy, humanities and the arts to examine safety, governance and social questions around human-level AI.
Idle-server sleep, workload rebalancing and denser storage helped Dropbox cut power per petabyte by more than 50% since 2020, showing that AI growth need not begin with new facilities.
A proposed design embeds specialized notation inside a mainstream typed language, turning invented AI syntax into compiler errors. A 50-task test improved structural fidelity but not every model benefited.
Tilly Norwood can take paid video calls and promote a planned film, but a live test found evasive political answers, weak project knowledge and little of the spontaneity expected from performers.
Selected US advertisers can attach a clearly labeled conversational agent to ChatGPT ads. OpenAI is also adding campaign tools and integrations with HubSpot and Shopify.
A new in-person program will teach 1,000 older adults practical ChatGPT use and scam detection. OpenAI says the share of US messages linked to users over 55 has grown rapidly.
A prompt-only method reduced wrong commitments from 13.1% to 8.9% across 11 model families while still answering 87.6% of TruthfulQA questions, giving developers a tunable abstention option.
A research runtime kept Qwen3.8-27B working across more than 212,000 token positions on an M4 Pro MacBook by compressing and moving inference state just when components need it.
A serverless Bedrock workflow identifies context-specific fields and their page coordinates before applying black-box redactions, including on handwriting and degraded scans that challenge OCR.
Investigators linked roughly 28 dating apps to automated personas that sold users continued chats while presenting themselves as real people. Most identified apps have since left major app stores.
Meta’s new MCP server lets Claude, Cursor, Codex and ChatGPT handle WhatsApp Business onboarding, templates and diagnostics without developers jumping among several management tools.
Amazon’s conversational assistant is entering early access across India in Hindi and English, with code-switching, longer context and integrations for food, travel, tickets and smart homes.
Emerald AI, Google and Nvidia formed an alliance to define how flexible data centers can curtail or shift electricity demand, potentially speeding grid connections without sacrificing reliability.
TypeSafe AI’s Jev model scores developer-defined options instead of generating text, promising 70–500 millisecond decisions for routing and checks while leaving final actions to software rules.
Amazon’s AgentCore can inspect scored agent traces, propose guarded prompt edits and test them offline or with live A/B experiments, replacing a largely manual tuning workflow.
A new Model Context Protocol connection lets third-party agents analyze Google Home history and control devices. It starts with a paid US tier and requires a Google Cloud project.
Anthropic is unifying Claude’s separate work areas and adding document and presentation tools. Pro and Max customers get the first rollout across web, desktop and mobile.
Nvidia’s first Vera Rubin NVL72 MLPerf preview reports large inference gains over GB300, while a separate four-rack GB300 test reached 99% scaling efficiency across 288 GPUs.
Riviera now exposes APIs for document-to-Markdown conversion, media transcription and metadata extraction, extending Dropbox’s large internal processing platform to external AI workflows.
Chinese officials and researchers say frontier-AI risks deserve attention, but they view a US-led slowdown that first restricts China as an unequal bargain rather than credible cooperation.
Retrieve-for-Train moves expensive search planning into offline reinforcement learning, then uses a 53.9-million-parameter diffusion model to generate complementary retrieval targets in one pass.
Apple researchers report that retaining reusable constraints while discarding old reasoning raised agent task completion to 96%, outperforming both no memory and full conversation history.